Вы находитесь на странице: 1из 187

Advances in Industrial Control

For further volumes:


www.springer.com/series/1412

Cristian Kunusch r Paul Puleston


Miguel Mayosky

Sliding-Mode
Control of PEM
Fuel Cells

Cristian Kunusch
Institut de Robtica i Informtica Industrial
CSIC-Universitat Politcnica de Catalunya
Barcelona
Spain
Paul Puleston
Depto. Electrotecnia
Laboratorio de Electrnica Industrial,
Control e Instrumentacin (LEICI)
Universidad Nacional
de La PlataCONICET
La Plata, Buenos Aires
Argentina

Miguel Mayosky
Depto. Electrotecnia
Laboratorio de Electrnica Industrial,
Control e Instrumentacin (LEICI)
Universidad Nacional de La Plata
Comisin de Investigaciones Cientcas
de la Provincia de Buenos Aires
(CICpBA)
La Plata, Buenos Aires
Argentina

ISSN 1430-9491
e-ISSN 2193-1577
Advances in Industrial Control
ISBN 978-1-4471-2430-6
e-ISBN 978-1-4471-2431-3
DOI 10.1007/978-1-4471-2431-3
Springer London Dordrecht Heidelberg New York
British Library Cataloguing in Publication Data
A catalogue record for this book is available from the British Library
Library of Congress Control Number: 2012930019
Springer-Verlag London Limited 2012
Apart from any fair dealing for the purposes of research or private study, or criticism or review, as permitted under the Copyright, Designs and Patents Act 1988, this publication may only be reproduced,
stored or transmitted, in any form or by any means, with the prior permission in writing of the publishers, or in the case of reprographic reproduction in accordance with the terms of licenses issued by the
Copyright Licensing Agency. Enquiries concerning reproduction outside those terms should be sent to
the publishers.
The use of registered names, trademarks, etc., in this publication does not imply, even in the absence of a
specic statement, that such names are exempt from the relevant laws and regulations and therefore free
for general use.
The publisher makes no representation, express or implied, with regard to the accuracy of the information
contained in this book and cannot accept any legal responsibility or liability for any errors or omissions
that may be made.
Printed on acid-free paper
Springer is part of Springer Science+Business Media (www.springer.com)

To my parents Zulma and Orlando, of course.


To Maria and all the people who encourage
me everyday, but specially to my tireless
co-authors.
To Mark, Ana, Tommy, Kate and my parents.
To Daniela, Matilde and the memory of
Norma and Alejandro.

Series Editors Foreword

The series Advances in Industrial Control aims to report and encourage technology transfer in control engineering. The rapid development of control technology
has an impact on all areas of the control discipline. New theory, new controllers,
actuators, sensors, new industrial processes, computer methods, new applications,
new philosophies, . . . , new challenges. Much of this development work resides in
industrial reports, feasibility study papers and the reports of advanced collaborative
projects. The series offers an opportunity for researchers to present an extended exposition of such new work in all aspects of industrial control for wider and rapid
dissemination.
Fuel cells are one of the devices promoted as a potential technological contributor to the green energy economy. Fuel cells are an energy conversion device that
uses the catalytic oxidation of hydrogen at an anode with the catalytic reduction of
oxygen at a cathode to release energy as electricity and heat with the non-polluting
by-product of water. So long as hydrogen and oxygen fuel is supplied (at the appropriate operating conditions, and with sufcient purity to avoid possible catalyst
poisoning), a fuel cell will provide a continuous energy output. In this way a fuel
cell does not suffer the fuel depletion that characterises the energy storage of battery
technology.
Several different fuel cell types arise from the use of a variety of different electrolyte and electrode media, with differing operating temperatures. To create a power
unit, fuel cells are then combined in series or parallel in planar or tubular stacks
and supplied with auxiliary equipment. Following on from this are requirements for
pressure regulation, temperature control and humidity-level maintenance. The complete assembly involves electrochemical, physical, chemical, and thermal processes
and is consequently, a highly nonlinear system. Finally, add to this that some internal variables are difcult to measure, that instrumentation overheads should be kept
low for portabilitys sake, and that system models have associated model and parametric uncertainty, and the fuel-cell-control problem clearly emerges as a highly
challenging one.
This technological prcis is the starting point for the authors Cristian Kunusch,
Paul F. Puleston, and Miguel A. Mayosky who make a very convincing case for
vii

viii

Series Editors Foreword

the benets that modern nonlinear closed-loop control has to offer fuel-cell technology. Their research is with the proton exchange membrane (PEM) fuel cell. Apparently, PEM fuel cells have a high power density, low weight, and volume that make
them particularly appropriate for portable or mobile applications. The authors chosen control paradigm is that of sliding-mode control, and in this monograph, they
give a comprehensive report of their research; however, what makes this monograph
particularly appropriate to the Advances in Industrial Control series is the way in
which the authors combine a wide of variety of skills, empirical nous, and technical
knowledge to develop a practical demonstration prototype for the fuel-cell-control
application.
The monograph is self-contained and opens with introductory material on fuel
cells, PEM fuel cells, and sliding-mode control (Chaps. 1, 2 and 3). For a rst
demonstration of their sliding-mode control proposals, the authors use their own
reformulation of a benchmark model that was previously reported in an earlier seminal Advances in Industrial Control monograph, Control of Fuel Cell Power Systems
by Jay T. Pukrushpan, Anna G. Stefanopoulou, and Huei Peng (ISBN 978-1-85233816-9, 2004). Thus the rst test of the sliding-mode control is through a detailed
simulation benchmark case study (Chap. 4).
The authors research with an experimental PEM-fuel-cell test bed is then reported in Chaps. 5 and 6 in the monograph. As so often occurs in process control case studies, the system modelling (Chap. 5) follows a process-based modular structure and the model is derived with a judicious blend of physically based
equations and experimental-data-supported empiricism. The trade-off between the
actual process behaviour and the accuracy that can be achieved with a tractable
model naturally introduces a level of system and parametric uncertainty and leads
to requirements for a robust and reliable control scheme. Chapter 6 presents the
authors second-order sliding-mode control design and the results from the experimental test bed facility. All is fully documented and offers an opportunity for further
independent study of the control system, its results, and the control outcomes. The
monograph closes with a chapter of conclusions and a discussion of future possible
research issues for this system and fuel-cell control per se.
A wide range of readers from the control community, and also those involved in
new energy technologies, will nd much of interest in this research from authors
Cristian Kunusch, Paul F. Puleston, and Miguel A. Mayosky. This volume is also a
very welcome complement to the earlier fuel-cell monograph mentioned above and
now provides the Advances in Industrial Control series with two excellent texts on
control for fuel-cell technologies.
Industrial Control Centre
Glasgow
Scotland, UK

M.J. Grimble
M.A. Johnson

Acknowledgements

There are many people and institutions to whom we would like to express our gratitude. Too many, in fact. And each one of us has our particular debts of gratitude.
Therefore, we are forced to just explicitly mention those that directly contributed to
the realisation of this book. To all the others, many thanks.
We wish to acknowledge the support of the National University of La Plata
(UNLP), CONICET and CICPBA from Argentina, the Technical University of Catalonia (UPC), CSIC and AECID from Spain, and the Seventh Framework Programme of the European Community through the Marie Curie actions.
We specially like to thank the Laboratorio de Electrnica Industrial, Control e
Instrumentacin (LEICI), Faculty of Engineering (UNLP) and its mentor, Professor Carlos F. Christiansen, who created an institution where we could initiate and
develop our careers in the eld of scientic-technological research.
We also wish to express profound thanks:
To Professor Jordi Riera Colomer for his enduring collaboration and generosity,
and to each and every member of the IRIs Fuel Cell Laboratory, whose prociency and dedication made the experimental phase of this investigation a reality.
To Professor Enric Fossas Colet, without whom this collaboration would not
have been at all possible. Thanks him for his continuous support and unbreakable friendship.
To Professor Leonid Fridman for his wise advice on Higher-Order Sliding-Mode
control and, moreover, for his unselsh help all the way.
To our friends and colleagues in the sliding mode community, who have supported and encouraged us during the different stages of our research.
A nal comment. We have made use of material from several books and papers
in this work. We are very grateful to their authors, and we do hope that we have
cited them abundantly and appropriately, acknowledging the full recognition they
deserve.

ix

Contents

Introducing Fuel Cells . . . . . . . . . . . . . . . . .


1.1 An Outlook of the Energy Perspective . . . . . . .
1.2 What is a Fuel Cell? . . . . . . . . . . . . . . . .
1.3 State-of-the-Art on Fuel Cell Systems Technology
1.3.1 Fuel Cell Units . . . . . . . . . . . . . . .
1.3.2 Fuel Cell Stacks . . . . . . . . . . . . . .
1.3.3 Auxiliary Components . . . . . . . . . . .
1.4 A Brief History of Fuel Cells . . . . . . . . . . .
1.5 Closed-Loop Operation of Fuel Cells. Why? . . .
1.6 Scope and Outline of the Book . . . . . . . . . .
References . . . . . . . . . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

1
1
3
3
4
5
6
6
9
10
10

PEM Fuel Cell Systems . . . . . . . . . . . . . . . . . . .


2.1 An Introduction to PEM Technology . . . . . . . . . .
2.2 Basics of PEM Fuel Cells Operation . . . . . . . . . .
2.3 Efciency and Power Conversion . . . . . . . . . . . .
2.4 State-of-the-Art in PEM Fuel Cells Technology . . . .
2.5 Components and Associated Devices . . . . . . . . . .
2.5.1 Polymeric Membranes . . . . . . . . . . . . . .
2.5.2 Electrodes . . . . . . . . . . . . . . . . . . . .
2.5.3 Gas Diffusion Layers . . . . . . . . . . . . . .
2.5.4 Sealing Gaskets . . . . . . . . . . . . . . . . .
2.5.5 Bipolar Plates . . . . . . . . . . . . . . . . . .
2.5.6 Auxiliary Devices . . . . . . . . . . . . . . . .
2.6 Available PEM Fuel Cell Models in the Open Literature
2.6.1 Control Oriented Models . . . . . . . . . . . .
2.6.2 Control Objectives and Challenges . . . . . . .
2.6.3 Recent Advances on PEM Fuel Cell Control . .
References . . . . . . . . . . . . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

13
13
14
17
18
19
19
20
21
22
22
24
27
28
29
30
31

Fundamentals of Sliding-Mode Control Design . . . . . . . . . . . .


3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

35
35
xi

xii

Contents

3.2 Variable Structure Control Preliminaries . . . . . . . . . . . . .


3.3 Fundamentals of Sliding-Mode Control . . . . . . . . . . . . . .
3.3.1 Diffeomorphisms, Lie Derivative and Relative Degree . .
3.3.2 First-Order Sliding Mode . . . . . . . . . . . . . . . . .
3.3.3 Equivalent Control Regularisation Method. Ideal Sliding
Dynamics . . . . . . . . . . . . . . . . . . . . . . . . .
3.3.4 Existence Conditions for the First-Order Sliding Regime .
3.3.5 Extension to Nonlinear Systems Non-afne in Control . .
3.3.6 Filippov Regularisation Method . . . . . . . . . . . . . .
3.3.7 Discontinuous Control Action in Classic Sliding-Mode
Control. Chattering Problem . . . . . . . . . . . . . . . .
3.4 Some General Concepts on Higher-Order Sliding Modes . . . . .
3.4.1 Denition of Differential Inclusion . . . . . . . . . . . .
3.4.2 Sliding Modes on Manifolds . . . . . . . . . . . . . . .
3.4.3 Sliding Modes and Constraint Functions. Regularity
Condition . . . . . . . . . . . . . . . . . . . . . . . . .
3.4.4 Closing Comments on Higher-Order Sliding Modes in
Control Systems . . . . . . . . . . . . . . . . . . . . . .
3.5 Design of Second-Order Sliding-Mode Controllers . . . . . . . .
3.5.1 Second-Order Sliding Generalised Problem . . . . . . . .
3.5.2 Solution of the Control Problem. SOSM Algorithms . . .
3.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . .
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4

.
.
.
.

35
36
37
39

.
.
.
.

40
42
45
45

.
.
.
.

47
48
49
50

51

.
.
.
.
.
.

53
54
54
56
69
69

Assessment of SOSM Techniques Applied to Fuel Cells. Case Study:


Electric Vehicle Stoichiometry Control . . . . . . . . . . . . . . . . . 73
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
4.2 Succinct Description of the Electric Vehicle Fuel Cell System . . . 74
4.2.1 Air Compressor . . . . . . . . . . . . . . . . . . . . . . . 75
4.2.2 Air Supply Manifold . . . . . . . . . . . . . . . . . . . . . 76
4.2.3 Air Humidier and Temperature Conditioner Subsystems . 77
4.2.4 Cathode Channels . . . . . . . . . . . . . . . . . . . . . . 77
4.2.5 Anode Channels . . . . . . . . . . . . . . . . . . . . . . . 78
4.2.6 Water Model of the Polymeric Membrane . . . . . . . . . 79
4.2.7 Return Manifold . . . . . . . . . . . . . . . . . . . . . . . 80
4.3 Electric Vehicle FCGS State Space Model for SOSM Control
Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
4.4 Control Objective and Sliding Surface . . . . . . . . . . . . . . . 82
4.5 Design of a SOSM Super-Twisting Controller for the Electric
Vehicle FCGS . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
4.6 SOSM Super-Twisting Controller Simulation Results . . . . . . . 88
4.7 Comparison with Other Control Strategies . . . . . . . . . . . . . 93
4.7.1 Different SOSM Control Algorithms . . . . . . . . . . . . 93
4.7.2 LQR Controller . . . . . . . . . . . . . . . . . . . . . . . 97
4.8 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102

Contents

xiii

Control-Oriented Modelling and Experimental Validation


of a PEMFC Generation System . . . . . . . . . . . . . . .
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . .
5.2 Air Compressor Subsystem . . . . . . . . . . . . . . . .
5.2.1 Air Compressor Motor Dynamic Equations . . . .
5.2.2 Diaphragm Vacuum Pump Characteristics . . . .
5.3 Air Supply Manifold Subsystem . . . . . . . . . . . . . .
5.4 Air Humidication Subsystem . . . . . . . . . . . . . . .
5.4.1 Step 1 . . . . . . . . . . . . . . . . . . . . . . .
5.4.2 Step 2 . . . . . . . . . . . . . . . . . . . . . . .
5.5 Fuel Cell Stack Flow Subsystem . . . . . . . . . . . . .
5.5.1 Cathode Channels . . . . . . . . . . . . . . . . .
5.5.2 Anode Channels . . . . . . . . . . . . . . . . . .
5.5.3 Membrane Water Transport . . . . . . . . . . . .
5.6 Electrical Characterisation of the Fuel Cell Stack . . . . .
5.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . .
References . . . . . . . . . . . . . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

105
105
106
107
111
112
113
114
115
117
117
120
121
123
126
128

SOSM Controller for the PEMFC-Based Generation System.


Design and Implementation . . . . . . . . . . . . . . . . . . . .
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . .
6.2 State Space Model of the Experimental Fuel Cell System for
Control Design . . . . . . . . . . . . . . . . . . . . . . . . .
6.2.1 Control Output . . . . . . . . . . . . . . . . . . . . .
6.2.2 Auxiliary Functions . . . . . . . . . . . . . . . . . .
6.3 Control Objective . . . . . . . . . . . . . . . . . . . . . . .
6.4 Controller Synthesis . . . . . . . . . . . . . . . . . . . . . .
6.4.1 Super-Twisting Algorithm . . . . . . . . . . . . . . .
6.4.2 Twisting Algorithm . . . . . . . . . . . . . . . . . .
6.4.3 Sub-Optimal Algorithm . . . . . . . . . . . . . . . .
6.4.4 Feedforward Term . . . . . . . . . . . . . . . . . . .
6.5 Simulation Results . . . . . . . . . . . . . . . . . . . . . . .
6.6 Experimental Set-up of the PEM Fuel Cell System . . . . . .
6.7 Experimental Tests . . . . . . . . . . . . . . . . . . . . . . .
6.7.1 Regulation Tests . . . . . . . . . . . . . . . . . . . .
6.7.2 Perturbation Tests . . . . . . . . . . . . . . . . . . .
6.8 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . .
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

129
133
133
134
137
139
140
140
141
141
143
144
144
148
150
152

Conclusions, Open Lines and Further Reading


7.1 Main Results . . . . . . . . . . . . . . . . .
7.2 Further Issues Related to PEMFC Control .
7.2.1 Adaptive Super-Twisting Algorithms
7.2.2 HOSM MIMO Control . . . . . . .
7.2.3 Model Predictive Control . . . . . .
7.2.4 Observers for Internal Variables . . .

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

153
153
154
154
155
155
156

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

. . . 129
. . . 129

xiv

Contents

7.3 Further Issues Related to Fuel-Cell-Based Systems


7.3.1 Hybrid Standalone Systems . . . . . . . .
7.3.2 Distributed Generation Systems . . . . . .
References . . . . . . . . . . . . . . . . . . . . . . . .

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

157
157
157
158

Appendix A Electric Vehicle PEM Fuel Cell Stack Parameters


A.1 Return Manifold Polynomial Fitting . . . . . . . . . . . .
A.2 Differential Equations Parameters . . . . . . . . . . . . .
A.3 Controller Parameters . . . . . . . . . . . . . . . . . . .
References . . . . . . . . . . . . . . . . . . . . . . . . . . . .

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

161
161
161
163
164

Appendix B

.
.
.
.

.
.
.
.

.
.
.
.

Laboratory FC Generation System Parameters . . . . . . . 165

Appendix C Laboratory FCGS State Space Functions and Coefcients


C.1 Expression of (x, u, t) . . . . . . . . . . . . . . . . . . . . . . .
C.2 Model Coefcients . . . . . . . . . . . . . . . . . . . . . . . . . .
C.3 Feedforward Control Action . . . . . . . . . . . . . . . . . . . . .

169
169
172
173

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175

Notation and Acronyms

FC
FCGS
LQR
PEM
SM
SOSM
ST
a
a0
a1
a2
a3
A0
A1
A00
A10
A20
A01
A11
A02
Aapp
Afc
b
B00
B10
B20
B01
B11
B02
Cp
C0

Fuel cell
Fuel cell generation system
Linear Quadratic Regulator controller
Proton Exchange Membrane
Sliding mode
Second-order sliding mode
Super-Twisting
Polarisation curve coefcient
Empirical coefcient of the membrane water content model
Empirical coefcient of the membrane water content model
Empirical coefcient of the membrane water content model
Empirical coefcient of the membrane water content model
Compressor model tting parameter
Compressor model tting parameter
Compressor model tting parameter
Compressor model tting parameter
Compressor model tting parameter
Compressor model tting parameter
Compressor model tting parameter
Compressor model tting parameter
Apparent fuel cell area
Fuel cell active area
Polarisation curve coefcient
Compressor model tting parameter
Compressor model tting parameter
Compressor model tting parameter
Compressor model tting parameter
Compressor model tting parameter
Compressor model tting parameter
Specic heat capacity of air
Cathode humidier model tting parameter
xv

xvi

C1
C3
cv,an
cv,ca
Dw
E
F
Ga
GH2
GN2
GO2
Gm,dry
Gv
i
Icp
i0
ia
Iload
Ist
J
Jg
Ksm
k, kI
Kca,out
kdct , kdcv , Rdc
Ksm
k
L
m
mH2 ,an
mH2 O,an,liq
mH2 O,ca,liq
mhum
mN2 ,ca
mO2 ,ca
msm
mv,an
mv,ca
mv2 ,an
mv2 ,ca
mw,an
n

Notation and Acronyms

Cathode humidier model tting parameter


Cathode humidier model tting parameter
Water concentration at the membrane surface on the anode
side
Water concentration at the membrane surface on the cathode
side
Diffusion coefcient of water in the membrane
Nernst voltage of a PEM fuel cell
Faraday constant
Dry air molar mass
Hydrogen molar mass
Nitrogen molar mass
Oxygen molar mass
Membrane dry molecular weight
Vapour molar mass
Stack current density
Armature current of the compressors DC motor
Exchange current density
Armature current of the compressors DC motor
Load current
Fuel cell stack current
Combined inertia of the compressor motor and the
compression device
Compressor and air manifold gathered or equivalent inertia
Supply manifold restriction constant
Proportional and integral gains of the LQR controller
Cathode output restriction
Compressor DC motor constants
Supply manifold restriction constant
DC motor constant
Electrical inductance of the DC motor stator winding
Polarisation curve coefcient
Hydrogen mass in the anode channels
Mass of liquid water in the anode channels
Mass of liquid water in the cathode channels
Cathode humidier mass of air
Nitrogen mass in the cathode channels
Oxygen mass in the cathode channels
Supply manifold mass of air
Vapour mass in the anode channels
Vapour mass in the cathode channels
Vapour mass in the anode channels
Vapour mass in the cathode channels
Mass of water inside the anode
Number of fuel cells in the stack

Notation and Acronyms

nd
nc
n0
n1
n2
Pamb
Pan
Pca
Pcp
PH2 ,an
PH2 ,lh,an
Phum
Phum,an
Phum,ca
Phum,diff
Pnet
PN2 ,ca
PN2 ,hum
PN2 ,lh
PO2 ,ca
PO2 ,hum
PO2 ,lh
Prm
Psat
Psm
Pst
Pv,an
Pv,ca
Pv,hum
Pv,lh
Pv,lh,an
q
r1 , r2
RH amb
R
R0
R1
Ra
RH an
RH ca
RH hum
RH lh
s
s0
Trm

xvii

Electro-osmotic drag coefcient


Polarisation curve coefcient
Empirical parameter of the electro-osmotic coefcient model
Empirical parameter of the electro-osmotic coefcient model
Empirical parameter of the electro-osmotic coefcient model
Atmospheric pressure
Stack pressure drop (anode)
Cathode pressure
Compressor power demand
Hydrogen partial pressure in the anode channels
Anode line heater hydrogen partial pressure
Cathode humidier pressure
Anode humidier pressure
Cathode humidier pressure
Cathode humidier pressure drop
FCGS net power
Nitrogen partial pressure in the cathode channels
Cathode humidier nitrogen partial pressure
Cathode line heater nitrogen partial pressure
Oxygen partial pressure in the cathode channels
Cathode humidier oxygen partial pressure
Cathode line heater oxygen partial pressure
Return manifold pressure
Vapour saturation pressure of a gas
Supply manifold pressure
Stack generated power
Vapour partial pressure in the anode channels
Vapour partial pressure in the cathode channels
Cathode humidier vapour partial pressure
Cathode line heater vapour partial pressure
Anode line heater vapour partial pressure
Integral state variable of the LQR
Twisting controller parameters
Relative humidity of the air
Electrical resistance of the DC motor stator winding
Resistance coefcient
Resistance coefcient
Specic gas constant of atmospheric air
Relative humidity of the gas in the anode channels
Relative humidity of the gas in the cathode channels
Relative humidity of the gas exiting the cathode humidier
Relative humidity of the gas exiting the cathode line heater
Sliding variable or output
Size of the validity region |s(t, x)| < s0
Return manifold temperature

xviii

Tamb
Tcp
Te
Thum
Thum,an
Thum,ca
Tl
Tl,amb
Tlh
Tsm
Tst
tm
Tp
Tl
U, , , sM
uff
ui = u + uff
u1 , u2
Vact
Vrm
Vca
Vconc
Vcpff
Vcpi
Vcpp
Vcp
Vfc
Vhum
Vnor
V0
Vohm
Vsm
Vst
Wa,hum
Wan
Wan,out
Wca
Wca,out
Wrm,out
Wair,ref
Wcp
Wdry,ref
W H2
WH2 ,an,in

Notation and Acronyms

Atmospheric temperature
Compressor air temperature
Compressor motor torque
Cathode humidier temperature
Anodes humidier temperature
Cathodes humidier temperature
Load torque
Compressor load torque at ambient pressure
Cathode line heater temperature
Manifold air temperature
Stack temperature
Thickness of the membrane
Unknown torque disturbance
Compressor extra load torque at pressures higher ambient
Sub-Optimal controller parameters
Feedforward control action
Implemented control action
Super-Twisting control action components
Activation losses
Return manifold volume
Volume of the cathode channels
Concentration losses
Compressor voltage feedforward term
Implemented compressor voltage command
Compressor voltage LQR pre-compensation term
Input voltage of the compressor DC motor
Voltage of a single fuel cell
Cathode humidier volume
DC motor voltage normalisation constant
Theoretical electrochemical potential of single
hydrogen/oxygen fuel cell
Ohmic losses
Supply manifold volume
Stack voltage
Mass ow of dry air exiting the cathode humidier
Mass ow entering the anode channels
Mass ow exiting the anode channels
Cathode input mass ow of air
Flow entering the return manifold
Outlet mass ow of the return manifold
Air mass ow reference
Compressor air mass ow
Reference dry air ow
Hydrogen mass ow
Mass ow rate of H2 entering the anode

Notation and Acronyms

WH2 ,an,out
WH2 ,react
Whum
Wl,ca,out
WN2 ,ca,in
WN2 ,ca,out
WO2 ,ca,in
WO2 ,ca,out
WO2 ,ca,ref
WO2 ,react
Wsm
Wv,an,in
Wv,an,out
Wv,ca,gen
Wv,ca,in
Wv,ca,out
Wv,hum
Wv,inj
Wv,mem
x
XO2 ,ca
XO2 ,ca,out

ca
an
O2
x 
x 
cp
dc
, ,

m , M ,
ca
an
m
O 2

amb
ca
hum

xix

Mass ow rate of H2 leaving the anode


Rate of H2 reacted
Cathode humidier output ow of air
Mass ow of liquid water exiting the cathode channels
Mass ow rate of N2 entering the cathode
Mass ow rate of N2 leaving the cathode
Mass ow rate of O2 entering the cathode
Mass ow rate of O2 leaving the cathode
Cathode oxygen mass ow reference
Oxygen ow consumed in the reaction
Manifold outlet mass ow
Mass ow rate of vapour entering the anode
Mass ow rate of vapour leaving the anode
Rate of vapour generated in the fuel cell reaction
Mass ow rate of vapour entering the cathode
Mass ow rate of vapour leaving the cathode
Mass ow of vapour entering the cathode humidier
Mass ow of vapour injected by the cathode humidier
Mass ow rate of water transfer across the membrane
LQR system state
Oxygen mass mole fraction of the air entering the cathode
channels
Oxygen mass mole fraction of the air exiting the cathode
channels
Charge transfer coefcient
Polarisation curve coefcient related to liquid water
accumulation in the cathode
Polarisation curve coefcient related to liquid water
accumulation in the anode
Molar fraction of oxygen in the air
Estimate of the linearised LQR system state
Linearised LQR system state
Compressor efciency
Compressor DC motor mechanical efciency
Super-Twisting controller parameters
Ratio of the specic heats of air
Bounds of s
Membrane water content on the cathode
Membrane water content on the anode
Membrane water content
Oxygen excess ratio or oxygen stoichiometry
Ancillary input of the expanded Relative Degree 1 system
Rate of vapour in the ambient air ow
Humidity ratio of the gas entering the cathode channels
Humidity ratio of the ambient air

xx

ca,out
an
an,out
cp
(t, x, u), (t, x, u)
m,dry

Notation and Acronyms

Humidity ratio gas exiting the cathode channels


Humidity ratio of the gas entering the anode channels
Humidity ratio of the gas exiting the anode channels
Compressor rotational speed
Functions of s
Membrane dry density

Chapter 1

Introducing Fuel Cells

1.1 An Outlook of the Energy Perspective


Increasing demands on pollution reduction is driving innovation on clean energy
sources. Among these, fuel cells (FCs) are regarded as one of the most promising
technologies, due to their efciency, compactness and reliability [3]. FCs are electrochemical devices that generate electrical current from hydrogen and oxygen, with
pure water and heat as by-products. Considering that hydrogen is widely available
and can be obtained from many renewable sources using solar and wind energy,
fuel cells represent an attractive, feasible alternative to reduce fossil fuel dependence. However, the widespread use of hydrogen as combustibleand the resulting
hydrogen economydespite its interesting possibilities, has some technological
issues to be resolved.
Approximately 70 percent of the electrical power we use today is generated
from fossil combustibles. Besides, almost all the transportation vehicles have internal combustion engines. Although in a complete hydrocarbon combustion the
only by-products are carbon dioxide and water, in practice other substances like
carbon monoxide (a poisonous gas), nitrogen oxides (responsible of most of the urban smog) and unburned hydrocarbons are also produced. Greenhouse gases like
carbon monoxide are responsible for increasing temperatures of the planet, which
is directly related to climate changes. The amount of greenhouse gases in the atmosphere is continuously increasing since the Industrial Revolution, and the nal ecological consequences are still hard to foresee. For this reason, most industrialised
countries are raising their pollution-reduction standards, in an attempt to slow down
the consequences of climate change. However, this is not the only problem with
fossil combustibles. There are also pollution hazards related with their transportation (oil spills, pipeline explosion and well res), as well as economic and political
concerns regarding dependence on oil-rich countries. A switch to hydrogen as the
main energy carrier could be a practical solution to some of these problems if it can
be produced in a clean and sustainable way. Molecular H2 has the highest energy
content per unit weight among the known gaseous fuels (143 MJ kg1 ) and is the
only carbon-free fuel which ultimately oxidises to water as a combustion product.
C. Kunusch et al., Sliding-Mode Control of PEM Fuel Cells,
Advances in Industrial Control,
DOI 10.1007/978-1-4471-2431-3_1, Springer-Verlag London Limited 2012

Introducing Fuel Cells

Most of the current production of hydrogen is used in the fertiliser (50%) and
petroleum (37%) industries. Industrial production of hydrogen is currently accomplished in four main ways:
Thermo-chemical technologies: Steam reforming is the most widely used process to produce hydrogen from natural gas, coal, methanol, ethanol and gasoline,
among others [4]. The process uses catalysts in one or more steps of the conversion process. Steam reforming of natural gas accounts for almost 50% of the
world feedstock for H2 production. Obviously, this technology does not avoid
either pollution (as one of the by-products is carbon monoxide) or fossil oil dependence (hydrocarbon is still required). In spite of these facts, the process is seen
by many as an intermediate step in the right direction. For instance, several car
manufacturers are currently testing prototypes using a reformer and a fuel cell to
drive an electric engine. This allows using the existing fuel distribution network
while the alternative fuels infrastructure is developed.
Photo-biological technology: This approach uses the natural photosynthetic activity of bacteria and green algae to produce H2 . A wide range of alternatives
are available, including direct and indirect biophotolysis, photofermentation and
dark-fermentation [2]. One major limitation of this promising technology is the
relatively slow production rates, which precludes its use in massive production.
Photo-electrochemical technology: This process produces H2 in one step, splitting water by illuminating a water-immersed semiconductor with sunlight [5].
This technology is still in early stages of development, and many obstacles must
be overcome before being economically viable.
Electrochemical technologies: Electrolysis is a simple process that only requires
water and electricity. It can be accomplished securely almost anywhere, even at
home. The combination of electrolysis and a fuel cell looks like a virtuous cycle: hydrogen is produced using water and electricity, while electricity and water
are obtained in a fuel cell combining hydrogen and oxygen (or air), without any
combustion or Carnot cycle involved. However, electrolysis is energy-intensive: it
takes approximately 50 kilowatt-hours of electricity per kilogramme of hydrogen
produced. As it was already said, most of the electricity is generated using fossil combustibles, and widespread use of electrolyses will increase electricity demands. For instance, if all transportation becomes electrical, electricity demands
to the power grid will double.
In the light of the above considerations it becomes clear that, in order to put hydrogen economy to work, H2 should be produced using non-pollutant energy sources
(e.g. solar or wind). However, renewable energy sources are often intermittent and
difcult to predict; therefore it is usually difcult to match the energy production
and the energy demand. This naturally introduces fuel cells (FC) in the energy conversion chain.

1.2 What is a Fuel Cell?

Fig. 1.1 Basic H2 /O2 fuel


cell

1.2 What is a Fuel Cell?


A fuel cell is an electrochemical device that produces electricity from hydrogen and
oxygen. To be specic, the catalytic oxidation of hydrogen in an anode and the catalytic reduction of oxygen in the cathode provide a voltage drop between electrodes
that can be used by an external circuit if the electrolyte isolates both electrodes but
allows ions mass and charge exchange (Fig. 1.1). Besides the energy released as
electricity, water and heat are also produced. Given that the electrochemical process does not involve a Carnot cycle, efciency can be very high. In applications
designed to capture and utilise the systems waste heat (co-generation), fuel use efciencies could reach 8085%. Fuel cells have no moving parts, make no noise and
are scalable.
The basic chemical equations in a H2 /O2 fuel cell are the following:
O2

2H2 4H+ + 4e
+ 4e H2 O
2H2 + O2 H2 O

+ 4H+

Anode half reaction


Cathode half reaction
Overall reaction

(1.1)

Although a fuel cell is similar to a typical battery in many ways, it differs in several aspects. The battery is an energy storage device where all the energy available is
stored within the battery itself. It will cease to produce electrical energy (discharge)
when the reactants are consumed. A fuel cell, on the other hand, is an energy conversion device to which fuel and oxidant are supplied continuously. It can produce
electric current for as long as fuel is supplied.

1.3 State-of-the-Art on Fuel Cell Systems Technology


Most fuel cell systems can be characterised according to three main aspects:

Introducing Fuel Cells

Cell units, the individual devices where electrochemical reactions take place.
Cell stacks, the combination of individual cell units to provide the desired voltage
and current.
Auxiliary components necessary to ensure fuel feeding, feed-stream conditioning,
stack temperature and humidity control, electrical output conditioning, etc. These
ancillary components usually have important dynamics that affect the overall system behaviour.

1.3.1 Fuel Cell Units


During its evolution, many fuel cell unit congurations were proposed, considering different electrolyte and electrode materials, catalysts and operation temperatures. Presently, there are six major fuel cell types: Alkaline fuel cell (AFC), Molten
Carbonate fuel cell (MCFC), Phosphoric Acid fuel cell (PAFC), Proton Exchange
Membrane fuel cell (PEMFC), Solid Oxide fuel cell (SOFC) and Direct Methanol
fuel cell (DMFC) [3].
Alkaline fuel cells were one of the rst technologies developed. They were used
by NASA for electricity and potable water generation in the manned missions
Gemini, Apollo and the Space Shuttle. AFCs use compressed hydrogen and oxygen as fuel and produce potable water as a by-product. The electrolyte used is
potassium hydroxide, which is relatively expensive and is susceptible to poisoning by carbon dioxide. The efciency can reach 70 percent. Operating temperatures of the early AFC were in the range of 100 C and 250 C, but newer designs
work at much lower temperatures (23 C to 70 C).
Molten Carbonate fuel cells: This is a promising technology that uses a molten
carbonate-salt-impregnated ceramic matrix as the electrolyte. They are used
mostly with natural gas and coal in industry applications. MCFCs operate at relatively high temperatures (650 C) and can reach efciencies of 85% when waste
heat is used. Due to the high operation temperature, no external hydrocarbon reforming is necessary. Fuel is converted to hydrogen within the fuel cell itself by
a process called internal reforming.
Phosphoric Acid fuel cell: PAFCs use liquid phosphoric acid as electrolyte. Acid
is contained in a Teon-bonded silicon carbide matrix. Electrodes are made of
porous carbon with a platinum catalyst. This is the rst commercial technology developed for stationary generation. Operating temperatures are in the 150
250 C range. Efciency can be as high as 85% in co-generation.
Proton Exchange Membrane fuel cells: Also called Polymer Electrolyte fuel cells
(same acronym), are characterised by high-power density and low weight and
volume, which make them a natural choice for mobile and portable applications.
PEM fuel cells use a solid polymer electrolyte and porous carbon electrodes containing a platinum catalyst. They need only hydrogen, oxygen from the air, and
water to operate and do not require corrosive uids like other fuel cell technologies. PEM cells are typically fuelled with pure hydrogen supplied from storage

1.3 State-of-the-Art on Fuel Cell Systems Technology

Fig. 1.2 Planar-bipolar


stacking

tanks or on-board reformers, and operate at relatively low temperatures, around


80 C.
Solid Oxide fuel cell: SOFCs use a hard, non-porous ceramic compound as the
electrolyte. Because the electrolyte is a solid, the cells do not have to be constructed in the plate-like conguration typical of other fuel cell types. SOFCs are
expected to have efciencies around 5060% when converting fuel to electricity
and 8085% in co-generation. Its operation at high temperatures, around 1000 C,
allows SOFCs to reform fuels internally, which enables the use of a wide variety
of fuels and reduces the cost associated with adding a reformer to the system.
Direct Methanol fuel cell: DMFCs are powered by pure methanol, which is mixed
with steam and fed directly to the anode. They do not have many of the fuel storage problems typical of other technologies, because methanol has a higher energy
density than hydrogen but less than gasoline or diesel fuel. It can be produced
easily from biomass (corn, sugar-cane, agricultural waste, etc.). Methanol is also
easier to transport and supply to consumers using the existing infrastructure because it is a liquid, like gasoline. It is a relatively new technology compared with
that of fuel cells powered by pure hydrogen. Operational temperatures are in the
90120 C range.

1.3.2 Fuel Cell Stacks


Individual fuel cell units are combined as modules in series or parallel congurations to provide desired voltage and output power. The mechanical arrangement
must ensure not only electrical contact among units, but also adequate circulation
of gases, allowing catalyst reactions to take place at the correct temperatures and
humidity levels. For at fuel cell units like PEM fuel cells, planar-bipolar stacking
is the most usual connection (Fig. 1.2). The interconnect is a separator plate, which
usually includes channels to distribute gases evenly, separating fuel and oxidant
ows of adjacent cells.
Specially for high-temperature cells (for instance, Solid Oxide fuel cells), tubular
cell stacks are usual (Fig. 1.3). They have some structural mechanical advantages,
although ensuring short paths in the electrical interconnect can be difcult. Flattened
tubes are sometimes used to reduce packing space.

Introducing Fuel Cells

Fig. 1.3 Tubular stacking,


tangential current ow. Other
congurations are possible

1.3.3 Auxiliary Components


Fuel cells require a number of ancillary subsystems to ensure proper operation.
Among these, the following are present in almost all FC technologies [3]:
Fuel preparation. As it was already mentioned, not all FC types need pure hydrogen as the combustible. Hydrogen can be obtained from almost any hydrocarbon
through reforming. This can be accomplished using an external reformer or, in
high-temperature FC units, via internal reforming. Reforming allows some FC
types to switch between different kind of fuels, an important issue regarding reliability. On the other hand, certain technologies are very sensitive to hydrogen
impurities, as they can degrade catalyst performance and lifetime (for instance,
PEM fuel cells require less than a few ppm impurities to avoid electrode poisoning). In this case, additional ltering stages are mandatory.
Air supply. This involves compressors, blowers or compressed air tanks, as well
as air lters. Usually compressor dynamics are not negligible in the overall fuel
cell model.
Thermal management. Precise temperature control is required to ensure proper
operation. This includes gas and stack temperature control.
Water management. Besides water being a by-product of fuel cell operation, in
certain FC types interacting gases must be pre-humidied. For instance, in PEM
fuel cells this is necessary to avoid drying of the membrane. On the other hand,
excessive water can affect fuel cell performance, due to the reduction of the effective reaction area. The amount of water that is generated depends on the electrical
power demand and must be properly controlled for every operating condition.
Electric power conditioning equipment. This relates to the electronics required to
full electric load requirements.

1.4 A Brief History of Fuel Cells


Since the discovering of electrolysis in 1800 by Nicholson and Carslisle, the possibility of an inverse mechanism for the production of electricity from hydrogen
and oxygen was actively investigated [1]. In 1838, William Robert Grove built an

1.4 A Brief History of Fuel Cells

Fig. 1.4 Groves gas battery


(Philosophical Magazine and
Journal of Science, 1843)

Fig. 1.5 Mond and Langer


fuel cell (Transactions of the
American Electrochemical
Society, 1905)

experimental set-up using two platinum electrodes partially immersed in a sulphuric


acid container (Fig. 1.4). Each electrode had one end immersed in acid and the other
end in a sealed container of hydrogen and oxygen, respectively. Grove observed a
constant current ow between the electrodes. The containers held water as well
as the gases, and he noted an increasing water level in both of them as the current owed. Groves gas battery was the rst experimental fuel cell. However, the
electrochemical reactions that took place in the cell were still poorly known.
Until the early twentieth century, many people tried to produce an FC that could
convert coal or carbon to electricity directly, without combustion. In 1889, the
chemists Ludwig Mond and Charles Langer attempted to build a practical device using air, coal gas and platinum electrodes, but failed because not enough was known
about materials or electricity (Fig. 1.5). It was not until 1893 that Nobel laureate
Friedrich Wilhelm Ostwald provided much of the theoretical understanding of how
fuel cells operate. He experimentally determined the interconnected roles of the
various components of the fuel cell: electrodes, electrolyte, oxidising and reducing
agents, anions and cations. Ostwald explained Groves gas battery operation, relating its physical properties and chemical reactions. His pioneering work was the
basis for modern fuel cell research. In his own words, as he wrote in 1834: If we
gave a galvanic element which directly delivers electrical power from coal and oxygen (. . . ), we are facing a technical revolution that must push back the one of the
invention of the steam engine. Imagine how (. . . ) the appearance of our industrial
places will change! No more smoke, no more soot, no more steam engine, even no
more re [6].

Introducing Fuel Cells

Fig. 1.6 Francis Bacon with


a prototype fuel cell (Source:
IEEE Global History
Network)

In 1932, Francis Bacon, an engineer working at Cambridge University in England, made a number of signicant modications to Mond and Langers design. He
replaced platinum electrodes with less expensive nickel gauze and substituted the
sulphuric acid electrolyte for alkali potassium hydroxide, a less corrosive substance.
In 1959, Bacon built a 5-kW alkaline fuel cell, capable to power a welding machine
(Fig. 1.6). Later in that year, Allis-Chalmers demonstrated the rst FC powered vehicle, combining 1008 cells to power a 20 hp tractor [1].
Aerospace applications boosted the research on FC technology. In the late 1950s,
US National Aero Space Agency (NASA) needed a compact way to generate electricity for manned space missions. Nuclear generation was too dangerous, conventional batteries were too heavy, and solar power was too expensive and complex at
the time. Fuel cells seemed to be an interesting alternative. Therefore, NASA went
on to fund 200 research contracts for FC technology. One of the principal achievements of this initiative was the development of Proton Exchange Membrane fuel
cells (PEMFCs). Indeed, in 1958 General Electric (GE) scientists Willard Thomas
Grubb and Leonard Niedrach devised a way to deposit platinum electrolyte on a
sulphonated polystyrene ion-exchange membrane. GE and NASA further developed
this technology for the Gemini space project (Fig. 1.7). This was the rst commercial use of a PEM fuel cell [7].
In an attempt to reduce FC weight and increase its operational lifetime, Pratt and
Whitney (an aircraft engine manufacturer) licensed Bacon patents for the Alkaline
fuel cell in the early 1960s. They made a number of modications to Bacons design,
which eventually led them to win a contract to provide FC technologies for the
Apollo missions, providing electric power and also potable water for the crew in
a compact assembly. An Apollo spacecraft carried three fuel cells in the service
module. Each unit housed 31 individual fuel cells connected in series, operating at
27 to 31 volts for a typical power output of 563 to 1420 watts, with a maximum of
2300 watts. Each unit measured 105 cm in height, 55 cm in diameter, and weighted
112 kg. Since then, similar cells have been used in most of the NASA manned
missions, including the Space Shuttle.

1.5 Closed-Loop Operation of Fuel Cells. Why?

Fig. 1.7 Gemini spaceship


PEM fuel cell assembly
(Source: NASA)

During the last decades, there has been important research on new materials
for electrolytes, electrodes and catalysts, in order to reduce costs and improve efciency and reliability. Commercial applications include stationary systems for energy backup in hospitals and schools, portable devices and vehicles. The rst bus
powered by a fuel cell was completed in 1993. Daimler Benz and Toyota launched
prototype fuel-cell powered cars in 1997. Nowadays, most of the major automotive
companies have prototypes of fuel-cell powered cars and utility vehicles.

1.5 Closed-Loop Operation of Fuel Cells. Why?


In spite of current advances in fuel-cell-based technologies, their relatively high
costs, moderate reliability, and reduced lifetime remain as major limitations. For this
reason, together with the continuous improvement of materials and components, the
incorporation of advanced control strategies embodies a major technological issue,
in order to achieve cost reduction, performance improvement and efciency optimisation. The design of control systems must be understood as a whole, taking into
account sensing devices, actuators and local control schemes for each subsystem, as
well as supervisory and fault tolerant strategies, for optimal energy management in
each operating condition.
By their very nature, fuel cell systems are complex devices. As it has been analysed in former sections, they are made of many interconnected subunits, comprising interdependent electrochemical, chemical and thermal phenomena. For proper
operation, precise pressure, temperature and humidity levels are required. Closedloop operation is therefore necessary to fully exploit fuel cells potential. Comburent

10

Introducing Fuel Cells

ows, moisture levels and temperatures must be controlled to full load demands,
while ensuring safe operation and avoiding internal components damage. Many factors make this a challenge.
Firstly, fuel cells are highly nonlinear devices. This precludes the use of conventional linear control techniques, especially if performance under wide operation
ranges is required. Besides, important internal variables are difcult to access, making their measurement expensive or cumbersome. Finally, there exist uncertainty
in system parameters, modelling errors and disturbances of many types. Reliable
control systems ensuring stability and performance, as well as robustness to model
uncertainties and external perturbations are of crucial importance for the advance of
fuel cell technology.

1.6 Scope and Outline of the Book


This book addresses the analysis and design of closed-loop controllers for fuel cell
systems. In particular, robust controllers based on the so-called sliding mode
paradigm are considered for the case of PEM FC systems. It resumes the current
state-of-the-art and can be used as a starting point for further research and development.
The text is conceptually divided in two parts. The rst section (Chaps. 2 and 3) is
devoted to theory of fuel cells and Sliding-Mode (SM) controllers. Fuel cell operation principles are discussed, in particular for the case of PEMFCs. A survey of the
dynamic models in the literature is presented, with special emphasis on their suitability for control design. The mathematical background relative to Sliding-Mode
controllers, and in particular Second-Order Sliding-Mode (SOSM) Controllers is
summarised, explaining design issues and discussing performance and dynamic response features. SOSM controllers allow reduction of chattering problems typically
associated with standard SM techniques.
The second part (Chaps. 4 to 6) presents simulation and experimental results
PEM fuel-cell-based systems. A modelling methodology is discussed for the related subsystems: compressor-based air supply, gases humidiers and line heaters,
fuel cell stack thermodynamics and electro chemistry. The proposed framework can
be used as a guide for similar FC congurations. A detailed design procedure for
different SOSM control strategies is outlined, and their results compared in simulations and experimentally. Performance and robustness issues are discussed in depth.
Finally, conclusions and open research lines are presented.
The intended audience for this book comprises control engineers, FC systems
developers and postgraduate students of Electronics and Control Engineering.

References
1. Collecting the history of fuel cells. Smithsonian National Museum of American History

References

11

2. Das D, Veziroglu TN (2001) Hydrogen production by biological processes: a survey of literature. Int J Hydrog Energy 26:1328
3. EG&G Technical Services, Inc (2004) Fuel cell handbook, 7th edn. U.S. Department of EnergyOfce of Fossil Energy, Morgantown
4. Haryanto A, Sandum F, Murali N, Adhikari S (2005) Current status of hydrogen production techniques by steam reforming of ethanol: a review. Energy Fuels 19:20982106.
http://americanhistory.si.edu/fuelcells/index.htm. Cited 4 Feb 2011
5. Kotay SM, Das D (2008) Bio-hydrogen as a renewable energy resourceprospects and potentials. Int J Hydrog Energy 33:258263
6. Kunze J, Stimming U (2009) Electrochemical versus heat-engine energy: a tribute to Wilhelm
Ostwalds visionary statements. Angew Chem Int Ed 48:92309237
7. Scott JH (2006) The development of fuel cell technology for electric power generation: from
NASAs manned space program to the hydrogen economy. Proc IEEE 94:18151825

Chapter 2

PEM Fuel Cell Systems

2.1 An Introduction to PEM Technology


Among the many different technologies summarised in Chap. 1, Proton Exchange
Membrane (PEM) fuel cells are extensively used for mobile and portable applications. This is due to their compactness, low weight, high power density and clean,
pollutant free operation. From the operational point of view, a relevant aspect is
their low temperature of operation (typically 6080 C), which allows fast starting
times. In a PEM Fuel Cell, a hydrogen-rich fuel is injected by the anode, and an
oxidant (usually pure oxygen or air) is fed through the cathode. Both electrodes are
separated by a solid electrolyte that allows ionic conduction and avoids electrons
circulation. Catalytic oxidation of H2 and catalytic reduction of O2 take place in the
negative and positive electrodes, respectively.
The standard electrolyte used in PEM Fuel Cells is a peruored solid polymer
composed by Teon-like chains. This material combines mechanical, chemical and
thermal stability with a high protonic conductivity, when properly humidied. Electrodes are typically made of a porous carbon compound coated with a catalyst such
as platinum or palladium, to improve the efciency of electrochemical reactions.
Catalysts are essential in this technology and constitute one of the most expensive
components of the cell. Besides, they are very sensitive to CO contamination, which
makes the use of high-purity hydrogen (CO  20 ppm) mandatory. This is a serious
limitation when H2 is obtained from hydrocarbon reforming. Alternative alloys, like
platinum/ruthenium, which are more resistant to CO poisoning, are currently under
development.
The output of a PEM Fuel Cell is electric energy, with water and heat as the
only by-products. Efciency can be high, as previously said, due to the absence of
a Carnot cycle. From the electrical point of view, the cell can be seen as a voltage source where the output impedance presents a highly nonlinear dependence to
operating conditions such as temperature, electric current, partial pressures and humidity levels of the incoming gases. Due to this nonlinear, multi-variable dependent
behaviour, precisely controlled conditions must be ensured for proper operation.
C. Kunusch et al., Sliding-Mode Control of PEM Fuel Cells,
Advances in Industrial Control,
DOI 10.1007/978-1-4471-2431-3_2, Springer-Verlag London Limited 2012

13

14

PEM Fuel Cell Systems

2.2 Basics of PEM Fuel Cells Operation


Catalytic reactions of hydrogen oxidation at the anode and oxygen reduction at the
cathode produce an electric potential difference between electrodes, that can be used
in an external circuit if the electrolyte allows ionic mass transport but isolates electrically both electrodes. On the anode side, the catalyst produces the dissociation of
hydrogen molecules into protons (H+ ) and electrons (e ). Protons cross the polymeric membrane, while electrons are forced to the external electric network. In the
cathodes surface the oxygen molecules react with electrons from the external circuit and protons from the membrane to produce water. In the process, the only byproduct is water, in vapour and liquid phases. The membrane must be properly humidied, because its protonic conductivity depends directly on its water content. To
accomplish this, input gases are previously humidied. The amount of energy produced in the electrochemical process can be calculated from changes on the Gibbs
free energy (gf ), that is, the difference between the Gibbs free energy of products
and reactants. In the particular case of PEM Cells fuelled with pure hydrogen, the
product is distilled water (H2 O), the reactants are hydrogen and oxygen, and then
gf = (gf )prod (gf )react = (gf )H2 O (gf )H2 (gf )O2

(2.1)

The Gibbs free energy represents the energy available for external work. The
values of gf depend on the reactants temperatures and pressures according to the
following expression:

1/2 
PH2 PO2
gf = gfo RT fc ln
PH2 O

(2.2)

where R is the universal constant for ideal gases, PH2 is the hydrogen partial pressure, PO2 is the oxygen partial pressure, PH2 O the water vapour partial pressure, and
gfo the change in the process gf at a standard working pressure (1 bar), which
in turn changes with the temperature of the fuel cell (Tfc ). Values of the Gibbs free
energy for standard pressure at different temperatures are shown in Table 2.1 [17].
A negative gfo implies that the reaction releases energy ((gf )reac > (gf )prod ).
If the electrochemical processes taking place in the cell were reversible, all the
Gibbs free energy could be converted into electrical energy for the external circuit. In this ideal case, for each mol of hydrogen, two moles of electrons circulate by the electric circuit making an electric work (charge voltage) 2F E,
where F is the Faradays constant or, equivalently the charge of an electron mol
(96485.309 C/mol), and E is the open circuit fuel cell voltage. This amount of electrical work is the net change in the Gibbs free energy:
gf = 2F E

(2.3)

Therefore, the reversible voltage of a PEM cell is expressed as


E=


1/2 
gfo
PH2 PO2
RT fc
gf
=
+
ln
2F
2F
2F
PH2 O

(2.4)

2.2 Basics of PEM Fuel Cells Operation


Table 2.1 Changes in gf
for a standard 1 bar pressure

Water phase

15
Temperature C

gfo (kJ/mol)

liquid

25

273.2

liquid

80

228.2

gaseous

80

226.1

gaseous

100

225.2

gaseous

200

220.4

gaseous

400

210.3

gaseous

600

199.6

gaseous

800

188.6

gaseous

1000

177.4

Expression 2.4 is the so called Nernst voltage of a PEM fuel cell. In practice, however, the open circuit voltage is smaller than what Eq. (2.4) predicts. In fact, the
term gfo /2F varies with temperature,and differs from its value E0 = 1.229 V at
standard conditions (25 C, 1 atm) according with the following expression:


gfo
S o

= 1.229 + (Tfc To )
(2.5)
2F
2F
where To is the standard temperature of reference (298 K), and S o is the entropy
change for the new operating conditions. Therefore, the last equation can be rewritten as


gfo
S o
298S o

= 1.229
+
(2.6)
Tfc
2F
2F
2F
Using the standard thermodynamical relations regarding entropy changes [17],
Eq. (2.4) can be written as


E = 1.229 0.85 103 (Tfc 298) + 4.3 105 Tfc ln(PH2 ) + 1/2 ln(PO2 )
(2.7)
Additionally, the cell voltage varies with electric load conditions. This is due to
electric losses, which can be classied as activation, ohmic and concentration or
diffusion losses.
Activation losses are important at low currents and reect the fact that the cell
requires a certain amount of energy to start electron circulation and create/break
chemical bondings, both in the anode and the cathode [19]. This produces an
important voltage drop at low current densities in both electrodes. It is worth
noting that hydrogen oxidation at the anode is considerably faster than oxygen
reduction at the cathode. Therefore, the dynamics of activation losses are always
dominated by the cathode. The relationship between activation losses and current
density can be described using the Tafel equation [18]
 
i
Vact = A ln
(2.8)
i0

16

PEM Fuel Cell Systems

where the constant A is higher for slow electrochemical reactions, and i0 is higher
for fast reactions. The value of i0 can be considered as the current density from
which the voltage drop becomes evident. It is called exchange current density and
its typical values are in the 102 108 A range. The Tafel equation is only valid
for i > i0 . For a pure hydrogen fuel cell, A is given by
A=

RT
2F

(2.9)

The constant is known as the charge transfer coefcient and represents the
amount of electric energy applied that is harnessed in changing the rate of an electrochemical reaction. Its value depends on the reaction involved and the material
of the electrode, and falls in the 01 range.
Ohmic losses are due to the resistance of the polymeric membrane to proton circulation, and also to the electrical resistance of electrodes and current collectors.
Therefore, these losses are proportional to electric current in a wide operational
range:
Vohm = i Rohm

(2.10)

The value of Rohm represents the internal resistance of the cell and has a strong
dependency with the cell humidity and temperature levels. It depends on the
membrane conductivity (m ) and dry thickness (tm) according to the following
expression [26, 30]:
Rohm =

tm
m

(2.11)

For standard operating conditions, the following empirical expression is frequently used [30]:
Vohm = (R0 R1 m )i

(2.12)

where m is the membrane water content, dened as the number of water


molecules per sulfonate group in the ionomer, and R0 and R1 are values to be
experimentally determined.
Diffusion losses are the result of changes in the concentration of reactants as
they are consumed by the electrochemical reaction. This effect is responsible of
an important voltage drop at high current densities. The non-uniform conditions
arising at the porous electrodes discourage this as a desirable operation zone.
A semi-empirical expression for this effect is [15]


PO2
(2.13)
Vconc = me(ni) + b ln
a
where m and n are empirical coefcients with typical values close to 3 105 V
and 8 cm2 /A, respectively [17]. The last term is included to take into account
concentration losses due to low oxygen stoichiometry values.

2.3 Efciency and Power Conversion

17

Fig. 2.1 Polarisation curve of a PEM fuel cell

Therefore, taking into account all the losses, the cell voltage can be written as
Vfc = E Vact Vohm Vconc

(2.14)

Replacing the values of the individual terms, the expression for (2.14) is

 

PO2
i
RT
ln
Vfc = E
(R0 R1 m )i me(ni) + b ln
(2.15)
2F
i0
a
where , i0 , R0 , R1 , m, n, b and a are empirical parameters that take into account the
different polarisation effects and are adjusted for a specic fuel cell stack, without
loss of generality. A systematic procedure for the determination of these constants
is outlined in Chap. 5. The resulting polarisation curve of a typical PEM fuel cell
is shown in Fig. 2.1. It can be seen that the open circuit voltage is close to 1 V. In
applications requiring higher voltage and power, several cells can be combined in
series/parallel congurations to full load demands.

2.3 Efciency and Power Conversion


The efciency of any energy conversion device is dened as the ratio between output
and input useful energy. In the PEM fuel cell case, the available energy at the input
of the device is the hydrogens enthalpy (measured as the amount of heat that can
be converted to work). If all the Gibbs free energy at the cell output is converted to
electric energy, the efciency results in
max =

gf
237.2
=
= 0.83
H
286

(2.16)

18

PEM Fuel Cell Systems

This is the maximum theoretical efciency that a PEM fuel cell can reach at
25 C. Using Faradays constant and considering that there are two electrons involved in the reaction, the efciency of a PEM fuel cell can be expressed as a quotient of voltages:
max =

gf /2F
gf
=
= 1.229/1.482 = 0.83
H
H /2F

(2.17)

where gf /2F = 1.229 V is the theoretical voltage of an open circuit cell, and
H /2F = 1.482 V is the value of the thermoneutral voltage (that is, the resulting
voltage if all the enthalpy of hydrogen is converted in electric energy). In this way,
the efciency of the cell at any condition can be obtained from the voltage at its
output terminals (Vfc ):
=

Vfc
1.482

(2.18)

2.4 State-of-the-Art in PEM Fuel Cells Technology


Current research efforts in PEM technology are mainly oriented to three basic directions: new components and materials, modelling of cell dynamics, and control.
Materials: evolution of new components and devices, capable of efcient operation under wider ranges of temperature, humidity and gas purity is required
to broaden the spectrum of applications currently devised for PEM fuel cells.
This includes aspects related to fabrication and operation of membranes and electrodes, manufacturing processes, design and characterisation of components. Research in membranes is oriented to thermostable polymers (polyetheretherketone,
polysulphone, etc.) and composite membranes capable to operate at temperatures
above 100 C and lower humidity levels than the actual commercial membranes
[31]. Regarding electrodes, most efforts are centred on reducing the amount of
platinum required in catalysts and improving gas diffusion layers. Advances in
the production, transport and storage technologies are also necessary, to make
hydrogen an economically viable alternative [15].
Cell Dynamics Modelling: better understanding of the processes involved in fuel
cell operation, both at the membrane level and the auxiliary subsystems (compressors, line heaters, humidiers, etc.), is helping engineers to develop dynamic
models suitable for the design of more reliable, compact and efcient devices.
This is also widening the spectrum of applications of PEM fuel cells, making
them attractive alternatives in elds traditionally reserved to other types of cells,
for instance, in stationary, high-power installations. On the other hand, accurate
models of gas distribution and uid dynamics inside cells are improving the predictability of electrochemical processes, with direct impact on design strategies.
From the automatic control perspective, efforts are conducted to the development of robust, nonlinear strategies capable to improve efciency and reliability
of PEM cells, avoiding permanent damage to membranes. Robustness is required

2.5 Components and Associated Devices

19

to take into account the intrinsic uncertainty in the system model and to avoid
expensive or cumbersome (even impossible) measurement of certain variables.
Additionally, nonlinear control strategies usually allow wider operation ranges
than local approaches that linearise the system around equilibrium points. For
commercial success, it is also important to develop models and strategies to diagnose and predict failure situations.

2.5 Components and Associated Devices


2.5.1 Polymeric Membranes
Membranes in a PEM fuel cell must have a relatively high protonic conductivity and also act as suitable mechanical barriers to avoid mixing of comburents
and reacting gases. Additionally, they must be chemically stable for the entire operation range. Standard membranes are made of peruorosulfonic acid (PFSA).
This material is, essentially, a copolymer of tetrauoroethylene (TFE) and several sulfonated peruor monomers. The most popular commercial membrane is
Naon , made by Dupont, which uses peruoro-2-(2-uorosulfonylethoxy) Propyl
Vinyl Ether (PSEPVE). Similar materials are produced by several companies, such
as Asahi Glass (Flemion ), Asahi Chemical (Aciplex ), Chlorine Engineers (C
Membrane ), etc. Dow Chemical has developed a composite membrane called
GoreSelect , using a Teon-like material.
Important features of membranes suited for PEM fuel cells are protonic conductivity, water transport properties, gas permeability, mechanical resistance and
dimensional stability. These parameters strongly depend on the membrane water
content. The water content is usually expressed as the weight rate of water and dry
polymer or, alternatively, the rate between the number of water molecules per sulfonic groups present in the polymer. The maximum amount of water in a given
membrane depends heavily on its previous preparation [11].
Regarding its critical role on protonic conduction, it is important to keep a proper
membrane water content at all possible operating conditions. Several mechanisms
affecting water transport are present in a PEM fuel cell, and their combined effects
determine the amount of water present. Among them, the most relevant are the following:
Water generation on the cathode side, at a rate proportional to the electric current
produced.
Electro-osmotic drag, produced by water molecules dragged by the proton ow
from anode to cathode.
Diffusion, due to the water concentration gradient across the membrane.
Water permeability, due to the pressure difference between anode and cathode
channels.
In thin membranes, diffusion can be compensated by the drying effect produced
by electro-osmotic drag at the anode. In thick membranes this drying effect can be

20

PEM Fuel Cell Systems

Fig. 2.2 PEM fuel cell


Membrane and cathode.
Effective area is 50 cm2 . IRI
(CSIC-UPC)

more important, especially at high current densities. An ideal electrolyte should be


impermeable to reactant gases, to avoid gas mixing inside the cell. However, due
to its porous structure, water content, and the solubility of hydrogen and oxygen
in water, small amounts of reactant gases can pass through the membrane. It is
worth noting that hydrogen has a permeability one order of magnitude higher than
oxygen [4].

2.5.2 Electrodes
Electrodes in PEM fuel cells are basically catalytic layers placed between the polymeric membrane and an electrically conducting substrate. This is where the electrochemical reactions take place. Given that gases, electrons and protons react in the
surface of electrodes, their efciency can be improved increasing its rugosity (effective area), reducing catalyst particle size and/or incorporating ionomeric material
(a one-ion polymer) in the catalyst. The later can be accomplished by painting the
electrode with a PFSA solution in a mix of alcohol and water, or simply mixing the
ionomer and catalyst in the preparation of the catalytic layer. A 30 percent ionomer
content in the catalytic layer is typical.
The catalyst most usually employed in both electrodes of PEM fuel cells is platinum. In the early stages of PEM fuel cells development, important quantities of
platinum were used (more than 28 mg/cm2 ). By the end of the 1990s this amount
was reduced to 0.30.4 mg/cm2 . Considering that the effective area of catalyst is of
paramount importance, it is crucial to achieve a ne dispersion of catalyst particles
in the support material, usually carbon powder [29].
The combination of electrodes and polymeric membrane is known as Membraneelectrode assembly (MEA). There are basically two different approaches for its con-

2.5 Components and Associated Devices

21

Fig. 2.3 Toray paper diffusion layers. IRI (CSIC-UPC)

struction. In the rst one, the catalyst is deposited on a porous substrate called gas
diffusion layer, which is typically a carbon bre paper. These are then placed at
both sides of the membrane using heat and pressure to ensure proper contact. In the
second approach, the catalyst is deposited directly on the membrane, an arrangement known as catalysed membrane. A porous substrate is then added, resulting in
a ve-layer MEA (Fig. 2.2). Several techniques can be used for deposition of catalyst on diffusion layers and membranes (spreading, spraying, sputtering, painting,
screen printing, decaling, electro-deposition, evaporative deposition, impregnation
reduction, etc.), and many other proprietary approaches exist.

2.5.3 Gas Diffusion Layers


The main purpose of gas diffusion layers is conducting and spreading reacting gases
from bipolar plate channels to the MEA. They are typically made of porous materials. Regarding desirable properties of diffusion layers, the following aspects must
be considered:
Porosity must be such that the ux of reactants and water is efcient. Note that
both ows are in opposite directions.
Electrical and thermal conductivity must be high. The contact resistance or interface is typically dominant versus the volume conductivity.
Given that the catalyst is a discrete material (small particles), diffusion layer pore
size cannot be excessively big.
They must provide proper mechanical support to the membrane. However, some
degree of exibility is required to provide good electrical contact.
Although conicting somewhat, these requirements are typically fullled with
carbon bre-based papers and cloths (Fig. 2.3). These materials are usually made
hydrophobic to avoid ooding in the structure. To achieve this, diffusion layers are
usually treated with polytetrauorethylene (PTFE), a polymer similar to polyethylene. To improve electrical properties, a microporous layer made of carbon or

22

PEM Fuel Cell Systems

Fig. 2.4 Sealing gasket


placed over a current
collector. IRI (CSIC-UPC)

graphite particles mixed with PTFE is added, resulting on pore sizes of 0.10.5 m,
which are much smaller than the carbon bre pores (2050 m) [21].

2.5.4 Sealing Gaskets


These components provide mechanical sealing among bipolar plates and diffusion
layers. Their purpose is twofold: on one hand, minimising leaking of gases to the
exterior of the cells and, secondly, avoiding the mixing of reactant gases near the catalyst areas (Fig. 2.4). However, it is worth noting that a certain amount of hydrogen
is expected to pass through the membrane by diffusion. This can be computed from
the number of cells, their width and type, effective area, partial pressures and working temperature. Sealing gaskets avoid direct combination of reactants, which in turn
can produce irreversible damages to cell components [12]. Most sealing gaskets are
made of silicone, neoprene or plastic polymers, with additional bre materials to
improve their mechanical properties.

2.5.5 Bipolar Plates


Bipolar plates, also called collectors or separators, have many functions in a fuel
cell system [4]. Among others, the most relevant are the following:
Electrical connection between individual cells of the stack.
Separation of gases among adjacent cells. Thus in PEM fuel cells they must be
impermeable to H2 , O2 and N2 .

2.5 Components and Associated Devices

23

Fig. 2.5 Graphite bipolar


plate. IRI (CSIC-UPC)

Structural support to the cell. Robustness and light weight are mandatory.
Efcient heat conduction.
In addition, bipolar plates must be corrosion resistant, which make them expensive. In most PEM cells pH is between 2 and 3, and temperatures are in the
6080 C range. These environmental conditions discourage the use of traditional
choices like aluminum, steel, titanium or nickel. Corrosion produces ions of metal
which can diffuse through the membrane, affecting its ionic conductivity, reducing
its lifetime and increasing its electrical resistance. For this reason, metallic plates are
usually coated with non-metallic conductive materials, such as graphite, diamond
carbon, conductive polymers, organic polymers, noble metals, metallic nitrides, tindoped indium, etc. Bipolar plates can also be made of thermoplastic materials, such
as polypropylene, polyethylene and polyvinyldene uoride. Thermosetting resins
(phenolic, epoxy, etc.) are also used, with the addition of graphite and bre reinforcing (Fig. 2.5).
An important property of bipolar plates is their electrical conductivity. In graphite
composites, typical values range between 50 and 200 S/cm. Although pure graphite
has a conductivity of 680 S/cm, metallic plates have values an order of magnitude
higher. Note that the overall conductivity of the cell is always lower than the material
conductivity, due to the contact resistance among components.
In Fig. 2.6, a schematic representation of a typical PEM fuel cells stack is presented, where all the components described above can be visualised.

24

PEM Fuel Cell Systems

Fig. 2.6 PEM fuel cell stack scheme [13]

2.5.6 Auxiliary Devices


2.5.6.1 Air/Oxygen and Hydrogen Supplies
In PEM fuel cell stacks, air can be supplied by means of a compressor, a fan or a
compressed air source, although the latter is mostly used only in laboratory settings.
Fans are customarily used in open-cathode designs working at ambient pressure. Air
compressors give autonomy and exibility to the system, allowing precise control
of working pressures. Usually, both compressors and fans are electrically connected
to the stack, becoming part of its losses or parasitic loads and thus reducing the
amount of energy available to external loads. This results in a signicant reduction
of the overall performance of the system as an energy conversion device.
At the time of this writing, PEM cells require a high-purity hydrogen supply
(CO < 20 ppm). Hydrogen storage for PEM cells is a matter of active research, the
main options being compressed gas, cryogenic liquids and solid metallic hydrides.

2.5.6.2 Water and Heat Management


Water management in individual PEM cells and cell assemblies is a technological
challenge. For proper operation, membranes must be totally saturated of vapour but,
in order to ensure optimum performance, excess water must be efciently removed
(especially from the cathode line). Ionic conductivity in the polymeric membrane is
directly related to its water content, which in turn affects the conversion efciency

2.5 Components and Associated Devices

25

Fig. 2.7 A PEM fuel cell


stack, showing the series
connection of individual
MEAs. IRI (CSIC-UPC)

at each operation point. The amount of liquid water and membrane humidication
levels can be modied by controlling the relative humidity of the reacting gases, as
well as their individual pressures and temperatures.
Temperature is as important as water content, and both magnitudes are closely related. PEM fuel cells are intended to operate at high power densities (>0.5 W/cm2 ),
but, apart of new high-temperature membranes, the most widespread technology
now prevents operation at temperatures above 100 C. This small gap between operational and ambient temperatures makes it difcult to remove the 1.3 W of heat
produced for each watt of electric power generated.

2.5.6.3 Electrical Conditioning


Output power of a fuel cell is not regulated, and its stability is a relevant issue. The
small voltage of each individual cell is heavily inuenced by changes in electric
current, partial gas pressures, reactants humidity level, gas speed and stoichiometry,
temperature and membrane water content.
According to Eqs. (2.6) and (2.7), the maximum voltage of a single PEM cell is
close to 1.229 V. Higher power and voltage is obtained connecting individual cells in
series/parallel congurations (Fig. 2.7). Then, an electronic conditioning system is
necessary to meet load requirements. For instance, DC/DC converters can be used
to extend the range of operating voltages or provide specic voltage values. The
combination of a DC/AC stage and a transformer can be used to ensure electric isolation between load and stack, converting DC to an intermediate AC frequency and
using a decoupling transformer and a rectier. Usually an auxiliary power source is
necessary to start the stack operation. In the special case of autonomous systems,
power converters associated with a supervisory electronic control are required to
ensure proper management of batteries or super-capacitors charge/discharge cycles.

26

PEM Fuel Cell Systems

2.5.6.4 Humidication
Proper humidication of polymeric membranes is an important matter. At moisture levels below recommended values, ionomer water content falls, reducing conductivity and adversely affecting the kinetics of reduction and oxidation reactions.
Water content of an MEA depends on many factors and is closely related with operating conditions. For instance, in open circuit and also at low current densities, the
small water production by oxygen reduction causes the MEA humidication levels
to decrease, although the reacting gases are saturated with vapour. This is because
water absorption in peruoro-sulfonated membranes is lower when the membrane
has been equilibrated with vapour instead of liquid water. Besides, at higher current
densities, the electro-osmotic dragging of water through the membrane tends to dry
the anode. An additional problem in this case is the eventual cathode ooding due
to the excess of water produced by the electrochemical reactions.
Under certain conditions, the amount of water produced by oxygen reduction
at the cathode sufces to keep proper membrane hydration. However, this fact does
not ensure an equilibrium point in water content. In most cases, a gas humidication
system is required, at least in the cathode line, to control this important performance
variable. Humidity control is a challenging task, because moisture levels in both
channels are coupled and many perturbation (known and unknown) exist. Besides,
humidication requirements can be in conict with other control objectives, such
as oxidant stoichiometry. For instance, in certain situations the optimal humidity
level could be incompatible with the most efcient reactant ux level required to
satisfy electric load demands. In such a case, a problem arises because it is difcult
to suspend instantaneously the cathode humidication if the cells start to ood. On
the other hand, it is also impossible to increase hydration levels at low currents when
the cell starts to dry.
Among the many techniques for gas humidication, the main alternatives are the
following:
Gas bubbling. This is a method mostly used in laboratory applications for relatively low ows, being seldom used in commercial devices. It involves circulating
air or hydrogen through a porous tube immersed in liquid water at a regulated
temperature. The resulting bubbles provide a relatively big contact area between
gas and water, allowing proper humidity transfer. Moisture level is controlled by
varying water temperature. In a well-designed system, emerging gases are saturated of vapour at the water temperature. The main drawback of this approach is
the presence of water droplets in the outcoming gas. This affects gas diffusion in
the MEAs, reducing the overall efciency.
Direct vapour injection. This is the most compact, efcient and easy to control
method. A ne mist is injected in the gas stream using a small pump. An additional heat source is usually required to produce complete water evaporation if
hot water enthalpy solely does not sufce. Moisture level is directly controlled by
varying the amount of vapour injected [4].

2.6 Available PEM Fuel Cell Models in the Open Literature

27

Water exchange through a permeable material. In this approach, gases are circulated on one side of a permeable membrane such as Naon . De-ionised,
temperature-controlled water circulates on the other side of the membrane. During the process, a moisture gradient is established, which allows water transfer
via diffusion through the membrane. The amount of water transfer is controlled
by varying its temperature.
Enthalpy wheel. In this passive method the water and heat content of a gas can be
transferred to other gas using a thermal process. It comprises a cylinder that interfaces two parallel conducts, where gases at different humidity and temperature
circulate in opposite directions. The cylinder is lled with a permeable material,
to provide a large interface area for energy transfer. As the wheel rotates between
the ventilation and exhaust gas streams, it takes heat energy and releases it into
the colder gas stream. The driving force behind the exchange is the thermal gradient between the opposing gas streams. The enthalpy exchange is accomplished
by the use of desiccants, which transfer moisture by adsorption. This effect is
predominately driven by the difference in the partial pressure of vapour between
the gas streams.
2.5.6.5 Gas Heating Lines
These devices are inserted in the gas path to provide temperature control, independently from the humidication process. Basically they are made of heating resistances, with protective stainless steel shielding. In PEM fuel cells, the main objective
of the heating lines is keeping gas temperature high enough to avoid condensation
inside cell channels.
The comprehensive block diagram presented in Fig. 2.8 shows a typical laboratory set up of a PEM-based generation system.

2.6 Available PEM Fuel Cell Models in the Open Literature


Accurate mathematical models of PEM cells behaviour are subject of current interest and active research. Dynamics of the electrochemical reactions that take place
in the MEAs and the ancillary devices required for operation of the cell constitute
a nonlinear, highly coupled multi-variable dynamic system. In control applications,
the interest is focused on the development of reduced-order nonlinear dynamical
models relating smooth vector elds. This is usually required in the design of many
nonlinear control strategies.
A suitable model should be capable of predicting the dynamic and stationary behaviour of the fuel cell in a wide range of operating conditions. Important variables
to be taken into account are temperature and relative humidity of interacting gases,
partial pressures (hydrogen, oxygen, nitrogen and water vapour), velocities of ows
in the MEAs channels, electric currents and voltages, etc. Although many models of
PEM fuel cells have been reported, only a few are suitable for their use in nonlinear
control design.

28

PEM Fuel Cell Systems

Fig. 2.8 PEM fuel-cell-based system scheme

2.6.1 Control Oriented Models


PEM fuel cell modelling has been studied by several recognised authors and with
different approaches [2, 20, 27, 28]. However, many of these models have not been
experimentally validated, and there is still a lack of rigorous studies on parameters
identication and their association with performance variables. On the other hand,
there are few models and methodologies specically oriented to control design. For
example, the rst models from the open literature, as the ones presented in [2] and
[30], are essentially electrochemical characterisations based on empirical relationships that do not consider gas dynamics. More recently, works such as [8, 9, 27, 28]
have presented extended equations, including gas dynamics and temperature effects
inside the cells. However, only [28] and [9] have proposed fully analytical control
oriented models. In [9], the model considers only three of the six states of a typical
air supply subsystem, the humidication phase is not included, and characterisation
of the other subsystems is only briey described. In [28] it is presented probably the
rst proposal of a PEM fuel cell stack model, fully validated and especially developed for control engineering, and it is the basis of numerous works such as [5, 14,
27]. Changes in the liquid water and oxygen concentrations, as well as temperature,
have signicant effects on the PEM fuel cell performance and may even affect its

2.6 Available PEM Fuel Cell Models in the Open Literature

29

durability. All these variables exhibit a spatial dependence along anode and cathode
channels, and therefore it is necessary to incorporate control mechanisms to keep
them within their nominal values [20].

2.6.2 Control Objectives and Challenges


PEM fuel cell systems have many advantages over traditional alternatives such as
internal combustion engines. However, a number of technical issues must be faced
to make them competitive. Among them, operational costs, lifetime and reliability
are of vital importance.
From the automatic control perspective, a PEM fuel cell is a nonlinear multipleinputmultiple-output (MIMO) dynamical system with strongly coupled internal
variables, external perturbations and parameter uncertainties. Its normal operation is
always associated with the generation and transport of liquid water, vapour and gas
mixtures, spontaneous electrochemical reactions, exothermic processes and thermal
conduction. By their very nature, they are sensitive to changes on operation conditions (power demand, partial pressures and relative humidity of reacting gases, temperatures, etc.) and also susceptible to potential damage. Three basic degradation
mechanisms can be clearly distinguished: mechanical, thermal and electrochemical. Among the mechanical processes that produce a signicant degradation, the
cycles of humidication/drying play a decisive role, as they cause membrane expansion/shrinking. This leads to mechanical stress of the membranes and gaskets.
Additionally, the thin polymeric membranes currently used make the system potentially vulnerable to abrupt pressure changes between channels, excessive temperature and low relative humidity conditions.
Thermal degradation arises when considerable temperature variations occur in
the stack, even within the range usually recommended by manufacturers of PEM
fuel cells (6080 C). These thermal cycles, which in some cases can be extreme
(e.g. those driven by cold starts and sudden high power demands), produce accumulative mechanical damages that affect the resistance contact between the membrane
and electrodes, as well as the mechanical resistance, conductivity and permeability
of the polymeric membranes. In fact, heat management is recognised as one of the
most important issues in high-power PEM cells. This is due to the fact that under
normal conditions, the cell produces as much heat energy as electricity. This means
that in an automotive 100-kW fuel cell it is necessary to provide a structure capable
of dissipating 100 kW of heat, which is particularly difcult if the operating temperature is only 80 C. This operating temperature is a restriction imposed by the
materials employed. Besides, temperature cannot be lower than 60 C, to avoid water condensation inside cell channels, which would result in a voltage drop, caused
by a reduction of gas mass transport to the membrane.
Regarding the electrochemical mechanisms of degradation, it is important to state
that the chemical reactions on the catalysts produce small amounts of peroxide radicals (HO) or hydroperoxide (HOO) that are responsible for the chemical degrada-

30

PEM Fuel Cell Systems

tion of the membrane and its catalyst. The production of such radicals is accelerated
when the fuel cells operate in open circuit or low humidity conditions.
Another issue which deserves special attention is oxygen control. If the oxygen
ow is too low, undesirable hot spots appear in the membrane, and output power
decreases because of the lack of reactants, a situation called cathode starvation. On
the other hand, if oxygen ow is too high, an excessive amount of water is pushed
to the cathode outlet, which in turn results in membrane drying, which affects its
ionic resistance. Besides, an increase of the air ow results in higher power demand
to the compressor that supplies it, reducing the overall system performance. Thus,
an efcient control system must be capable of regulating air ow properly, avoiding
irreversible damages to the membrane and delivering enough oxygen to meet the
electric power demand in a reliable and efcient way.

2.6.3 Recent Advances on PEM Fuel Cell Control


During the last years, several control proposals have been made for PEM fuel-cellbased systems. Many examples can be mentioned. For instance, in [38] fuel cell
power output is directly regulated by limiting its hydrogen feed. This is achieved
using a PID control that varies the internal resistance of the membrane-electrode assembly in a self-draining fuel cell with the efuents connected to water reservoirs.
In [35] cathode oxygen is regulated through a feedforward loop, and temperature
is controlled using a proportional control, to ensure stack performance around an
optimal operation point, where net power is maximised. In [23] a MIMO system is
considered, with hydrogen and coolant as inputs and power density and temperature as outputs. Those variables were selected from a steady-state analysis using a
relative gain array (RGA) technique. Two PID controllers were used, and simulation results suggest that the design can be accomplished from two decoupled SISO
systems. In [34] and [3] predictive control approaches are considered, allowing improvements on the response of the air supply and efciency optimisation in fuel
cell stacks. In [28] a dynamic model of the air supply subsystem of a PEM cell is
presented. Based on this model, an LQR controller was designed to decouple the
air mass ow from cathode pressure. Reported results favourably compare against
a standard PI controller. A substantial improvement was made in [37] where, from
a linear identication of a fuel cell system, a H controller was designed to regulate the cell output resistance and control output voltage, manipulating input gases
ows.
Alternatively, several proposals have been made regarding hybrid power generation congurations. In these systems, a fuel cell stack is usually combined with
different energy storage devices, to provide a more reliable power source. Ultracapacitors [33] can be included to improve power transients using, for instance, PID
controllers to regulate DC bus voltage. In [18] and [7] a hybrid system made from
a fuel cell and batteries is proposed. In this approach, fuel cells are used for low
power demands, while battery banks supply additional energy for higher power requirements, when the cell tends to reduce its output voltage. In such cases power

References

31

converters can also be used to improve efciency and extend operating ranges of the
stack. For instance, in [36] resonant soft-switching techniques are used to adjust the
output voltage of a 250-W PEM fuel cell.
Most of the above approaches are based on linear models, providing interesting
solutions that efciently solve different control objectives [10, 27, 39]. However,
an important issue such of robustness against parametric uncertainty and external
perturbations have not been addressed in depth. Besides, the validity range of these
controllers is local, and results do not extend to the entire operation range. Solutions
to this problem can be found in the nonlinear control eld, such as [1] and [24, 25],
where strategies based on parametric cerebellar model articulation and exact linearisation were addressed, respectively. More recently, in [6] a methodology based
on an energetic macroscopic representation of the fuel cell stack is proposed. Good
global performances were attained with these nonlinear techniques, but, regretfully,
their applicability in real systems is still limited because the algorithms demand
considerable computational burden.
Therefore, despite these advances in fuel cell controllers, it is evident that to meet
the expected enhanced capabilities, a substantial R&D work is still necessary, and
a control approach particularly suitable to cope with their challenging features is of
high interest. At this point, Sliding-Mode control emerges as an especially apt technique to tackle the complex characteristics inherent to fuel cell systems (e.g. high
nonlinear dynamics, inaccessible variables, model uncertainties and disturbances).
In addition, the on-line computational burden of the resultant algorithms can be conveniently low. Promising results have been obtained with sliding mode controllers
for fuel cell systems [16, 22, 32], strongly encouraging the prosecution of research
in this direction.

References
1. Almeida PE, Godoy Simoes M (2005) Neural optimal control of PEM fuel cells with parametric CMAC networks. IEEE Trans Ind Appl 41(1):237245
2. Amphlett J, Baumert R, Mann R, Peppley B, Roberge P (1995) Performance modeling of the
Ballard Mark IV solid polymer electrolyte fuel cell. J Electrochem Soc 142(1):915
3. Arce A, del Real A, Bordons C, Ramirez D (2010) Real-time implementation of a constrained
MPC for efcient airow control in a PEM fuel cell. IEEE Trans Ind Electron 57(6):1892
1905
4. Barbir F (2005) PEM fuel cells: theory and practice. Elsevier, Amsterdam
5. Bao C, Ouyang M, Yi B (2006) Modeling and control of air stream and hydrogen ow with
recirculation in a PEM fuel cell systemI. Control-oriented modeling. Int J Hydrog Energy
31(13):18791896
6. Boulon L, Hissel D, Bouscayrol A, Pera MC (2010) From modeling to control of a PEM fuel
cell using energetic macroscopic representation. IEEE Trans Ind Electron 57(6):18821891
7. Feroldi D, Serra M, Riera J (2009) Design and analysis of fuel cell hybrid systems oriented to
automotive applications. IEEE Trans Veh Technol 58(9):47204729
8. Gao F, Blunier B, Simes M, Miraoui A (2010) PEM fuel cell stack modeling for real-time
emulation in hardware-in-the-loop applications. IEEE Trans Energy Convers 26(1):184194
9. Grasser F, Rufer A (2007) A fully analytical PEM fuel cell system model for control applications. IEEE Trans Ind Appl 43(6):14991506

32

PEM Fuel Cell Systems

10. Golbert J, Lewind R (2004) Model-based control of fuel cell. J Power Sources 135:135151
11. Gottesfeld S, Zawodzinski T (1997) Polymer electrolyte fuel cells. Wiley, New York
12. Husar A, Serra M, Kunusch C (2007) Description of gasket failure in a 7 cell PEMFC stack.
J Power Sources 169(1):8591
13. Husar A (2008) Dynamic water management of an open-cathode self-humidied PEMFC system. PhD proposal at Universitat Politctica de Catalunya
14. Khan MJ, Iqbal MT (2005) Modelling and analysis of electro-chemical, thermal, and reactant
ow dynamics for a PEM fuel cell system. Fuel Cells 5(4):463475
15. Kim B, Chang L, Gadd G (2007) Challenges in microbial fuel cell development and operation.
Appl Microbiol Biotechnol 76:485494
16. Kunusch C, Puleston P, Mayosky M, Riera J (2009) Sliding mode strategy for PEM fuel cells
stacks breathing control using a super-twisting algorithm. IEEE Trans Control Syst Technol
17(1):167174
17. Larminie J, Dicks A (2000) Fuel cell systems explained. Wiley, West Sussex
18. Lee J, Lalk T, Appleby A (1998) Modeling electrochemical performance in large scale proton
exchange membrane fuel cell stacks. J Power Sources 70:258268
19. Lee H, Jeong K, Oh B (2003) An experimental study of controlling strategies and drive forces
for hydrogen fuel cell hybrid vehicles. Int J Hydrog Energy 28:215222
20. Mann R, Amphlett J, Hooper M, Jensen H, Peppley B, Roberge P (2000) Development and
application of a generalized steady-state electrochemical model for a PEM fuel cell. J Power
Sources 86:173180
21. Mathias M, Roth J, Fleming J, Lehnert W (2003) Handbook of fuel cells. Fundamentals,
technology and applications. Wiley, New York
22. Matraji I, Laghrouche S, Wack M (2010) Second order sliding mode control for PEM fuel
cells. In: 49th IEEE conference on decision and control CDC 2010, pp 27652770
23. Methekar R, Pradas V, Gudi R (2007) Dynamic analysis and linear control strategies for
proton exchange membrane fuel cell using a distributed parameter model. J Power Sources
165:152170
24. Na W, Gou B (2008) Feedback-linearization-based nonlinear control for PEM fuel cells. IEEE
Trans Energy Convers 23:179190
25. Na W, Gou B, Diong B (2007) Nonlinear, control of PEM fuel cells by exact linearization.
IEEE Trans Ind Appl 43:14261433
26. Nguyen T, White R (1993) A water and heat management model for proton-exchangemembrane fuel cells. J Electrochem Soc 140(8):21782186
27. Pukrushpan J, Stefanopoulou A, Peng H (2004) Control of fuel cell breathing. IEEE Control
Syst Mag 24(2):3046
28. Rodatz P (2003) Dynamics of the polymer electrolyte fuel cell: experiments and model-based
analysis. PhD thesis, Swiss Federal Institute of Technology Zurich
29. Sammes N (2006) Fuel cell technologies. Engineering materials and processes. Springer,
Berlin
30. Springer C, Zawodzinski T, Gottesfeld S (1991) Polymer electrolyte fuel cell model. J Electrochem Soc 138(8):23342342
31. Squadrito G, Barbera O, Giacoppo G, Urbani F, Passalacqua E (2008) Polymer electrolyte
fuel cell stack research and development. Int J Hydrog Energy 33:19411946
32. Talj R, Hissel D, Ortega R, Becherif M, Hilairet M (2010) Experimental validation of a PEM
fuel-cell reduced order model and a moto-compressor higher order sliding-mode control. IEEE
Trans Ind Electron 57(6):19061913
33. Thounthong P, Ral S, Davat B (2006) Control strategy of fuel cell/supercapacitors hybrid
power sources for electric vehicle. J Power Sources 158:806814
34. Vahidi A, Stefanpoulou A, Peng H (2004) Model predictive control for starvation prevention
in hybrid fuel cell systems. In: Proceedings of the American control conference, pp 834839
35. Vega-Leal A, Palomo F, Barragn F, Garca C, Brey J (2007) Design of control systems for
portable PEM fuel cells. J Power Sources 169:194197
36. Wai R, Duan R, Lee J, Liu L (2005) High-efciency fuel-cell power inverter with softswitching resonant technique. IEEE Trans Energy Convers 20:482492

References

33

37. Wang F, Chen H, Yang Y (2008) Multivariate robust control of a proton exchange membrane
fuel cell system. J Power Sources 177:393403
38. Woo C, Benziger J (2007) PEM fuel cell current regulation by fuel feed control. Chem Eng
Sci 62:957968
39. Yuan R, Cao G, Zhu X (2005) Predictive control of proton exchange membrane fuel cell
(PEMFC) based on support vector regression machine. In: International conference on machine learning and cybernetics, pp 40284031

Chapter 3

Fundamentals of Sliding-Mode Control Design

3.1 Introduction
This chapter provides an introduction to Variable Structure Control (VSC) theory
and its extension to the so-called Sliding-Mode (SM) control. Note that the presentation is not intended as a comprehensive survey of the state-of-the-art in the
eld, but to merely supply the basic concepts on SM control required to understand
the developments to come in this book. Readers well acquainted with this subject
matter may omit this chapter. On the other hand, rst-timers can use this material
as a straightforward, but incomplete, introduction to the eld of SM control, and
are strongly encouraged to search for further and more substantial reading in the
seminal works cited in the bibliography (useful introductory material could be, for
instance, [5, 12, 14, 25, 26, 38, 44, 45]).
The chapter is divided in two parts. In the rst one, Sect. 3.3, a general analysis
of the classic or rst-order SM control is formulated, which is the natural background to the subsequent generalisation known as Higher-Order Sliding-Mode control (HOSM). This section is mainly based on the inuential works [38, 42, 44] and,
to a lesser extent, on contributions from a series of classic survey papers such as
[12, 45].
In the second part, Sects. 3.4 and 3.5, a general study of systems operating in
sets of arbitrary sliding-mode order is presented. This section outlines the fundamentals of Higher-Order Sliding-Mode control theory, particularly focusing on
Second-Order Sliding-Mode (SOSM) controllers. To a great extent, this part has
been inspired in the works and results from [5, 25, 31, 33].

3.2 Variable Structure Control Preliminaries


The variable structure control and associated sliding modes were rstly proposed
and developed by Stanislav Emelyanov and Vadim Utkin in the early 1950s in the
Soviet Union [16, 43]. The most relevant feature of the SM control is its ability to
C. Kunusch et al., Sliding-Mode Control of PEM Fuel Cells,
Advances in Industrial Control,
DOI 10.1007/978-1-4471-2431-3_3, Springer-Verlag London Limited 2012

35

36

Fundamentals of Sliding-Mode Control Design

generate robust control algorithms that are invariant under certain conditions. Briey
speaking, the concept of invariance indicates that the system remains completely
insensitive to certain types of disturbances and uncertainties [13].
Since the 1990s, the control of systems subject to external disturbances and
model uncertainty has been the focus of increasing interest. Among the different
existing alternatives, the SM control has proven to be an attractive option to implement in systems electronically controlled, proving to be highly robust and even
insensitive against certain system uncertainty and perturbations. The feasibility and
benets of SM control applied to electronically controlled actuators have been extensively demonstrated in the literature, such as [44]. In addition, the SM control
allows a relatively simple design approach, even dealing with nonlinear systems,
admitting a successful combination with other nonlinear control techniques such
as energy shaping and model predictive control. As a result, the research and development of SM control design methods have been greatly accelerated, both in
theoretical and practical elds [1, 6, 15, 36, 46].
One of the most distinctive aspects of the SM is the discontinuous nature of its
control action. Its primary function consists in performing a switching between two
different structures in order to get a desired new dynamics in the system, known as
sliding-mode dynamics. This feature allows the system to have an enhanced performance, including insensitivity to parametric uncertainties and rejection to disturbances that verify the so-called matching condition [13, 38]. When the concept of
parametric uncertainties is considered, it is referred to both external and internal uncertainties in the parameters as the product of the process of model reduction used
in control design [14, 42].
However, a great deal of the success to full the control objectives depends
on the capability of the sliding-mode controller design to reduce chattering. The
term chattering describes the phenomenon of nite-frequency, nite-amplitude oscillations appearing in many sliding-mode implementations. These oscillations are
caused by the high-frequency switching of a sliding-mode controller under practical
(non-ideal) operating conditions, such as unmodelled dynamics in the closed-loop
or nite switching frequency [9, 22, 23, 44].
A successful alternative to reduce this undesired phenomenon, currently addressed by many control researchers and engineers, is to use the so-called HigherOrder Sliding-Mode control. In this case, from the denition of a continuous control
action, the HOSM generalises the notion of sliding surface or manifold while keeping the main advantages of the original approach of SM for Lipschitz continuous
uncertainty/perturbations. In particular, there are several promising results related
to Second-Order Sliding-Mode control, existing several algorithms that solve the
robust stabilisation of nonlinear uncertain systems, while guaranteeing a nite-time
convergence of the sliding variable [5, 24, 29, 34].

3.3 Fundamentals of Sliding-Mode Control


The SM control is a strategy based on output feedback and a high-frequency switching control action which, in ideal conditions, is innite. Essentially, this high-speed

3.3 Fundamentals of Sliding-Mode Control

37

control law can lead the system trajectories to a subspace of the state space (commonly associated to a sliding surface or manifold). If a system is forced to constrain its evolution on a given manifold, the static relationships result in a dynamical behaviour determined by the design parameters and equations that dene the
surface [38]. On average, the controlled dynamics may be considered as ideally
constrained to the surface while adopting all its desirable geometrical features.
Thus, making an appropriate design of the sliding surface (i.e. embedding the
control objectives into the control function that gives rise to such manifold), it is
possible to achieve conventional control goals such as global stability, optimisation,
tracking, regulation, etc.
In the sequel, the basics of the theory of classical sliding-mode control are introduced, focusing on Single-Input Single-Output (SISO) systems. Note that in most
sections of this chapter, the possible explicit dependence on time of the dynamical
system has been omitted for the sake of clarity and economy of notation. In the
present approach, this compacted notation can be used without loss of generality,
provided that in the case of a non-autonomous system, it could be rewritten as autonomous by treating t as an additional dependent variable, with its trivial evolution
given by the ctitious equation t = 1 (obviously, at the expense of increasing the
dimension by one).

3.3.1 Diffeomorphisms, Lie Derivative and Relative Degree


Firstly, it is useful to review some mathematical tools and procedures that will be
necessary later. Let a control afne nonlinear system be given by
x = f (x) + g(x)u
y = h(x)

(3.1)

with x X Rn , f : Rn Rn and g : Rn Rn smooth vector elds (innitely


differentiable) with g(x)
= 0, h(x) smooth scalar eld and u : Rn R possibly
discontinuous. These systems are linear in the control, so they are called control
afne systems or analytical linear systems.
A diffeomorphism is dened as a coordinate transformation of the form z = (x)
with : Rn Rn vector eld with inverse 1 . In particular, we only consider
transformations such that and 1 are C n (i.e. with n continuous derivatives).
This last condition ensures that the transformed system preserves the original system
structure.
After making the proposed change of coordinates, the dynamical system (3.1)
looks as follows:

x =
f (x) +
g(x)u
(3.2)
z = (x)
=
x
x
x
(x) (x)
(x)
Note that
x = [ x1 x2 xn ] gives the direction of the gradient vector of
(x), (x). So the system (3.1) can be written in terms of the new variable z:

38

Fundamentals of Sliding-Mode Control Design

z = f(z) + g(z)u

y = h(z)

(3.3)


f(z) =
f (x) 1
x= (z)
x



g(z)
=
g(x) 1
x= (z)
x
h(z) = h(x) x= 1 (z)

(3.4)

where

To simplify the notation, it is necessary to dene the concept of directional


derivative or Lie derivative [40], which is expressed as
(Lf h)(x) = Lf h(x) : Rn R

(3.5)

and represents the derivative of a scalar eld h(x) : Rn R in the direction of a


vector eld f (x) : Rn Rn ,
Lf h(x) =

h
f (x)
x

(3.6)

Lf is a rst-order differential operator, while the composition Lf Lg , which is usually written as Lf Lg , is a second-order operator. Moreover, the directional derivative can be applied recursively:
Lkf h(x) =

k1
Lf h(x) f (x)
x

(3.7)

In this way, a compact notation for the derivatives of scalar functions in the direction of vector elds is obtained. Either in the direction of a single vector eld (f )
or more (f and g):
Lg Lf h(x) =


Lf h(x) g(x)
x

(3.8)

Finally, assuming a smooth output h(x) of system (3.1), the relative degree of
h(x) at the vicinity of a given point x is dened as the smallest positive integer r, if
one exists, with the property that
Lg Lif h = 0

0 i r 2

(3.9)

and
Lg Lr1
f h
= 0

(3.10)

Therefore, a system output h(x) with relative degree r implies, in a simplied


way, that u explicitly appears for the rst time at the rth time derivative of h(x). In
short, r gives an idea about how directly the control inuences the output.

3.3 Fundamentals of Sliding-Mode Control

39

3.3.2 First-Order Sliding Mode


Consider the nonlinear dynamical system (3.1), with control action u : Rm R
(possibly discontinuous), f and g smooth vector elds with g(x)
= 0 x X. Let s
be dened as a smooth constraint function s : X R, designed according to the
desired control objectives (i.e. the specications are fullled when s is constrained
s
non-null on X [38]. Then the set
to zero), with gradient s = x


(3.11)
S = x X Rn : s(x) = 0
denes a locally regular manifold in X (of dimension n 1 in the case of a SISO
system), called sliding manifold or, simply, switching surface. This order reduction
feature is a characteristic of SM control systems (rst and HOSM) and indicates that
the subspace on which the sliding movements occur have non-zero co-dimension,
meaning that after reaching the sliding regime, the trajectories of the system will
remain within a subspace of lower dimension than the space generated by n states.
The results obtained below are of a local nature, restricted to an open neighbourhood
of X Rn , having a non-empty intersection with the sliding manifold S [16, 28,
42].
In order to attain the sliding motion in such manifold, a variable structure control
law can be proposed by imposing a discontinuous control action u, which takes one
of two possible feedback values, depending on the sign of s(x). For example,
+
u (x) if s(x) > 0
with u+
= u
(3.12)
u=
u (x) if s(x) < 0
The upper and lower levels of u (u+ (x) and u (x), respectively) are smooth
functions of x. Moreover, without loss of generality, it can be assumed that u+ (x) >
u (x) holds locally in X. Note that if u+ (x) > u (x) for any point x, then the
inequality holds for every x, given that the functions are smooth and do not intersect.
Suppose that, as a result of the control law (3.12), the constraint function locally
satises the following inequalities in the neighbourhood of S :

s (x) < 0 if s(x) > 0
(3.13)
s (x) > 0 if s(x) < 0
Under these conditions, the system will reach the sliding manifold S and thereafter will remain conned in a vicinity of S (see Fig. 3.1). Then, it is considered
that a sliding regime is established on S whenever (3.13) holds.
Using the notation of the directional derivative, s (x) can be expressed as follows:
s (x) = Lf +gu s = Lf s + Lg s u

(3.14)

Note that the output s(x) must have relative degree 1 with respect to u, i.e.
Lg s
= 0, to ensure that the discontinuous control action is able to inuence the sign
of s (x).
Expression (3.13) can also be written as follows:

lim Lf +gu+ s < 0


s+0
(3.15)
lim Lf +gu s > 0
s0

40

Fundamentals of Sliding-Mode Control Design

Fig. 3.1 Sliding manifold and system trajectories

meaning that the rate of change of the constraint or SM function s(x), evaluated in
the direction of the control eld, is such that a crossing of the surface is guaranteed
from each side of the surface, by using the switching law (3.12). This can be graphically interpreted with the help of Fig. 3.1, analysing the projection of the controlled
eld f + gu onto the gradient vector s at both sides of S .
To conclude this subsection, a succinct nal remark regarding the SM control
nite reaching time is pertinent. Note then that the explicit condition (3.13) can be
condensed as s (x)s(x) < 0. From this it is simple to understand that, to achieve
nite reaching time, the control law (3.12) must be designed to full the previous
inequality, but in a more strict way, that is according to the scalar sufcient condition
s (x)s(x) < |s(x)| with > 0 (or, similarly, s (x) sign s(x) < ). This means that
the system should always be moving toward the switching surface with non-zero
speed. This can be straightforwardly proven by taking V = 12 s 2 (x) as a Lyapunov
function.

3.3.3 Equivalent Control Regularisation Method. Ideal Sliding


Dynamics
From a methodological and systematic point of view, it is convenient to develop a
regularisation method for deriving the sliding-mode equations for system (3.1). Assuming that the state vector is in the manifold S (s(x) = 0) and the sliding mode occurs with the state trajectories conned to this manifold for t > 0, one way to dene
the ideal sliding mode is using the so-called equivalent control method [44]. Since
the motion in the sliding mode implies s(x) = 0 for t > 0, it may be assumed that
ds/dt = s = 0 as well. Hence, in addition to s(x) = 0, the time derivative s (x) = 0
may be used to characterise the state trajectories during the sliding mode.

3.3 Fundamentals of Sliding-Mode Control

41

In summary, the equivalent control action is dened by the following invariance


conditions on the switching manifold S [38]:

s(x) = 0
(3.16)
s (x) = Lf s + Lg s ueq = 0
where ueq (x) is a smooth control law called equivalent control that makes S a local
invariant manifold of system (3.1). Therefore, the equivalent control ueq (x) can be
obtained from Eq. (3.16):

Lf s 
(3.17)
ueq (x) =
L s
g

s(x)=0

Thus, once s = 0 is attained, ueq (x) would provide the continuous control action
required to maintain the system conned in the sliding surface.
The ideal sliding-mode dynamics, i.e. the closed-loop dynamics on the manifold S , is obtained by substituting ueq for u into (3.1):

Lf s 
x = f (x) + g(x)ueq |s(x)=0 = f (x) g(x)
(3.18)
Lg s s(x)=0
Note that the state variables are related by the algebraic equation s(x) = 0, reducing
the order of the closed-loop system dynamics to n 1.
Substituting the Lie derivative and operating in (3.18), we have




s 1 s
g
x = I g
f (x) = (x)f (x)
(3.19)
x
x
Evaluated in s(x) = 0, (3.19) gives an idealised version of the motions occurring on
the sliding manifold S , constituting an average description of the trajectories of
system (3.1) controlled with the VSC law (3.12).
The geometrical representation presented in Fig. 3.2 can be of help for a better
understanding. In accordance with ueq being the control action that makes the system remain on S , the vector f = f + gueq must lie in Tx , the tangent plane to
S (i.e. normal to the gradient s, as can be seen in Fig. 3.2). Mathematically this
is expressed as
(x)f (x) ker(s) Tx

(3.20)

Consequently, the matrix (x) can be considered as a projection operator that,


applied to the vector f (x), projects it onto the plane tangent to surface S at the
point x.
To conclude, it is of interest to briey consider the effect of projector over
any vector collinear with g(x). Let be a general vector of arbitrary amplitude,
possibly a function of x, such that span(g):
(x) = g(x)(x)

with (x) : Rn R

(3.21)

Note then that the application of the operator (x) to this vector projects to
the origin. In fact,




s 1 s
g
g(x)(x) = 0
(3.22)
(x)(x) = I g
x
x

42

Fundamentals of Sliding-Mode Control Design

Fig. 3.2 Sliding surface and detail of the vector elds and projections

The cancellation of expression (3.22) can be interpreted in Fig. 3.2 as follows: the
operator (x) projects any vector in the direction of g(x) onto the tangent subspace
of S . Hence, the projection of any vector that belongs to span(g) would result in
just a point on Tx .

3.3.4 Existence Conditions for the First-Order Sliding Regime


3.3.4.1 Existence of Equivalent Control
It can be stated that the equivalent control is well dened if ueq exists and is uniquely
determined from the invariance conditions (3.16) [38].
Lemma 3.1 The equivalent control is well dened if and only if the following condition is satised locally in S :
s(x)
g(x)
= 0
(3.23)
x
This condition is known as transversality condition and may be inferred
from (3.14). The proof of the lemma can be found in [38].
Lg s(x) =

Geometrically, this lemma states that the vector eld g cannot be tangential to
the sliding manifold (S : g
/ ker(s)); otherwise it could not force the system to
cross the surface. The transversality condition represents just a necessary condition
for the existence of a rst-order sliding mode.

3.3 Fundamentals of Sliding-Mode Control

43

3.3.4.2 Necessary Conditions for the Existence of a First-Order Sliding


Regime
Based on the transversality condition, the following necessary condition for the existence of a sliding regime can be stated.
Lemma 3.2 A necessary condition for the existence of a local sliding mode in S
is that the equivalent control action ueq (x) must be well dened.
Indeed, if ueq is not well dened, i.e. Lg s = 0 at some point, the existence conditions of the sliding mode (3.15) cannot be satised simultaneously.
Lemma 3.3 Assume, without loss of generality, that u+ (x) > u (x). Then the following condition is necessary for the existence of a sliding regime on S :
Lg s(x) =

s
g(x) = sg(x) < 0
x

(3.24)

The proof, direct from (3.15) and (3.16), is given in [38] and can be easily inferred from Fig. 3.2 by analysing the sign of the projection of g(x) onto s(x).

3.3.4.3 Necessary and Sufcient Condition for the Existence of a First-Order


Sliding Regime
A necessary and sufcient condition for the local existence of a sliding mode in S
is that, for x S ,
u (x) < ueq (x) < u+ (x)

(3.25)

This condition can also be proved from (3.15) and (3.16) [38].
Then, ueq can be interpreted as the averaged control signal resulting from the
implementation of the maximum and minimum control actions, with an innitesimal
duty cycle resolution (in ideal sliding mode). However, in practice, several model
imperfections and nite switching frequency make the state oscillate in a vicinity of
the manifold [44].

3.3.4.4 Robustness of the First-Order SM


The behaviour of SM controlled systems under the effect of disturbances is discussed briey. To this end, consider system (3.1) perturbed as follows:
x = f (x) + g(x)u + (x)

(3.26)

with (x) a vector of lumped perturbations that may take into account parametric perturbations of the nominal drift eld or unstructured external disturbances

44

Fundamentals of Sliding-Mode Control Design

Fig. 3.3 Sliding surface and detail of the vector elds with perturbations

[38]. The vector (x) can be uniquely decomposed into two components, one in
the span(g), g(x)(x), and the other, (x), onto the tangent plane Tx (see Fig. 3.3):
(x) = g(x)(x) + (x)

(3.27)

It is said that the disturbances that belong to span(g) satisfy the matching condition, and the SM control is not merely robust to them but exhibits a strong invariance property. Effectively, as it can be observed in Fig. 3.3, if u+ and u are strong
enough, the component g(x)(x) can be completely annihilated by the control, simply generating a new innitesimal duty cycle and, consequently, a new equivalent
control for the disturbed system (ueqp (x) = ueq (x) (x)). In addition, the undisturbed or nominal sliding dynamics suffers no modications.
On the other hand, it can be appreciated (in Fig. 3.3) that the tangential component of the disturbances, (x), cannot be rejected. However, it does not compromise
the local existence of the sliding motion, but denitely inuences the ideal sliding
dynamics.
In accordance with this analysis, it can be stated that a necessary and sufcient
condition for the local existence of a sliding mode in the perturbed system is
u (x) < ueqp (x) = ueq (x) (x) < u+ (x)
A detailed demonstration is provided in [38].

(3.28)

3.3 Fundamentals of Sliding-Mode Control

45

3.3.5 Extension to Nonlinear Systems Non-afne in Control


Consider a generic system described by the following differential equation:
x = F (x, u)

(3.29)

Using the same SM function s(x), the controlled system can be again decomposed into two subsystems or structures, depending on whether s(x) > 0 or
s(x) < 0:

F (x, u+ ) = F + if s(x) > 0
x = F (x, u) =
(3.30)
F (x, u ) = F if s(x) < 0
Along the trajectories of the system, the dynamics of s(x) has the following
expression:


 x1
s
s s
s

x1 +
x2 + =
x2 = LF s(x)
(3.31)
s (x) =
x1
x2
x1 x2
..
.
In the same way as in system (3.30), in the time derivative of the SM function
two cases can be distinguished:
if s > 0 s = LF + s(x)
if s < 0 s = LF s(x)

(3.32)

As the trajectory has to converge to the manifold, when s(x) > 0, the states
should move towards s(x) = 0 (i.e. s (x) < 0, so s(x) decreases), and conversely
in the reciprocal case. This means that the establishment of the sliding mode on
s(x) = 0 is fullled with a condition similar to (3.13):
if s > 0 s = LF + s(x) < 0
if s < 0 s = LF s(x) > 0

(3.33)

3.3.6 Filippov Regularisation Method


Besides the equivalent control method presented in Sect. 3.3, at this point it is of
interest to introduce another regularisation method, also capable of dealing with
discontinuous systems. In particular, the underlying concept behind this method will
be of use in the higher-order SM control strategies to come.
Recall that conventional theory of differential equations is limited to continuous
state functions, hence when dealing with discontinuous systems, it does not answer
even fundamental questions, such as the existence and uniqueness of the solution.
Strictly speaking, most conventional methods require the right-hand side of the differential equation (3.29) to satisfy the Lipschitz condition


F (x1 ) F (x2 ) < L x1 x2
(3.34)

46

Fundamentals of Sliding-Mode Control Design

Fig. 3.4 Hysteresis of the


switching device

with L being some positive value, known as Lipschitz constant, for any x1 and x2 .
This condition implies that the function does not grow faster than some linear function [44]. Nevertheless, this is not the case for discontinuous functions if x1 and x2
are close to a discontinuity point.
So as previously stated, in situations where conventional methods are not applicable, the common approach is to employ different methods of regularisation like the
equivalent control method proposed in [42]. Another useful regularisation method
usually applied to general nonlinear systems as (3.29) is the so-called Filippov
method [20]. This procedure consists of considering that the discontinuous control is implemented with a switching device with small imperfections. In particular,
if a hysteresis loop of width 2 is considered, then the state trajectories oscillate in
a -vicinity of the switching surface when the control takes one of the two extreme
values, u+ (x) or u (x) (see Fig. 3.4).
is considered small enough, so the state velocities F + = F (x, u+ ) and F =
F (x, u ) are assumed to be constant for some point x on the surface s(x) = 0
within a short time interval [t, t + t]. Let the time interval t consist of two sets
of intervals t1 and t2 such that t = t1 + t2 , u = u+ during t1 and u = u
during t2 . Then, the increment of the state vector once t is elapsed is found as
x = F + t1 + F t2

(3.35)

and the average velocity of the state vector is given by the convex average of the
velocity vectors:
x
= F + + (1 )F
(3.36)
x =
t
where the convex average factor = t1 /t can be understood as the percentage
of time that the control takes the value u+ , while (1 ) is the percentage corresponding to u , with belonging to the closed set [0, 1]. Now, the procedure
to get the state vector movement x is to make t tend to zero. Nevertheless, this
limit is intrinsic to the assumption that the state velocity vector, or equivalently the
vector eld F (x), is constant within the time interval t. Then, for the Filippov
regularisation method, the convex expression
x = F + + (1 )F

(3.37)

3.3 Fundamentals of Sliding-Mode Control

47

Fig. 3.5 Filippovs


regularisation method

represents the motion during the rst-order sliding mode (see the convex closure in
Fig. 3.5, as a graphical interpretation of the Filippov method). Accordingly, since
the trajectories during the sliding mode are on the manifold s(x) = 0, the following
equation holds:


s = s(x)x = s(x) F + + (1 )F = 0
(3.38)
so the parameter should take a value that allows the state velocity of the system
(3.37) to lie on the tangent plane (see Fig. 3.5). From (3.38) it can be easily inferred
that such value of must be
=

s(x)F
.
s(x)[F F + ]

(3.39)

Note For control afne systems, the resultant sliding equations derived from the
Filippov regularisation method are the same as those obtained from the Utkin equivalent control.

3.3.7 Discontinuous Control Action in Classic Sliding-Mode


Control. Chattering Problem
One of the main drawbacks of the rst-order sliding-mode control in certain applications is the direct use of discontinuous control actions. In actual implementations,
the discontinuous control law, together with unmodelled dynamics and nite switching frequency, may produce fast oscillations in the outputs of the system. This effect
is known as chattering phenomenon.
During the mid-1980s, the following three main approaches to reduce chattering
in sliding-mode controlled systems were proposed [8]:
The use of a saturation control instead of the discontinuous action [10, 39]. This
well-established approach allows the control to be continuous, restraining the system dynamics not strictly onto the sliding manifold, but within a thin boundary
layer of the manifold. This method ensures the convergence to the boundary layer,
whose size is dened by the slope of the saturation linear region.

48

Fundamentals of Sliding-Mode Control Design

The observer-based approach [9, 44]. This method allows bypassing the plant
dynamics by the chattering loop. This approach successfully reduces the problem
of robust control to the problem of exact robust estimation. However, in some
applications it can be sensitive to the plant uncertainties, due to the mismatch
between the observer and plant dynamics [45].
The Higher-Order Sliding-Mode approach (HOSM) [18, 29]. This method allows
the nite-time convergence of the sliding variable and its derivatives. This approach was actively developed since the 1990s [2, 4, 5, 29, 32, 37], not only providing chattering attenuation, but also robust control of plants of relative degree
one and higher. Theoretically, an r-order sliding mode would totally suppress
the chattering phenomenon in the model of the system (but not in the actual system) when the relative degree of the model of the plant (including actuators and
sensors) is r. Yet, no model can fully account for parasitic dynamics, and, consequently, the chattering effect cannot be totally avoided. Nevertheless, theoretical
results in HOSM, especially Second-Order Sliding-Mode algorithms, have been
successfully proven in practice, encouraging the progress of the research activities.
Then, it is of natural interest the study of sliding-mode control alternatives that,
smoothing the control action, reduce the chattering effects and avoid unnecessary
requirements on the actuators. This is particularly relevant in fuel cell control, as
there are mechanical actuators involved that may suffer when exposed to control
actions of high frequency and amplitude.
It should be noted that when using Higher-Order Sliding Modes, it is not possible
to maintain the invariance properties against matched disturbances as in the original
approach. However, different control schemes that guarantee robust stability of the
system can be achieved, satisfying the condition s(x) = 0 (and even zeroing higherorder derivatives of s(x)) in nite time [18, 25].
In the sequel, a brief introduction to Higher-Order Sliding-Mode control applied
to uncertain nonlinear systems is presented. Then, Second-Order Sliding-Mode control and in particular three different algorithms are analysed in detail.

3.4 Some General Concepts on Higher-Order Sliding Modes


As discussed in Sect. 3.2, rst-order sliding-mode control has certain properties that
make it particularly attractive to apply to uncertain nonlinear systems. Among them,
it can be highlighted nite convergence to the surface, system order reduction and
robustness against certain disturbances. In this context, Higher-Order Sliding-Mode
control will inherit some of these properties. This control approach generalises the
idea of rst-order sliding mode, by acting on the higher-order derivatives of the constraint function s(x), instead of inuencing the rst derivative (as in (3.14)). Keeping the main advantages of the original approach, the HOSM control works with
continuous action over s (x), relegating the discontinuous control to operate on the
higher derivatives of s(x). This weakens the effect of chattering in the output, providing greater accuracy in realisation. Additionally, in some applications (namely,

3.4 Some General Concepts on Higher-Order Sliding Modes

49

plants with relative degree 1 with respect to s), the resultant physical control input
to the plant is continuous, contributing to the longer service life of certain actuators.
A signicant number of these controller proposals can be found in [2, 7, 18, 21, 25,
29, 30, 35].
An important concept in HOSM is the notion of sliding order. If the goal is to
maintain a constraint given by s(x) = 0, the sliding order is dened as the number
of continuous time derivatives of s(x) (including the zero-order one) in the vicinity
of a sliding point. With these considerations, a sliding mode of order r is determined
by the following equalities:
s = s = s = = s (r1) = 0

(3.40)

Expression (3.40) represents an r-dimensional condition in the dynamic system,


which implies an order reduction of r (that is, (3.40) species r algebraic equations
that bond the state variables).

3.4.1 Denition of Differential Inclusion


As in rst-order sliding-mode control, the HOSM scheme forces a movement on a
set of discontinuity, demanding an approach to the problem capable to deal with differential equations with right-hand side single-valued, but discontinuous, functions.
Such an approach can be found in the Filippov concepts introduced in Sect. 3.3.6.
The basic idea behind Filippovs method was not to focus on the value of the vector
function precisely at the discontinuity point, but on its values in the points immediate neighbourhood. Then, the function at the point is replaced with an average
function, taken from a set generated by the convex combination of the values at
each side of the discontinuity point.
This replacement can be interpreted as including or covering the discontinuous
right-hand side single-valued function with a more comprehensive set-valued function (as would be the convex closure presented in the case treated in Sect. 3.3.6).
This idea of inclusion, or better expressed differential inclusion, will be of help
when designing SOSM controllers for dynamical system with uncertainties.
To better formalise this mathematical concept, consider a general differential
equation of the form
z = v(z, t)

(3.41)

where the generic variable z Rn , and v(z, t) is a piecewise-continuous singlevalued function in a domain G with some points of discontinuity in a set M of
measure zero. Note that in the framework of the SM control problem the generic
variable z could be particularised to be a suitable variable of the dynamical system
under control (for instance, x or, through a diffeomorphism, s, s and any appropriate
internal variables), while the discontinuity set M could be the sliding manifold.
Next, for each point (z, t) of the domain G, a set-valued function V (z, t) in an
n-dimensional space must be considered. Note that just as the single-valued function takes a point in its domain into a single point (direction) in another space, the

50

Fundamentals of Sliding-Mode Control Design

set-valued function take a point in its domain into a set of points (directions) in
another space [11]. In this particular case, for points (z, t) where function v(z, t) is
continuous, the set V (z, t) trivially consists of one point (direction) which coincides
with the single value of v(z, t) at this point. On the other hand, if (z, t) is a point
of discontinuity of function v(z, t), then V (z, t) comprises a set of directions rather
than a single specic one. Now, in accordance with Filippovs denition, the discontinuous differential equation (3.41), can be formally replaced by an equivalent
differential inclusion of the form
z V (z, t)

(3.42)

The expression above, is called a Filippov differential inclusion if the set V (z, t)
is non-empty, closed, convex, locally bounded and upper-semicontinuous. In this
way, V (z, t) can cover the situation in which the state derivative belongs to a set of
directions, not to a single one. In the simplest case, i.e. when v(z, t) is continuous
almost everywhere, V (z, t) is the convex closure of the set of all possible limits of
v(t, zcont ) as zcont z, while zcont are continuity points of v(z, t).
Note that this denition veries the description of V (z, t) previously given.
When zcont approaches a continuity point, the limits converge to a single value,
so, as expected, V (z, t) effectively coincides with the continuous value of v(z, t).
Conversely, when zcont approaches a discontinuity point, limits are different, and
V (z, t) comprises a set of directions.
It can be stated, then, that a solution z(t) of the differential equation (3.41) is
understood as a solution in the Filippov sense, if it is an absolute continuous function
in an interval and satises the differential inclusion (3.42) almost everywhere on
such interval [19, 20].
Summarising, the Filippov denition replaces the discontinuous differential
equation (3.41) by the differential inclusion (3.42). Removing sets of zero measure
(discontinuity points) from the values taken by v(z, t) corresponds to purposely ignoring possible misbehaviour of the right-hand side in (3.41) on small sets.

3.4.2 Sliding Modes on Manifolds


The notion of sliding mode manifold acquired with the rst-order SM can be extended to HOSM. The progression that generates the successive sliding manifolds
can be described as follows. Let S be the smooth manifold dened from a smooth
function s(x) (see Eq. (3.11)). The set of points x for which the set of possible velocities entirely lies in the subspace Tx tangent to S is dened as a second-order
sliding set with respect to S (recall that in a rst-order SM the set of possible velocities of the system does not lie in Tx , but intersects it. See Fig. 3.2). The former concept means that once S is reached, the Filippov solutions of Eq. (3.41) fall within
the tangent space of the manifold S . This set of points is denoted as S2 . Assuming that S2 can be considered as a manifold smooth enough, the same construction
can be performed for S2 , calling S3 to the corresponding set of second-order sliding solutions with respect to S2 or third-order sliding set with respect to S . Thus,

3.4 Some General Concepts on Higher-Order Sliding Modes

51

continuing this way, one can nd sliding sets of any order [25]. Summarising, it is
said that there is an rth-order sliding mode on the manifold S in a neighbourhood
of an rth-order sliding point x Sr if in a neighbourhood of this point x, the set
Sr is an integral set, and this means that the set of trajectories is understood in the
Filippov sense.

3.4.3 Sliding Modes and Constraint Functions. Regularity


Condition
3.4.3.1 Denition of Regularity Condition
At this point, it is useful to briey introduce the denition of the regularity condition
and its relation with other concepts, such as the normal form of nonlinear systems.
Hence, reconsider the constraint given by s(x) = 0, where s : Rn R is a function
smooth enough. Assume also that the time derivatives of s(x), i.e. s , s, . . . , s (r1)
exist and are single-valued functions of x (which is not trivial in discontinuous dynamical systems). Recall that the discontinuity does not appear in the rst r 1
derivatives of the constraint function s, or analogously, s is an output of relative
degree r with respect to the discontinuous input, according with (3.9) and (3.10).
When these assumptions hold, the sliding set of order r will be unequivocally determined by Eqs. (3.40), implying that the reduced system dynamics has order n r.
Denition 3.1 Consider the non-empty rth-order sliding set (3.40) and assume that
it is a set locally integrable in the sense of Filippov (i.e. consisting of Filippov trajectories of the discontinuous dynamical system). Then, the corresponding motion
that satises (3.40) is called rth-order sliding mode with respect to the constraint
function s.
To show the relationship of this denition with other control denitions, consider
a manifold S given by the equation s(x) = 0. Suppose that s, s , s, . . . , s (r1) are
smooth functions of x and


rank s, s , s, . . . , s (r1) = r
(3.43)
holds locally. Then, since all Si , i = 1, . . . , r 1, are smooth manifolds, Sr is a
differentiable manifold determined by (3.40). Recall that the rank of a set of vectors
indicates the dimension of the subspace they dene.
Equation (3.43), together with the requirement that the corresponding time
derivatives of s are smooth functions of x, is referred to as the sliding regularity condition [25, 31].
This is a useful denition because if condition (3.43) is reached, new local coordinates y1 = s can be taken, and the system can be described through the following
set of equations:

52

Fundamentals of Sliding-Mode Control Design

y1 = y2

y2 = y3

..
.

yr = (y, )

= (y, ) with Rnr

(3.44)

Note that this is similar to the normal form of nonlinear systems; the only difference is that in the normal form yr = a(y) + b(y)u [31].
Additional Remark It is sometimes mentioned that the higher-order sliding
modes differ depending on the number of total derivatives of s which are extinguished when reaching the manifold S . However, this number cannot be considered as a feature of the HOSM, since formally all orders of derivatives are cancelled
at S [4]. The most important feature of a sliding mode is the number of successive
continuous derivatives of s in the neighbourhood of the manifold. In other words,
the value of r is taken from computing the rst discontinuous or non-existent time
derivative of s. The sliding order r is understood in this sense.

3.4.3.2 Connection with Other Well-Known Results in Control Theory


Let the control afne nonlinear system (3.1) be recalled as

x = f (x) + g(x)u
s = s(x) R

uR

(3.45)

with f , g and s sufciently smooth vector functions.


Assuming that the output s(x) has relative degree r, according to (3.9) and (3.10),
this means that in the neighbourhood of a given point,
Lg s = Lg Lf s = = Lg Lr2
f s = 0;

Lg Lr1
f s
= 0

(3.46)

hence s (i) = Lif s for i = 1, . . . , r 1, and the regularity condition (3.43) is automatically satised. For this reason, a direct analogy between the relative degree notion
and the regularity condition of sliding mode can be established. In general terms, it
can be stated that the regularity condition (3.43) means that the relative degree of
system output with respect to the discontinuity is at least r. Similarly, the notion of
rth-order sliding-mode dynamics is analogous to the zero dynamics concept dened
in [27]. The nominal stability of the controlled system can be guaranteed if the stability of (3.44) holds when y = 0, i.e. when the reduced system = (y, ) with
Rnr is stable.

3.4 Some General Concepts on Higher-Order Sliding Modes

53

3.4.4 Closing Comments on Higher-Order Sliding Modes in


Control Systems
3.4.4.1 An Observation Regarding the Accuracy of Real Sliding Modes
It is necessary to clarify that when referring to a system operating in sliding mode,
it can be both ideal (nominal) sliding, which takes place when the switching imperfections are neglected and the restriction is maintained accurately, or real sliding,
which occurs when the switching imperfections are taken into account. In the latter
case the restriction can be satised only approximately.
The quality of the control design is related to the sliding accuracy. It is worth
mentioning that in practice, there are no design methods that can ideally maintain
the desired constraint s(x) = 0. Therefore, there is a need to introduce some sort
of comparison between different control systems. Further details and proofs can be
found in [25].
Strictly speaking, any ideal sliding mode should be understood as the limit of
movements when the imperfections disappear and the switching frequency tends to
innity. Therefore, if is taken as a measure of these imperfections, the accuracy of
any sliding-mode control design can be characterised by its asymptotic behaviour
as 0 [29].
For example, to obtain a real sliding mode of order r (with discrete switching), it
is required to satisfy an order r of ideal sliding (at innite switching frequency). So,
most of the real second-order algorithms come from discretising ideal second-order
algorithms [17, 29].
A special discrete switching algorithm that provides second-order real sliding
was presented in [41]. Another example of a second-order real sliding controller is
the Drift Algorithm [29]. Moreover, a real third-order sliding controller that only
uses measures of s has been presented in [3].

3.4.4.2 HOSM Convergence Time


Prior to entering the section devoted to the design of specic Second-Order SlidingMode (SOSM) algorithms, a nal general comment concerning the convergence
time is of interest. Convergence in HOSM can be either asymptotic or in nite time.
Examples of asymptotically stable sliding-mode algorithms of arbitrary order are
well known in the literature [24]. On the other hand, fewer examples can be cited
for r-sliding controllers that converge in nite time. For instance, these can be found
for r = 1 (which is trivial), for r = 2 [2, 5, 29] and for r = 3 [24, 31]. Despite the
fact that some arbitrary-order sliding-mode controllers of nite-time convergence
have already been presented [33], its implementation is not yet widespread.

54

Fundamentals of Sliding-Mode Control Design

3.5 Design of Second-Order Sliding-Mode Controllers


3.5.1 Second-Order Sliding Generalised Problem
This nal section is focused on the design of SISO second-order sliding-mode controllers, aiming to explain the specic algorithms that will be used in this book.
To this end, consider the uncertain nonlinear system (initially, not necessarily
afne in the control), explicitly dened as

x = F (x, u, t)
(3.47)
s = s(x, t) R

u = U (x, t) R
with x Rn , u the single control input, and F and s smooth functions. Note that
in this section the possible direct dependence on t has been explicitly manifested in
system (3.47), in order to better explain the subsequent SOSM design procedure.
As always, the ultimate control objective would be steering the sliding output
s to zero. However, the SOSM approach enables not only that s = 0 and its time
derivative s = 0, but also nite time stabilisation of both, as long as s is of relative
degree 1 or 2 with respect to the control input u. Moreover, in the former case the
physical control action synthesised by the SOSM algorithm is continuous.
The SOSM design procedure depends on the bounds of certain functions that
constitute the second time derivative of the sliding output s. Hence, as a rst step,
s is differentiated twice, and the following general expressions are derived:

s(x, t) +
s(x, t)F (x, u, t)
t
x

(3.48)

s (x, t) +
s (x, t)F (x, u, t) +
s (x, t)u(t)

t
x
u

(3.49)

s =
s =

Then, two different cases will be addressed, depending on the relative degree of
s with respect to input u. Systems with relative degree 1 and relative degree 2 will
be considered, respectively.
Case 1 Systems with relative degree 1.
In relative degree 1 systems, u appears in s , thus in the expression of s the derivative u is explicitly presented in afne form, as in (3.49). Therefore expression (3.49)
can be given as follows:
s = (x, u, t) + (x, u, t)u(t)

(3.50)

with (x, u, t) and (x, u, t) uncertain but uniformly bounded functions in a


bounded domain. In order to specify the control problem, the following conditions
must be assumed [29]:

3.5 Design of Second-Order Sliding-Mode Controllers

55

1. There are bounds m and M such that within the region |s(x, t)| < s0 the following inequality holds for all t, x X , u U :
0 < m (x, u, t) =

s (x, t) M
u

(3.51)

The constant s0 denes a region of operation around the sliding manifold, where
the bounds are valid. Note that, eventually, an appropriate control action has to be
included in the controller, in order to attract the system into this validity region.
2. There is also a bound such that, within the region |s(x, t)| < s0 ,




(x, u, t) = s (x, t) + s (x, t).F (x, u, t)
(3.52)


t
x
for all t, x X , u U .
With these bounds at hand, the following differential inclusion can be proposed
to replace (3.50) [33]:
s [, ] + [m , M ]u

(3.53)

This is a very important relation when considering robustness. As it will be


demonstrated in the following subsection, many SOSM controllers ensure nitetime stabilisation of both s(x, t) = 0 and s (x, t) = 0, not merely for the nominal
original system, but for (3.53). Since inclusion (3.53) does not remember whether
or not the original system (3.47) is perturbed (it will include both cases, as far as
perturbations had been computed into the bounds), then such a controller will be
obviously robust with respect to any perturbation or uncertainty existing in (3.47)
and, consequently, translated to (3.50).
Case 2 Systems with relative degree 2.
In relative degree 2 systems, u is not present in s , hence the derivative u does not
appear in s (i.e. the third term of (3.49) is null), resulting in
s =

s (x, t) +
s (x, t)F (x, u, t)
t
x

(3.54)

In this case, we will limit the analysis to afne in the control nonlinear systems
of the form:
x = F (x, u, t) = f (x, t) + g(x, t)u(t)

(3.55)

Therefore, expressions (3.54) and (3.55) can be combined as follows:


s =  (x, t) +  (x, t)u(t)

(3.56)

Once again  (x, t) and  (x, t) are uncertain but uniformly bounded functions in a
bounded domain.

56

Fundamentals of Sliding-Mode Control Design

Fig. 3.6 System trajectory


on the plane (s, s )

As in the relative degree 1 case, analogous conditions must be assumed:


1. There are bounds m and M such that, within the region |s(x, t)| < s0 , the
following inequality holds for all t, x X , u U :
0 < m  (x, t) M
and s0 denes the region of validity around the sliding manifold.
2. There is also a bound  such that, within the region |s(x, t)| < s0 ,

 
 (x, t) 

(3.57)

(3.58)

for all t, x X , u U .
In this case, the following differential inclusion can be proposed to replace (3.56):
 


(3.59)
s  ,  + m , M u
Robustness considerations similar to the prior case can be established.

3.5.2 Solution of the Control Problem. SOSM Algorithms


There is a wide variety of proposals for second-order sliding-mode controllers that
provide solutions to the aforementioned problem of nite-time convergence and robust stability. In this section three of the most widely-known algorithms will be
reviewed. Two of them, the Twisting and Sub-Optimal algorithms, are devised for
relative degree 2 systems, and the other one, the Super Twisting algorithm, is for
relative degree 1 systems.

3.5.2.1 Twisting Algorithm


This is one of the rst SOSM mode algorithms, and it is primarily intended for
relative degree 2 systems. Once the initialisation phase is elapsed (i.e. the region

3.5 Design of Second-Order Sliding-Mode Controllers

57

Fig. 3.7 Expanded system

|s| < s0 is reached by using an appropriate extra control action), it generates system
trajectories that encircle the origin of the plane (s, s ) an innite number of times
(Fig. 3.6), converging to it in nite time [29].
Consider system (3.47) under the conditions of Case 2 of Sect. 3.5.1; these are
of relative degree 2 with respect to s and afne in the control form, as in (3.55). Additionally, the bounds dened in Case 2, conditions 1 and 2, exist and are available
for the designer. Then, the Twisting algorithm can be written as
u = r1 sign(s) r2 sign(s )

(3.60)

where r1 and r2 are the controller parameters, to be tuned based on the system
bounds. It will be demonstrated in the sequel that if they simultaneously satisfy the
conditions
r1 > r2 > 0
m (r1

+ r2 )  > M (r1 r2 ) + 
m (r1 r2 ) > 

(3.61)

then the Twisting controller (3.60) generates a second-order sliding mode that attracts the trajectories of the system to s = s = 0 in nite time [25].
Twisting Adaptation to Relative Degree 1 Systems Prior to proving the convergence of the Twisting algorithm, it would be useful to explain how to apply it to
relative degree 1 systems.
It is rather obvious that algorithms intended for relative degree 2 can straightforwardly be adapted for its implementation on relative degree 1 systems. The procedure would be as simple as to articially increase the relative degree to 2 by
expanding the system. To this end, an integrator is incorporated prior to input u, and
a new articial input is created (see Fig. 3.7) in accordance with the differential
equation below:
u = (t)

(3.62)

Now, u has become a new internal state variable (xi = u) of the following expanded system with articial or auxiliary input :
xe = Fe (xe , , t) = fe (xe , t) + ge (xe , t)(t)

(3.63)

58

Fundamentals of Sliding-Mode Control Design

with xe = [x T xi ]T Rn+1 , the new input of the expanded system, and Fe , fe and
ge smooth functions.
The expanded system (3.63) is then of relative degree 2 with respect to s and,
consequently, fulls the conditions required for the design of the Twisting algorithm.
Following the steps previously described, a Twisting SOSM control signal would be
synthesised for the relative degree 2 input:
= u = r1 sign(s) r2 sign(s )

(3.64)

To conclude, it is pertinent to remark on the bounds required for the design. It is


straightforward to infer, according with (3.56), that the expression of s in terms of
the expanded system becomes
s =  (xe , t) +  (xe , t)(t)

(3.65)

In turn, if the expression of s of the original system, i.e. the expression for relative
degree 1 systems (3.50), is also written in terms of the expanded system variables,
it results in
s = (x, u = xi , t) + (x, u = xi , t)u(t)

= (x, xi , t) + (x, xi , t)(t)

(3.66)

Comparing (3.65) and (3.66), it can be appreciated that in this case, in which the
relative degree 2 system comes from an articially expanded relative degree 1 system, the functions to be bounded are exactly the same to those of the latter (i.e.
 (xe , t) = (x, u, t) and  (xe , t) = (x, u, t)). So, in this particular case, this
would allow the use of the same bounds for designing either relative degree 1 or
2 SOSM controllers. In this way, the computations involved in the design procedure
of this type of systems are signicantly alleviated.
Proof of Convergence of the Twisting Algorithm Auxiliary System To analyse
the convergence of the algorithm, consider a suitable auxiliary system, namely a
double integrator s = u. This system has been chosen given that it will behave as a
majorant of the system under consideration (3.56), i.e. it will give the worst case
for the region of convergence in the phase plane (s, s ). Then, controlling the auxiliary system with a control law of the Twisting form u = b1 sign(s) b2 sign(s ),
yields
s =

d s d s
= s = b1 sign(s) b2 sign(s )
dt
ds

(3.67)

with constant parameters b1 and b2 satisfying b1 > b2 > 0.


Now, the task in this rst stage will be to show that the solutions of (3.67) converge to the origin (s = s = 0) in nite time. This result will be used at the end of
the subsection to prove the convergence of the uncertain system under study.

3.5 Design of Second-Order Sliding-Mode Controllers

59

Fig. 3.8 Vector eld

Analysing the differential equation it can be noticed that


b1 b2
if s > 0, s > 0

b
+b
d s 1s 2 if s > 0, s < 0
= b b
1
2
ds
if s < 0, s > 0

b1 +b2
if s < 0, s < 0
s

(3.68)

Then, the system vector eld will take the shape shown in Fig. 3.8.
Taking the initial conditions P1 = (0, s0 ), the solution of (3.67) for the rst quadrant can be found as follows:
 s
 s
s d s =
(b1 + b2 )ds
(3.69)
0

s1

1 2
s = (b1 + b2 )(s1 s)
2
s 2
2(b1 + b2 )
For the second quadrant, a similar expression can be found:

s = s1

(3.70)

(3.71)

s 2
(3.72)
2(b1 b2 )
Taking two xed points of this trajectory, P1 = (0, s0 ) and P2 = (0, s1 ) (see
Fig. 3.9), it can be concluded that
s = s1

s02 = s1 2(b1 + b2 )
s12 = s1 2(b1 b2 )

(3.73)

60

Fundamentals of Sliding-Mode Control Design

Fig. 3.9 Phase trajectory of


the Twisting algorithm

|s1 |
=
|s0 |

b1 b2
=q <1
b1 + b2

(3.74)

Extending this reasoning to the rest of the trajectory, it is veried that the following inequality always holds:
|si+1 |
=q <1
|si |

(3.75)

Therefore, the algorithm converges to the origin.


Time of Convergence Considering t1+ as the time that takes the auxiliary system
trajectories to go from the point s0 to s1 . Integrating Eq. (3.67) in the rst quadrant,
it results in
 t
 s
d s =
(b1 + b2 )dt
(3.76)
s0

Knowing that

s (t1+ ) = 0,

s (t) = (b1 + b2 )t + s0

(3.77)

we get that
t1+ =

s0
b1 + b2

(3.78)

The computation of t1 (time that takes the system to go from s1 to s1 ) is similar


to the previous step:
s (t) = s1

s 2
[(b1 b2 )t]2
(b1 b2 )t 2
= s1
= s1
2(b1 b2 )
2(b1 b2 )
2

(3.79)

3.5 Design of Second-Order Sliding-Mode Controllers

61

Taking into account that s(t1 ) = 0 and using (3.73), we can state that


2s
1
1
t1 =
s0
=
b1 b2
(b1 b2 )(b1 + b2 )

(3.80)

This means that the time interval t1 of the trajectory s0 s1 s1 is


t = t1+ + t1 = s0
where

(3.81)


1
=
+
b1 + b2

1
(b1 b2 )(b1 + b2 )

(3.82)

More generally, we can state that


ti = |si1 | = q i1 s0

(3.83)

So the total convergence time of the auxiliary system is given by


T=

ti =

i=1

|si1 | =

i=1

q i1 s0 =

i=1

s0
1q

(3.84)

In the sequel, it will be demonstrated that, with an adequate selection of b1


and b2 , the auxiliary system effectively acts as a majorant of (3.56) and (3.84) is
a bound for its convergence time.
Convergence of the Uncertain System Once the convergence of the auxiliary
system has been demonstrated, the last stage consists of using this result to prove
the convergence of the relative degree 2 uncertain system under study.
To this end, reconsider the differential equation
s =  (x, t) +  (x, t)u
with the aforementioned bounds

 
 (x, t)  ,

0 m  (x, t) M

(3.85)

(3.86)

and the differential inclusion


 


s  ,  + m , M u

(3.87)

provided that the solutions of (3.85) are understood in the Filippov sense.
The Twisting law (3.60) with r1 and r2 simultaneously satisfying
r1 > r2 > 0
m (r1

+ r2 )  > M (r1 r2 ) + 
m (r1 r2 ) > 

is used to control the system.

(3.88)

62

Fundamentals of Sliding-Mode Control Design

Fig. 3.10 Twisting algorithm


and majorant trajectories

Then, to prove that the auxiliary system (3.67) is a majorant for (3.85) controlled
with the Twisting algorithm, the following appropriate selection for constants b1 >
b2 > 0 is proposed:
b1 + b2 = m (r1 + r2 )  ,

b1 b2 = M (r1 r2 ) + 

(3.89)

With this parameter selection, trajectories s0 s1 s1 and s0 sM sM in Fig. 3.10 correspond to (3.85) and (3.67), respectively, with common initial conditions s = 0 and
s = s0 > 0 for t = 0.
In particular, s0 sM sM is the so-called majorant curve of the system. One of its
points, P (sp , sp ), in the rst quadrant is considered. If the trajectories of the system
(3.85) pass through this point, controlled by (3.60), the system velocity would have
the coordinates (sp , sp ). Note that the horizontal component of velocity (sp ) is the
y-coordinate of the point P (positive value).
On the other hand, using the control (3.60) in system (3.87) on the rst quadrant,
the following will be fullled:

 

s  ,  + m (r1 r2 ), M (r1 r2 )
(3.90)
Moreover, due to the fact that inequalities (3.88) hold, the vertical component is
kept within the following limits:
 M (r1 + r2 ) s  m (r1 + r2 ) < 0

(3.91)

This implies that the velocity of system (3.85), (3.60) at P (sp sp ), will always
point to the interior of the region bounded by the axes s = 0, s = 0 and the surrounding curve (3.67), (3.89). Then, the trajectory of system (3.85), (3.60) intersects
+
the axis s = 0 in a point s1 sM in nite time t1+ tM
.

3.5 Design of Second-Order Sliding-Mode Controllers

63

Considering the trajectories s1 s1 and sM sM of systems (3.85), (3.60) and (3.67)


in the second quadrant (s 0, s 0), a point C(sC , sC ) where similar properties are
veried can be taken. As it can be inferred from inequality (3.88), in this quadrant
the module of the vertical component of velocity vector (3.85), (3.60) is smaller
than the surrounding system (3.67), while the horizontal component is sC :
 M (r1 r2 ) s  m (r1 r2 ) < 0

(3.92)

This means that the trajectories of system (3.85), (3.60) are inside the surrounding system. On the other hand, the time required to cover the vertical segment (0, s1 )
is the same, but the surrounding trajectory must also cover the vertical segment

s1 , sM . This is the reason why t1 tM


, where t1 is the time it takes the system to

cover the trajectory s1 s1 , and tM is that for sM sM .


Then, the trajectory of system (3.85), (3.60) intersects s = 0 at s1 . Moreover,
|s1 | |sM |, and therefore the evolution time of the uncertain system is bounded by
the majorant as t1 tM .

3.5.2.2 Super Twisting Algorithm


The Super Twisting algorithm, one of the most widely used algorithms of the family,
is particularly intended for systems with relative degree 1 [25, 29]. Highly suitable
for real implementation, with a proper choice of parameters, this algorithm converges in nite time after describing a trajectory similar to the one of the Twisting
algorithm (Fig. 3.11). The most distinctive features of the Super Twisting algorithm
are the aforementioned direct applicability to relative degree 1 systems, the synthesis of continuous control actions and the absence of a measurement of s in the
control law. This makes it more immune to output measurement noise and possible
errors in the estimation of s .
The control action u(t) of the Super Twisting algorithm is composed of two continuous terms, even though the rst one is given by the integral of a discontinuous
action. Once the validity region |s| < s0 is attained, with the help of an appropriate
reaching control action, the Super Twisting control is given by
u(t) = u1 (t) + u2 (t)

(3.93)

with
u 1 (t) = sign(s)
u2 (t) = |s| sign(s)

(3.94)

where > 0, > 0 and (0, 1/2] are the parameters of the controller. The restrictions for their design are based on the bounds dened in Case 1 and conditions 1
and 2 in Sect. 3.5.1. The following are sufcient conditions for convergence in nite

64

Fundamentals of Sliding-Mode Control Design

Fig. 3.11 Phase trajectory of


the Super Twisting algorithm

time to the sliding manifold [25, 29]:

m
2 (m + )2
2 > 2
m (m )
>

(3.95)
if = 0.5

Note that if = 1 and and / are large enough, it can be even proven that
there will be a stable second-order sliding mode. In this case, |s| + |s | would tend
to zero with exponential upper and lower bounds.
Super Twisting Adaptation to Relative Degree 2 Systems In similar way that
an algorithm intended for relative degree 2 can be adapted for its implementation on
relative degree 1 systems, the Super Twisting algorithm can be adapted for application to relative degree 2 systems. In this case a differentiator should be incorporated
into the system. This is not at all a specic subject matter of this book, and therefore it will not be addressed here. Nevertheless, to give an idea of the procedure,
consider that the articial input is created in this case such that u = , i.e. differentiating the new input instead of integrating it (as it was done in the Twisting
analogous case). Then, substituting for u in the expression of s for the relative
degree 2 system (3.55), it would become afne with respect to the derivative of the
input, specically, the articial input . Displaying that form, the differential inclusion proposed for relative degree 1 is applicable to (3.56), hence the standard Super
Twisting algorithm could be used to synthesise the control signal for the articial
input . The actual implementation method and the proper use of differentiators are
much more elaborate and, as previously stated, far exceed the scope of this book.
The interested reader is strongly encouraged to read the specialised literature on this
topic (e.g., [4, 32]).

3.5 Design of Second-Order Sliding-Mode Controllers

65

Proof of Convergence of the Super Twisting Algorithm Consider now the controlled process (3.47) with output s of relative degree 1 with respect to u. As was
said, if the control problem satises conditions (3.51), (3.52), the system solutions
can be understood in the sense of Filippov, and Eq. (3.50) can be replaced by the
differential inclusion (3.53).
Then, substituting (3.94) into the differential inclusion (3.53), the overall system
performance and the majorant curves that limit the evolution of the system trajectories can be evaluated. Consider the case where |s| < |s0 | and the trajectory of the
system is within the rst quadrant (s > 0 and s > 0):

(3.96)
s [, ] + [m , M ] s 1 s
Due to the fact that in this quadrant s > 0, in order to decrease the value of s and
ensure that the system trajectories cross s = 0, the condition s < 0 must be achieved
in the entire quadrant. The worst possible scenario is when (x, t) = (maximum
positive value that the vector eld can take) and (x, t) = m (lower dominance
of control in the system dynamics):

(3.97)
s = + m s 1 s < 0
To keep the sign of s negative, even when s 1 s 0, the relation m <
0 must be satised. This imposes the rst convergence condition of the algorithm:
>

(3.98)

To improve clarity in the analysis, the second convergence condition of the algorithm will be obtained with the help of the following change of notation: s = z1 and
s = z2 . Then, the dynamics of the planar system in the rst quadrant is given by

z 1 = z2

(3.99)
1
z 2 = + z1 z2
Considering the worst-case scenario for this quadrant ( = and = m ), system (3.99) will have a limit trajectory in the solutions of the following planar system:

z 1 = z2

(3.100)
z 2 = + m (z1 )1 z2 +
The analytical solutions of this nonlinear system cannot be straightforwardly
found, but numeric tools can be used to predict the solution from an initial condition.
Figure 3.12 presents the results of a numerical evaluation of the limit trajectories of
system (3.99) with a set of parameters arbitrarily chosen as an example ( = 10,
m = 1, M = 1.7, = 50, = 15 and = 0.5).
The case of slower decrease of s takes place as s 1 s 0. Thus, in the rst
quadrant, the majorant curve is governed by the following expression:
s = m

(3.101)

Integrating successively this equation, the following general expression with initial conditions (0, s0 ) can be found, which represent the majorant curve of the sys-

66

Fundamentals of Sliding-Mode Control Design

Fig. 3.12 Example of convergence of the Super Twisting algorithm

tem trajectories in the rst quadrant:


s = (s s0 )2

1
s0
+ (s s0 )
2K
K

(3.102)

with
K = m < 0

(3.103)

In the example of Fig. 3.12, the external line shows the majorant curve of the
system trajectories (3.99), with the corresponding parameters listed above. It is important to note that, although the trajectories can be closer to the majorant curve at
certain points, the majorant curve does not represent a possible path of the system,
which would imply s 1 s = 0 for all (s, s ).
Using an analogous analysis on the fourth quadrant, similar equations for the
system dynamics can be found:

z 1 = z2

(3.104)
z 2 = + (z1 )1 z2 +
In this case, there are no restrictions on the sign of s , but it can be known from
the continuity of trajectories that, in the rst section, s < 0. Then, from an isocline
analysis, the area where s is zero can be determined. Taking s = 0 in the worst-case
scenario of the quadrant ( = and = m ), we obtain that
s = s 1

(3.105)

with
=

+ m
m

(3.106)

3.5 Design of Second-Order Sliding-Mode Controllers

67

Therefore, when crossing the curve (3.105), (3.106), the system trajectories will
do it with s = 0, yielding a local minimum at s and a change of sign in s (see
Fig. 3.12). Since in this quadrant s < 0, the function s(x) decreases monotonically.
This ensures an intersection point with the curve (3.105), (3.106). To dene a majorant curve of the system, the minimum possible value of s can be analysed, which
is determined by the value taken by (3.105), (3.106) when s = s (see Fig. 3.12):

1
s02
+ m
1
s = s =

(3.107)
m
2( m )
Finally, to ensure the algorithm convergence to the origin, it is necessary to satisfy that |s | < |s0 |:

1
s02
|s | + m
1
=
<1
(3.108)
|s0 |
m
2(m )
s0
If = 0.5, this condition is reduced to
2

2 (m + )2
m2 (m )

(3.109)

For the case where |s| > s0 , a simplied expression similar to (3.102) can be
found, which ensures that the controlled system trajectories arrive at the area |s| < s0
in nite time.

3.5.2.3 Sub-Optimal Algorithm


Similarly to the Twisting algorithm, this SOSM controller is primarily intended for
relative degree 2 systems. It was developed as a Sub-Optimal feedback implementation of the classical time-optimal control for the uncertain double integrator problem [2].
In this case, the system trajectories on the plane (s, s ) are conned within limit
parabolic arcs that include the origin. So both twisting and leaping behaviours are
possible. A most important feature is that the coordinates of the successive trajectory
intersection with axis s decrease in geometric progression (see Fig. 3.13).
Given that the Sub-Optimal algorithm is a controller primarily intended for relative degree 2 systems, we consider similar conditions to the ones established for
Twisting algorithm (i.e., conditions of Case 2 of Sect. 3.5.1). Then, once the region
of validity is reached, the Sub-Optimal algorithm is dened by the following control
law:
u = (t)U sign(s sM )

1
if (s sM )sM 0
(t) =

if (s sM )sM < 0

(3.110)

where > 1 is the so-called modulation factor, 0 < 1 is the anticipation factor,
and U > 0 is the minimum control magnitude. sM corresponds to the last singular

68

Fundamentals of Sliding-Mode Control Design

Fig. 3.13 Phase trajectory of


the Sub-Optimal algorithm

value of the function s, i.e. the last value of s at the time its derivative s reaches
zero (see Fig. 3.13). This value must be refreshed every time s zeroes, thus the SubOptimal algorithm is particularly suitable for plants that incorporate a special device
or software code to detect the singular values (sM ) (i.e. peak detector).
The parameters of the controller, , and U , should be tuned based on the
bounds dened in conditions 1 and 2 of Case 2, Sect. 3.5.1. The sufcient conditions
for the nite-time convergence and robust stabilisation of the system are
U>


m


 + (1 )M U
; +
[1; +)
m U

(3.111)

which are known as the dominance condition and the convergence condition, respectively. The former ensures that the control has sufcient authority to affect the
sign of s . The latter guarantees the sliding-mode stability and determines the rate
of convergence. A detailed analysis and proof of convergence can be found in [2]
and [5].
Sub-Optimal Algorithm Adaptation to Relative Degree 1 Systems Similarly to
the Twisting algorithm case, the application of the Sub-Optimal algorithm to relative
degree 1 systems can be achieved by articially increasing the relative degree, with
the incorporation of an integrator prior to the system input u. The procedure to be
followed is exactly the same as that described in the Twisting algorithm Sect. 3.5.2.1.

3.6 Conclusions

69

3.6 Conclusions
Fundamentals of sliding-mode control have been introduced in this chapter. This
control theory has proven to be capable of successfully dealing with nonlinear systems, presenting several attractive characteristics. Among them, nite convergence,
system order reduction and robustness against certain disturbances are the most relevant. In this context, the extension known as Higher-Order Sliding Modes adds
chattering reduction to the list of positive features, improving accuracy in realisation and, in several plants, contributing to extending the service life of the actuators.
These, together with relatively low on-line computational cost, make the HOSM
technique specially suitable for implementation.
To this end, several algorithms have been developed, particularly the SecondOrder Sliding-Mode ones. In this chapter, three of the most widely used SOSM
controllers have been reviewed, namely Twisting, Super Twisting and Sub-Optimal.
There are many others that robustly solve the problem of convergence, each one with
its own advantages and features (e.g., Drift algorithm [18, 29], Global algorithm [5]
and Prescribed convergence law algorithm [29]).
Then, the challenge now is to tackle the specic problem of PEM fuel cell control
and assess the applicability of this control technique, with the objective of enhancing
fuel cell efciency and increasing their service life. This will be the topic of the next
chapter.

References
1. Barbot JP, Perruquetti W (eds) (2002) Sliding mode control in engineering. Dekker, New
York
2. Bartolini G, Ferrara A, Usai E (1998) Chattering avoidance by second order sliding mode
control. IEEE Trans Autom Control 43(2):241246
3. Bartolini G, Levant A, Pisano A, Usai E (1999) 2-Sliding mode with adaptation. In: Proceedings of the 7th IEEE Mediterranean conference on control and systems, Haifa, Israel
4. Bartolini G, Levant A, Pisano A, Usai E (2002) Higher-order sliding modes for outputfeedback control of nonlinear uncertain systems. In: Variable structure systems: towards the
21st century. Lecture notes in control and information sciences, vol 274, Springer, Berlin,
pp 83108 (Chap 6)
5. Bartolini G, Pisano A, Punta E, Usai E (2003) A survey of applications of second-order sliding
mode control to mechanical systems. Int J Control 76(9/10):875892
6. Bartolini G, Fridman L, Pisano A, Usai E (eds) (2008) Modern sliding mode control theory.
New perspectives and applications. Lecture notes in control and information sciences, vol 375.
Springer, Berlin
7. Boiko I, Fridman L (2005) Analysis of chattering in continuous sliding mode controllers.
IEEE Trans Autom Control 50(9):14421446
8. Boiko I, Fridman L, Pisano A, Usai E (2007) Analysis of chattering in systems with secondorder sliding modes. IEEE Trans Autom Control 52(11):20852102
9. Bondarev A, Bondare S, Kostyleva N, Utkin V (1985) Sliding modes in systems with asymptotic state observers. Autom Remote Control 46(5):679684
10. Burton J, Zinober A (1986) Continuous approximation of variable structure control. Int J Syst
Sci 17:875885

70

Fundamentals of Sliding-Mode Control Design

11. Cortes J (2008) Discontinuous dynamical systemsa tutorial on solutions, nonsmooth analysis, and stability. IEEE Control Syst Mag 28(3):3673
12. DeCarlo R, Zak S, Matthews G (1988) Variable structure control of nonlinear multivariable
systems: a tutorial. Proc IEEE 76(3):212232
13. Drazenovic B (1969) The invariance conditions in variable structure systems. Automatica
5:286295
14. Edwards C, Spurgeon S (1998) Sliding mode control: theory and applications, systems and
control book series. Taylor & Francis, London
15. Edwards C, Fossas Colet E, Fridman L (eds) (2006) Advances in variable structure and sliding
mode control. Lecture notes in control and information sciences, vol 334. Springer, Berlin
16. Emelyanov S (1957) Variable structure control systems. Nauka, Moscow
17. Emelyanov S, Korovin S (1981) Applying the principle of control by deviation to extend the
set of possible feedback types. Sov Phys Dokl 26(6):562564
18. Emelyanov S, Korovin S, Levantovsky L (1986) Higher order sliding regimes in the binary
control systems. Sov Phys Dokl 31(4):291293
19. Filippov A (1960) Differential equations with discontinuous right-hand side. Mat Sb (in Russian). English translation: Trans. Am Math. Soc. 62 (1964)
20. Filippov A (1988) Differential equations with discontinuous righthand sides. Mathematics and
its application. Kluwer Academic, Dordrecht
21. Floquet T, Barbot J-P, Perruquetti W (2003) Higher-order sliding mode stabilization for a
class of nonholonomic perturbed systems. Automatica 39(6):10771083
22. Fridman L (2001) An averaging approach to chattering. IEEE Trans Autom Control
46(8):12601265
23. Fridman L (2003) Chattering analysis in sliding mode systems with inertial sensors. Int J
Control 76(9/10):906912
24. Fridman L, Levant A (1996) Higher order sliding modes as a natural phenomenon in control
theory. In: Robust control via variable structure and Lyapunov techniques. Lecture notes in
control and information sciences, vol 217. Springer, London, pp 106133 (Chap 7)
25. Fridman L, Levant A (2002) Higher order sliding modes. In: Sliding mode control in engineering. Dekker, New York, pp 53101 (Chap 3)
26. Hung JY, Gao W, Hung JC (1993) Variable structure control: a survey. IEEE Trans Ind Electron 40(1):222
27. Isidori A (1995) Nonlinear control systems. Springer, Berlin
28. Khalil H (2002) Nonlinear systems, 3rd edn. Prentice Hall, New York
29. Levant A (1993) Sliding order and sliding accuracy in sliding mode control. Int J Control
58(6):12471263
30. Levant A (1998) Arbitrary-order sliding modes with nite time convergence. In: Proceedings
of the 6th IEEE Mediterranean conference on control and systems, Alghero, Sardinia, Italy.
31. Levant A (2001) Universal SISO sliding-mode controllers with nite-time convergence. IEEE
Trans Autom Control 46(9):14471451
32. Levant A (2003) Higher order sliding modes, differentiation and output feedback control. Int
J Control 76(9):924941
33. Levant A (2005) Homogeneity approach to high-order sliding mode design. Automatica
41:823830
34. Levant A (2007) Construction principles of 2-sliding mode design. Automatica 43(4):576
586
35. Plestan F, Glumineau A, Laghrouche S (2008) A new algorithm for high order sliding mode
control. Int J Robust Nonlinear Control 18(45):441453
36. Sabanovic A, Fridman L, Spurgeon S (eds) (2004) Variable structure systems: from principles
to implementation. IET control engineering series. IET, London
37. Shtessel Y, Shkolnikov I, Brown M (2003) An asymptotic second-order smooth sliding mode
control. Asian J Control 5:498503
38. Sira-Ramirez H (1988) Differential geometric methods in variable structure control. Int J Control 48(5):13591390

References

71

39. Slotine J (1991) Applied nonlinear control. Prentice Hall, New York
40. Sontag E (1998) Mathematical control theory: deterministic nite dimensional systems.
Springer, Berlin
41. Su W, Drakunov S, Ozguner U (1994) Implementation of variable structure control for
sampled-data systems. In: Proceedings of IEEE workshop on robust control via variable structure and Lyapunov techniques, Benevento, Italy, pp 166173
42. Utkin V (1978) Sliding modes and their application in variable structure systems. Mir,
Moscow
43. Utkin V (2002) First stage of VSS: people and events. Springer, Berlin, pp 133 (Chap 1)
44. Utkin V, Gulder J, Shi J (1999) Sliding mode control in electro mechanical systems. Taylor &
Francis, London
45. Young K, Utkin V, Ozguner U (1999) A control engineers guide to sliding mode control.
IEEE Trans Control Syst Technol 7(3):328342
46. Yu X, Xu JX (eds) (2002) Variable structure systems: towards the 21st century. Springer,
Berlin

Chapter 4

Assessment of SOSM Techniques Applied


to Fuel Cells. Case Study: Electric Vehicle
Stoichiometry Control

4.1 Introduction
As it was introduced in Chap. 2, the design of procient control algorithms for the
fuel-oxidant coordination problem in PEM fuel cells under variable load operation
can reduce the fuel consumption while preventing performance deterioration, oxygen starvation and, eventually, irreversible damage in the polymeric membranes.
However, the stoichiometry problem is complex and cannot be successfully tackled
by using traditional techniques. A number of reasons turn fuel cell stacks into a major control challenge, e.g. high-order nonlinear equations are required to describe
hydrogen-air fed stack dynamics, experimental data-based models usually incorporate look-up tables and piecewise functions, as well as a wide variety of model uncertainties [8]. Besides, many internal variables are inaccessible for measurement,
and there exist disturbances that affect the system operation. It is clear that, in order to attain an efcient controller for the FC system, a special control technique
capable of coping with such challenges is required.
Therefore, the main objective of this chapter is to evaluate the feasibility of
SOSM techniques for oxygen stoichiometry control in PEM fuel-cell-based systems. According to the features reviewed in Chap. 3, SOSM algorithms are especially suited candidates to deal with the aforementioned challenges. In general
terms, they are capable of solving the nonlinear robust stability of the fuel cell system, converging to the reference in nite time and avoiding chattering effects, due to
the absence of direct discontinuous components in the rst derivative of the output.
Moreover, if the system is of relative degree 1, then the control action applied to the
actual input of the plant will be continuous. Another feature of relevance for fuel
cell applications is the control design based on nonlinear models. Given the highly
nonlinear nature of the stack, this feature guarantees robust operation and performance in a wider range than those achieved by control methods based on model
linearization [8, 9]. In addition, the SOSM control laws are relatively simple, relying on a reduced set of parameters and few measurable variables, rather than on
the knowledge of the full state vector. Thus, the implementation requirements and
on-line computational burden are considerably alleviated.
C. Kunusch et al., Sliding-Mode Control of PEM Fuel Cells,
Advances in Industrial Control,
DOI 10.1007/978-1-4471-2431-3_4, Springer-Verlag London Limited 2012

73

74

Assessment of SOSM Techniques Applied to Fuel Cells

The assessment of the SOSM controllers suitability is performed here through


computer simulations, using a benchmark model of a fuel cell system for automotive applications. To this end, the model developed by J. Pukrushpan et al. at the
Mechanical Department of the Michigan University [8] has been selected, since it
is one of the most complete and accurate models available in the open literature
(Sect. 4.2). Therefore, the case study under consideration in this chapter is a 75-kW
PEM FC stack where the oxygen is fed by an air turbo compressor used in the Ford
P2000 fuel cell electric vehicle [8].

4.2 Succinct Description of the Electric Vehicle Fuel Cell System


The rst stage in the controllers design procedure is to reformulate the model description according to the design requirements of the SOSM control techniques.
A nonlinear state model of the form x = F (x, u, t) with F a smooth vector function,
at least C 1 , is required (see Sect. 4.3). The original benchmark model comprises
look-up tables and switched piecewise functions, continuous but not differentiable.
Hence, in the proposed design model, they are replaced with smooth functions using adequate polynomial approximations. Additionally, an order reduction based on
control considerations of the system physics can be done. Any minor degree of error
or uncertainty generated by the aforementioned model adaptation procedure can be
added to the inherent uncertainty of the original model. So, such lumped uncertainties, together with external disturbances, are taken into account in the design of the
proposed controllers.
Then, the aim of this section is twofold: on the one hand, to present a sixth-order
nonlinear state space model of the FC ow system in the SOSM control design
form; on the other hand, to review its constituent subsystems, giving a mathematical
characterization in accordance with the approach followed for the model construction in [7, 8]. This will be particularly useful to readers unacquainted with such
models. However, it is not the intention of this section to give an exhaustive description, but to briey provide the necessary background to understand the subsequent
developments in this book. Therefore, the readers interested in a detailed description and discussions on the benchmark model are referred to the original works of
Pukrushpan et al.
The fuel cell generation system (FCGS) roughly comprises ve main subsystems: the air ow (breathing), the hydrogen ow, the humidity, the stack electrochemistry and the stack temperature subsystem, respectively. In [8], it is assumed
that the input reactant ows are efciently humidied and the stack temperature is
well regulated by dedicated controllers. In addition, it is considered that sufcient
compressed hydrogen is available, and therefore the main attention is focused on
the air management. In Fig. 4.1 a schematic view of the FCGS under consideration is represented. The most relevant components related to the FC ow system are
succinctly characterized in the sequel.

4.2 Succinct Description of the Electric Vehicle Fuel Cell System

75

Fig. 4.1 Schematic diagram of a PEM fuel-cell-based generation system (FCGS)

4.2.1 Air Compressor


A 14-kW Allied Signal air turbo compressor is considered in the case under study.
Its control input is the command voltage of the DC motor (Vcp ), while the shaft
rotational speed is the only dynamic state variable. The model is divided into two
parts. One part is characterized by a static compressor map, which determines the
air ow rate through the compressor. The compressor air ow Wcp is computed
using the rotational speed and the pressure ratio across the compressor. The inlet
air is atmospheric, so constant pressure and temperature (Pamb = 1 atm, Tamb =
25 C) are assumed. The other part represents the dynamics of the air compressor
subsystem. Neglecting the armature inductance, a dominant rst-order mechanical
dynamic equation can be obtained by applying Newtons second law to the rotating
parts, assuming lumped rotational parameters:
dcp
1
= (Te Tl )
(4.1)
dt
J
where cp is the compressor rotational speed, J is the combined inertia of the compressor motor and the compression device, Te is the compressor motor torque, and
Tl the load torque. Te can be computed using the simplied DC motor equation
dc kdct (Vcp kdcv cp )
(4.2)
Rdc
the compressor DC motor constants, and dc its mechanical
Te =

with kdct , kdcv and Rdc


efciency.

76

Assessment of SOSM Techniques Applied to Fuel Cells

The load torque Tl depends on the compressor speed, the supply manifold pressure Psm , the compressor efciency (cp ) and the compressor air ow (Wcp ). Its
expression can be obtained from turbine thermodynamics equations given by [2, 3]
Tl =

Cp Tamb
cp cp



Psm
Patm

 1


1 Wcp

(4.3)

with Cp the specic heat capacity of air, and the ratio of the specic heats of
air. Wcp is rst given as a ow map determined from the pressure ratio across the
compressor and the motor shaft speed. Then, the look-up table is replaced by an
interpolation given by the Jensen and Kristensen nonlinear curve tting method [6].
Thus, in the current model, Wcp is given in the form of a continuously differentiable
bivariate function of Psm and cp . In turn, the compressor efciency was originally
represented as a double input look-up table depending on Wcp and Psm . However,
it has been proved that the use of the mean value of cp is a sufciently good approximation for modeling and control design purposes. Under these conditions, the
right-hand side of (4.1) is in the whole operation range, at least, a C 1 function.

4.2.2 Air Supply Manifold


The model of the air supply manifold takes into consideration the piping between the
compressor and the stack (cooler and humidier included). Two dynamic equations
are required. The rst one is given by the mass conservation principle:
dmsm
= Wcp Wsm
(4.4)
dt
where msm is the mass of air accumulated in the supply manifold, and Wcp and Wsm
are the manifold inlet and outlet mass ows, respectively. The former comes directly
from the compressor subsystem, while the latter, assuming a small pressure gradient
between the cathode and the manifold, can be taken as
Wsm = Ksm (Psm Pca )

(4.5)

with Ksm the supply manifold constant or restriction.


The second dynamic equation results from applying the energy conservation
principle and the ideal gas law. Assuming that the air temperature changes inside
the manifold, the pressure dynamic equation is given by
dPsm
Ra
=
(Wcp Tcp Wsm Tsm )
dt
Vsm

(4.6)

with Ra the gas constant of air, Vsm the supply manifold volume, and Tsm the manifold air temperature that can be computed using (4.4), (4.6) and the ideal gas law:
Tsm =

Psm Vsm
msm Ra

(4.7)

4.2 Succinct Description of the Electric Vehicle Fuel Cell System

77

4.2.3 Air Humidier and Temperature Conditioner Subsystems


Typically PEM fuel cell systems require air temperature and humidity conditioning
before entering the cathode. According to [8], in this electric vehicle FC stack, the
temperature of the air entering the supply manifold is high (>90 C), and thus it
has to be cooled to prevent damage to the fuel cell membranes. Then, it is assumed
that an ideal cooler conditions the air temperature to the stack operating temperature
(80 C), without producing an appreciable pressure drop.
Additionally, the cathode air must have a high humidity level to maintain the
hydration of the polymeric membranes, but excessive water amounts should be
avoided. Dry membrane and ooded fuel cells are both undesirable situations that
produce an efciency reduction and may cause irreversible damage. Hence, in this
system a humidier injects the proper amount of water into the stream, to adjust
the cathode inlet ow relative humidity to stay close to 100%. It is assumed that
the temperature of the air ow is constant and the humidier volume is negligible
compared to the supply manifold volume.
In order to compute the change in air humidity, due to the additional injection
of water, a static model was used. Based on the air ow conditions released by the
cooler and by using thermodynamic equations, the mass ows of dry air and vapor
can be determined (please refer to [7]). In particular, the vapor saturation pressure
of gases is calculated from the ow temperature using the following equation:
log10 (Psat ) = 1.69 1010 T 4 + 3.85 107 T 3
3.39 10

T + 0.143T 20.92
2

(4.8)
(4.9)

where the units of Psat are [kPa] and [K] for T .

4.2.4 Cathode Channels


The cathode model represents the lumped volume of the cathode channels of the 381
stacked cells. Using the mass conservation principle and thermodynamic properties
of the air, the model has been developed under the following assumptions: (1) gases
are ideal, (2) the stack and the cathode ow temperatures are constant and uniformly
distributed (80 C), (3) the ow rate of liquid water leaving the cathode is zero, and
(4) temperature, pressure, humidity and oxygen mole fraction are equal inside and
exiting the cathode. Then, the mass balance of oxygen, nitrogen and water inside
the cathode gives
dmO2 ,ca
= WO2 ,ca,in WO2 ,ca,out WO2 ,react
dt
mO2 ,ca : mass of O2 inside the cathode.
WO2 ,ca,in : mass ow rate of O2 entering the cathode.
WO2 ,ca,out : mass ow rate of O2 leaving the cathode.

(4.10)

78

Assessment of SOSM Techniques Applied to Fuel Cells

WO2 ,react : rate of O2 reacted.


dmN2 ,ca
= WN2 ,ca,in WN2 ,ca,out
dt

(4.11)

mN2 ,ca : mass of N2 inside the cathode.


WN2 ,ca,in : mass ow rate of N2 entering the cathode.
WN2 ,ca,out : mass ow rate of N2 leaving the cathode.
dmw,ca
= Wv,ca,in Wv,ca,out Wv,ca,gen + Wv,mem
dt

(4.12)

mw,ca : mass of water inside the cathode.


Wv,ca,in : mass ow rate of vapor entering the cathode.
Wv,ca,out : mass ow rate of vapor leaving the cathode.
Wv,ca,gen : rate of vapor generated in the fuel cell reaction.
Wv,mem : mass ow rate of water transferred across the membrane.
The partial pressures of the oxygen, nitrogen and vapor inside the cathode channels can be computed using the ideal gas law for each species.
At this point, an additional analysis can be performed to simplify the cathode
model. After a thorough simulation study, it can be established that in this modeled
system an important amount of water is transferred across the membrane and, consequently the cathode gas is saturated at almost every operating condition. Therefore,
the dynamics of the mass of water inside the cathode (mw,ca ) can be neglected, turning Eq. (4.12) into an algebraic relationship. Thus, the model dynamics is reduced
by one order [4].

4.2.5 Anode Channels


Several assumptions are made in [8] to develop the model of the anode subsystem.
The most relevant are: (1) a compressed hydrogen tank and a valve instantaneously
ensure sufcient hydrogen ow for the fuel cell reaction, (2) the stack and the anode
ow temperatures are the same, (3) one lumped volume is considered, and (4) temperature, pressure and humidity of the hydrogen outlet ow are similar to those of
the gas inside the anode channels. In the same way as in the cathode, the gas humidity and partial pressures are computed by balancing hydrogen and water mass ows
inside the anode.
dmH2 ,an
= WH2 ,an,in WH2 ,an,out WH2 ,react
dt
mH2 ,an : mass of H2 inside the anode.
WH2 ,an,in : mass ow rate of H2 entering the anode.
WH2 ,an,out : mass ow rate of H2 leaving the anode.

(4.13)

4.2 Succinct Description of the Electric Vehicle Fuel Cell System

79

WH2 ,react : rate of H2 reacted.


dmw,an
= Wv,an,in Wv,an,out Wv,mem
dt

(4.14)

mw,an : water mass inside the anode.


Wv,an,in : mass ow rate of vapor entering the anode.
Wv,an,out : mass ow rate of vapor leaving the anode.
Wv,mem : mass ow rate of water transfer across the membrane.
In this model the dynamics of the mass of water (mw,an ) can also be neglected,
given that in the model it only affects the hydration of the membrane, which is
assumed to be 100% humidied. Then, Eq. (4.14) becomes a static relationship,
reducing another order of the model dynamics.

4.2.6 Water Model of the Polymeric Membrane


The hydration model of the membrane allows the computation of the membrane
water content and the mass ow rate of water transfer across the membrane. Both
are functions of the stack current and the relative humidity of the cathode and anode
ows. The dynamics involved in the PEM subsystem are considerably faster than
those of the gas channels and the gas diffusion layers, so they have been modeled
with static equations. One possible approximation to obtain the membrane water
content is to calculate the average between the water content in the anode ow
and the cathode ow. On the other hand, the total stack mass ow rate across the
membrane (Wv,mem ) depends on two different phenomena: the electro-osmotic drag,
i.e. the water molecules dragged across the membrane by the hydrogen proton, and
the back diffusion of water from cathode to anode, caused by the concentration
gradient [10]:


i
(cv,ca cv,an )
Wv,mem = n Af c Gv nd Dw
(4.15)
F
tm
n: number of fuel cells in the stack.
Af c : fuel cell active area.
Gv : vapor molar mass.
nd : electro-osmotic drag coefcient.
F : Faraday constant.
Dw : diffusion coefcient of water in the membrane.
cv,ca : water concentration in the cathode.
cv,an : water concentration in the anode.
tm : thickness of the membrane.
More details and general ideas about PEM fuel cell modeling can be found in [8]
and in Chap. 5 of this book.

80

Assessment of SOSM Techniques Applied to Fuel Cells

4.2.7 Return Manifold


The return manifold model takes into consideration the pipeline dynamics at the
fuel cell stack exhaust. The temperature of the air leaving the stack is relatively low
(close to 80 C), so the changes of air temperature in this manifold can be neglected.
Therefore, the manifold dynamics follow the isothermic relation
dPrm Ra Trm
(Wca,out Wrm,out )
=
dt
Vrm

(4.16)

Prm : return manifold pressure.


Trm : return manifold temperature.
Vrm : return manifold volume.
Wca,out : ow entering the return manifold.
Wrm,out : outlet mass ow of the return manifold.
The relation between the multiple output mass ow, Wrm,out , and the pressure
within its volume, Prm , is modeled through the following piecewise continuous
function [8]:

1
A P ( Pamb )1/ [1 Pamb ]1/2 if Pamb > A
1
rm Prm
2
P
Prm
rm
Wrm,out =
(4.17)

A3 Prm
if PPamb

A
2
rm
where A1 , A2 , A3 , Pamb and are model constants.
Expression (4.17) is not C 1 , so it cannot be included in the model for secondorder sliding-mode design. Therefore, in order to include the return manifold in the
control design model of Sect. 4.3, Eq. (4.17) was replaced by the following fthorder polynomial [4]:
2
3
4
5
Wrm,out = p0 + p1 Prm + p2 Prm
+ p3 Prm
+ p4 Prm
+ p5 Prm

(4.18)

The values of the polynomial constants in Eq. (4.18) can be found in Appendix A,
Table A.1.

4.3 Electric Vehicle FCGS State Space Model for SOSM Control
Design
To conclude the rst stage of the control design procedure, the dynamic equations
required for oxygen stoichiometry control must be rearranged in the SOSM design
form. Considering that in this model the anode subsystem is decoupled from the
cathode subsystem and does not enter in the oxygen stoichiometry control loop, its
dynamics can be neglected, and the system order is reduced by one.
As a result of the reduction and rebuilding work performed on Pukrushpan et al.s
model, the following sixth-order control design model is proposed:
x = F (x, u, t) = f (t, x) + g(t, x, u)
x R6 ;

u R;

f : R6 R6 ;

g : R6 R6

(4.19)

4.3 Electric Vehicle FCGS State Space Model for SOSM Control Design

81

with f and g sufciently smooth vector functions, where the coordinates of the state
vector
x = [cp

Psm

msm

mO2 ,ca

mN2 ,ca

Prm ]T

(4.20)

can be summarized as follows:


x1 = cp : angular speed of the compressor motor that feeds the stack cathode
through the supply manifold.
x2 = Psm : total pressure inside the supply manifold, consisting of the sum of the
partial pressures of the gases that constitute the air (oxygen, nitrogen and water
vapor).
x3 = msm : total mass of air in the supply manifold, consisting of the sum of the
instantaneous masses of oxygen, nitrogen and water vapor.
x4 = mO2 ,ca : instantaneous oxygen mass in the stacks cathode channels. This
state is affected by the oxygen consumed in the reaction, the amount of oxygen
coming from the supply manifold and the oxygen mass outgoing through the
return manifold.
x5 = mN2 ,ca : instantaneous mass of nitrogen inside the stacks cathode channels.
It only relies on the incoming nitrogen from the supply manifold and the outgoing
nitrogen that leaves the stack through the return manifold.
x6 = Prm : total pressure inside the return manifold, consisting of the sum of the
partial pressures of the gases that constitute the air.
The input u(t) is the normalized input voltage of the compressor DC motor Vcp .
The normalization constant (Vnor = 180 V) is the maximum value of Vcp in the
operating region.
The detailed expressions of the system model are given by

x1 = B3 n(x) 1 d(x) + B1 u2 x11 B2 u


1
x2 = B6 1 d(x) x1 1 + n(x)n1
cp x2 B7 x2 B8 x2 x5 B9 x2 x4 B10 x3

x3 = B11 1 d(x) x1 x2 B12 + B13 + x5 B14 + x4 B15



x4 = (x2 B59 B60 x5 B61 x4 B62 )(x2 x2 B21 )1 e(x)

+ (x2 B63 B64 x5 B65 x4 B66 )e(x) k(x) x4 (B25 + x5 B26


1
(4.21)
+ x4 B27 x6 B24 )j (x)x41 j (x)B67 + GN2
m(x) Ist B32

1
x5 = (x2 B50 B51 x5 B52 x4 B53 )(x2 x2 B21 ) e(x) + (x2 B54


B55 x5 B56 x4 B57 )e(x) k(x) 1 j (x)B30 j (x)B68 + GN2
(B35 + x5 B36 + x4 B37 x6 B34 )m(x)
x6 = B47 + x5 B48 + x4 B49 x6 B46 B39 c(x)5 B40 c(x)4 B41 c(x)3
B42 c(x)2 B43 c(x) B44

82

Assessment of SOSM Techniques Applied to Fuel Cells

with the auxiliary functions


a(x) = a1 (x) + a21 (x)a22 (x)x1
a1 (x) = 1 d(x)
a21 (x) = d(x)
a22 (x) = 2B69 x13 n(x)
b(x) = b1 (x)b2 (x)x1
b1 (x) = a21 (x)
b2 (x) =

B69 B4 x2B4 1

B4
x12 Pamb
c(x) = x6 B45
B ((

x2

)B4 1)x 2

1
d(x) = e 69 Pamb

1
x2 B20
e(x) = 1 +
x2 x2 B21
x4
j (x) =
x5 B28 + x4 B29

1
B22
k(x) = 1 +
x2 x2 B21 + B23

1

m(x) = 1 + B58 j (x)B68 + GN2
j (x)x41


x2 B4
n(x) =
1
Pamb

(4.22)

The compressor air ow can be written in terms of the states as follows:


(4.23)
Wcp = B11 1 d(x) x1
The constants are given in Appendix A, Table A.2.

4.4 Control Objective and Sliding Surface


The second stage in the SOSM control design procedure is to establish the control
objective and, accordingly, dene the sliding surface. In this case, the proposed objective is the optimization of the energy conversion of the FCGS, maximizing the net
power (Pnet ) generated by the system under different load conditions. We assume
that the system net power results as follows:
Pnet = Pst Pcp

(4.24)

Pst : stack generated power.


Pcp : compressor power demand.
It can be shown that the efciency optimization can be achieved by regulating the
oxygen mass ow entering to the stack cathode. If an adequate comburent ow is

4.4 Control Objective and Sliding Surface

83

Fig. 4.2 Steady state analysis of the system performance under different load conditions

ensured through the stack, the load demand is satised with minimum fuel consumption. Additionally, oxygen starvation and irreversible membrane damage are averted.
Accomplishing such an optimal comburent ow is equivalent to maintaining the
oxygen excess ratio of the cathode at an optimal value. The oxygen excess ratio or
oxygen stoichiometry is dened as
O2 =

WO2 ,ca,in
WO2 ,react

(4.25)

where WO2 ,ca,in is the aforementioned oxygen partial ow entering the cathode,
which depends on the air ow released by the compressor Wcp . Recall that WO2 ,react
is the oxygen ow consumed in the reaction, so it is directly related to the stack
current:
WO2 ,react = GO2

nIst
4F

(4.26)

with GO2 the molar mass of oxygen.


The optimum value of O2 is determined from a thorough off-line analysis of
the open-loop system, considering changes in the current demand and a wide range
of oxygen stoichiometry values. In Fig. 4.2 it can be observed that the optimum

84

Assessment of SOSM Techniques Applied to Fuel Cells

value of O2 (O2 ,opt ) depends on Ist . However, in the electric vehicle FCGS under
consideration, O2 ,opt presents minor variations over the operation range. Therefore, a constant value of O2 ,opt was used when computing the system reference.
This condition may not hold for other PEM fuel cell systems, and hence the variable O2 ,opt could be easily expressed as a function of Ist . Alternatively, a possible
solution would be to apply an extremum seeking algorithm. In that case, the proposed problem is solved by a local minimum algorithm, which does not rely on a
detailed model description, and it is considered as a nonlinear on-line optimization
for unconstrained problems [1].
Another important advantage of stoichiometry control is the avoidance of oxygen starvation on the cathode channels, which would occur if O2 is allowed to go
below 1. Oxygen starvation represents a signicant problem in PEM fuel cell systems because if it persists for a long time, it may produce hot spots and consequent
irreversible damage in the catalyst layers and polymeric membranes.
Once O2 ,opt is determined, the objective of keeping the oxygen excess ratio
within optimal values can be written in terms of controlling the oxygen mass ow
(WO2 ,ca,in ). Then, the following cathode oxygen mass ow reference can be obtained from (4.25) and (4.26):
WO2 ,ca,ref = O2 ,opt GO2

nIst
4F

(4.27)

where tracking WO2 ,ca,ref effectively implies O2 = O2 ,opt .


Nevertheless, due to the fact that WO2 ,ca,in is an inaccessible internal variable of
the FCGS, it is not practical to include it in the control algorithm. This problem can
be successfully circumvented by inferring information of WO2 ,ca,in from an accessible variable of the system, such as the air mass ow delivered by the turbo compressor (Wcp ). Under the aforementioned xed humidication assumption, this variable
is directly related to WO2 ,ca,in through the supply manifold dynamics. Furthermore,
once the manifold transient is nished, the relation between Wcp and WO2 ,ca,in remains xed in all operating conditions. Therefore, the operation of the stack close
to its maximum efciency points can be successfully achieved by posing the control
objective in terms of a tracking control problem of Wcp (in this way, O2 = O2 ,opt
is ensured for every load condition, once the supply manifold transient expires) [5].
As it has been reviewed in Chap. 3, in the framework of the sliding-mode theory,
such a control objective is formalized dening the sliding variable
s(x, t) = Wcp Wair,ref

(4.28)

and steering s to zero. The expression of the air mass ow reference Wair,ref can be
readily obtained from the cathode oxygen mass ow reference (4.27). Given that the
molar fraction of oxygen in the air is known (O2 = 0.21), the desired mass ow of
dry air can be directly computed from (4.26):
Wdry,ref =

nIst
1
1
WO2 ,ca,ref =
O ,opt GO2
O2
O2 2
4F

(4.29)

4.5 Super-Twisting Controller Design

85

Then, taking into account the relative humidity of the air (RH amb ), the nal expression of the air mass reference results in
Wair,ref = (1 + amb )

nIst
1
O2 ,opt GO2
O2
4F

(4.30)

amb being the rate of vapor in the ambient air ow,


amb =

Psat (Tamb )RH amb


Gv
Ga Pamb Psat (Tamb )RH amb

(4.31)

Ga : molar mass of dry air.


Gv : molar mass of vapor.
Note that for stable ambient conditions, the reference Wair,ref only depends on
one single measurable variable, the stack current (Ist ).

4.5 Design of a SOSM Super-Twisting Controller for the Electric


Vehicle FCGS
Once the FCGS dynamic model has been re-modeled in the SOSM design form and
the sliding variable s has been dened, the nal stage of the design procedure, i.e.
the synthesis of the SOSM control law, can be completed in terms of O2 ,opt and the
measurable system variables Wcp and Ist .
According with the SOSM design procedure presented in Chap. 3, the controller
design requires the computation of the time derivative of s and global bounds for
the second derivative. To this end, using (4.23) and (4.28), we can write the sliding
variable as
s(x, t) = Wcp Wair,ref

nIst
1
= B11 1 d(x) x1 (1 + amb )
O ,opt GO2
O2 2
4F

(4.32)

and directly derive, through standard mathematical computations, the expressions


of the rst and second time derivatives of s:

s (x, t) = s(x, t) +
s(x, t). f (x, t) + g(x, u, t)
(4.33)
t
x

s(x, t) = s (x, t) +
s (x, t). f (x, t) + g(x, u, t)
t
x

+
s (x, t).u(t)
= (x, u, t) + (x, u, t)u(t)

(4.34)
u
The vector elds f and g are smooth enough in the stack operation range, and
information about s is assumed to be available. Besides, note that s does not explicitly depend on the control input, but u does appear in s via the expression of x1 ;
consequently, the sliding variable s has relative degree 1 with respect to u.

86

Assessment of SOSM Techniques Applied to Fuel Cells

Then, analyzing the system equations, it is veried that the conditions required
for SOSM control design are satised by the fuel cell system model (4.19) with
output s given by (4.32). In accordance with the general framework established in
Chap. 3, the system is under the conditions established in Case 1 of Sect. 3.5.1,
which for the particular FCGS under study are:
1. Given the total second time derivative s , there are bounds m and M such that,
within the region |s(x, t)| < s0 , the following inequality holds for all t, x X ,
uU :

s (x, t) M
(4.35)
u
For the fuel cell stack under consideration, such bounds were obtained from a
detailed analysis of the system structure together with comprehensive simulation studies, considering the addition of an appropriate feedforward control that
leads the system to the validity region. As a result, the following bounds were
determined for s0 = 5e3 :
0 < m (x, u, t) =

m = 0.5,

M = 0.9

2. Analogously, a positive constant ( = 0.01 for this FCGS) can also be calculated, such that, within the region |s| < s0 , the following inequality holds for all
t, x X , u U :





 s (x, t) + s (x, t). f (x, t) + g(x, u, t) 
(4.36)
 t

x
with Ist covering the whole operation range of the stack (1 A to 300 A).
Therefore, the stabilization problem of the electric vehicle FCGS (4.19) with
inputoutput dynamics (4.34) is solved through the solutions of the following equivalent differential inclusion by applying SOSM:
s [ = 0.01, = 0.01] + [m = 0.5, M = 0.9]u

(4.37)

and the nal parameters of the robust controller can be designed based on , m
and M .
It is interesting to stress that the bounds for functions (x, u, t) and (x, u, t)
were calculated considering the bounded perturbations and uncertainties existing in
the FCGS. In this way, (4.37) covers their effects, and hence the design based on
these values results in controllers which are naturally robust to such disturbances.
In the sequel, a control law based on the Super-Twisting (ST) algorithm is developed. As it was explained in Chap. 3, this algorithm only requires the knowledge of
the sign of s during on-line operation, and it is specially suited for plants of relative
degree 1, like the FC system under consideration. According to (3.93) and (3.94),
the control u is given as the sum of two components:
u(t) = u1 (t) + u2 (t)

(4.38)

4.5 Super-Twisting Controller Design

87

Fig. 4.3 Schematic block diagram of the FCGS and the proposed SOSM control set up

u 1 (t) = sign(s)
u2 (t) = |s| sign(s)

(4.39)

where the controller parameters , and are designed to fulll the sufcient
conditions for convergence to s(x) = s (x) = 0 in nite time, i.e.

m
2 (m + )2
2 > 2
m (m )
>

(4.40)

= 0.5
In this way it is ensured that the control has sufcient authority to affect the sign
of s , guaranteeing the establishment of a SOSM.
To complete the control set up to be implemented, as previously mentioned, an
appropriate extra control action is added, to steer the sliding variable within the
validity region. In the FCGS, a feedforward action (uff ) proved to be effective to
speed up the reaching phase. Therefore, uff was included into the control scheme,
and the implemented control action (ui (t)) comprises two terms:
ui (t) = u(t) + uff (t)

(4.41)

where u corresponds to the SOSM control action particularized by Eqs. (4.38) and
(4.39). The expression of uff is computed via a simple polynomial, a function of
the measurable current Ist , obtained from an off-line test along the entire operation
range of the FCGS [8].
A schematic block diagram of the control set-up proposed for implementation
can be appreciated in Fig. 4.3.
To rene the nal tuning of the Super-Twisting controller, it is recommended to
consider not only the behavior of the controlled variable (in this case the oxygen
stoichiometry), but also any other involved variable that can affect the overall performance of the plant. For instance, in this particular system, the electric quality of
the FCGS net power should be also taken into account. This is due the direct effect of the control on the electric power (a relative degree 0 output). Therefore, the
Super-Twisting controller parameters were selected aiming to smooth the control

88

Table 4.1 Variations of


system parameters

Assessment of SOSM Techniques Applied to Fuel Cells

Parameter

Variation

Stack temperature (Tst )

+10%

Cathode volume (Vca )

+5%

Supply manifold volume (Vsm )

10%

Atmospheric pressure (Patm )

+10%

Ambient temperature (Tamb )

+10%

Return manifold volume (Vrm )

10%

Motor constant (Kv )

10%

Electric resistance of the motor (Rcm )

+5%

Compressor diameter (dc )

+1%
+10%

Motor inertia (J )

action and to have a low content of high-frequency components. After an iterative


rening procedure, the most adequate set of parameters resulted as follows:
= 2,

= 3,

= 0.5

(4.42)

In the next section it will be shown that, under the inuence of the control (4.41),
the phase portrait ((s, s ) plane) of the controlled system presents the characteristic
non-monotonous behavior of the homogeneous Super-Twisting algorithm. The trajectories converge to the origin in nite time, twisting around the center during the
reaching mode.

4.6 SOSM Super-Twisting Controller Simulation Results


To evaluate the efciency of the proposed controller (4.41) dealing with model uncertainties, external disturbances and a wide range of current demand, a number
of simulation studies were performed. To assess the SOSM controller performance
under realistic operating conditions, simulation tests were conducted using the comprehensive ninth-order nonlinear model developed in [8]. The simulation model
incorporates the original look-up tables representing parameter characteristics obtained from experimental data. In addition, extra uncertainty has been incorporated
in several parameters of the system (see Table 4.1).
Moreover, an unknown torque disturbance, modeled as a noisy quadratic function
of the angular speed (cp ), was included in the same test (see Fig. 4.4). This friction
term is set to start at t = 15 sec, and its expression is
Tp = cp (t)2 B1 + cp (t)B2 + e(t)
109

[N m/s2 ],

106

(4.43)

with B1 =
B2 = 20
[N m/s] and e(t) corresponding to a
band-limited noise component.
Then, the features of the designed controller are examined through simulation
tests, which aim to demonstrate its nominal performance and its robust tracking

4.6 SOSM Super-Twisting Controller Simulation Results

89

Fig. 4.4 Torque disturbance

Fig. 4.5 Load current

characteristics in the presence of the aforementioned unknown disturbances and


uncertainties. To this end, series of load current (ranging from 60 A to 300 A) were
designed in order to illustrate the air regulation performance in a wide range of
operation. The sequence of current variations, generated from a ltered steps series,
is shown in Fig. 4.5. Note that abrupt and signicant changes in the amplitude of the
load demand were considered to test the prociency of the controller under exacting
operating conditions.

90

Assessment of SOSM Techniques Applied to Fuel Cells

Fig. 4.6 Oxygen excess ratio (oxygen stoichiometry)

In Fig. 4.6 and the sequel, the black line depicts the system variables evolution
in the presence of uncertainties and friction disturbances, while the grey line shows
the corresponding variables of the undisturbed system. Note that in Figs. 4.6, 4.8
and 4.9 both lines are superimposed during the entire test.
The behavior of the oxygen stoichiometry (O2 ) under different load conditions
is depicted in Fig. 4.6. In accordance with the explanation of Sect. 4.4, the values
of O2 that enable the system to work in its maximum net power points were determined performing off-line simulations (Fig. 4.2). In this electric vehicle FCGS,
due to the relatively small variation of O2 ,opt for different values of stack current,
a unique value O2 ,opt = 2.05 was adopted. When O2 ,opt is reached, it can be assured that the system is working in the neighborhood of its maximum power generation points. It is observed in Fig. 4.6 that the excess oxygen ratio satisfactorily
follows that reference in spite of uncertainties and disturbances.
Figures 4.7 and 4.8 present the dynamic behavior of the system net power (Pnet )
and the stack voltage (Vst ). It can be observed that, effectively, the former displays
a desirable low content of high-frequency components. Besides, remark that the
noticeable difference between the net power of the nominal and the disturbed system (after t = 15 sec) is not a controller aw, but the expected consequence of the
addition of the torque disturbance of Fig. 4.4. To compensate this spurious load,
the compressor must increase its consumption, hence for given Ist , the resultant net
power decreases. In spite of all disturbances, the ST robustly controls the output s,
as attested by Fig. 4.9.
Figure 4.9 shows the evolution of the sliding output s. Once the brief reaching
mode elapses, the controller maintains the system operating on the sliding surface
(s = 0, i.e. regulating O2 at its optimum value), despite the coexistence of parameters uncertainties, external perturbations and important load variations. Recall that

4.6 SOSM Super-Twisting Controller Simulation Results

91

Fig. 4.7 Net power delivered by the system

Fig. 4.8 Stack output voltage

the stability of the closed-loop system is guaranteed because the differential inclusion (4.37) has been computed to hold in the presence of such disturbances.
Note that s exhibits the typical overshoots of a system controlled by an ST algorithm (Fig. 4.9). Effectively, the characteristic twisting behavior imposed by the
Super-Twisting controller can be better appreciated in Fig. 4.10, where the phase
portrait of the system is plotted for a representative lapse, ranging in time from
24.2 sec to 24.3 sec.

92

Assessment of SOSM Techniques Applied to Fuel Cells

Fig. 4.9 Sliding variable

Fig. 4.10 s vs. s

Finally, in Fig. 4.11 the time evolution of the control signals are presented, in
terms of the actual input compressor voltage units (i.e. denormalizing the control
actions by multiplying by Vnor ). The implemented control action ui that drives the
fuel-cell-based system, together with the two constituent components u and uff , are
depicted. Note the increase in the control effort after the appearance of the friction
disturbance. To counteract its effect, the SOSM controller is in charge of providing
the necessary action to achieve the regulation objective.

4.7 Comparison with Other Control Strategies

93

Fig. 4.11 Control signal

4.7 Comparison with Other Control Strategies


4.7.1 Different SOSM Control Algorithms
In this section, the electric vehicle FCGS performance is evaluated under the action
of other control algorithms. Given the encouraging results obtained in the previous
section that proved the suitability of a SOSM ST controller for oxygen stoichiometry control, it is naturally of interest to explore the capabilities of other SOSM
algorithms. Therefore, two SOSM controllers based on Twisting and Sub-Optimal
algorithms, respectively, are synthesized and assessed in the present subsection. According to the description in Chap. 3, these SOSM algorithms are originally intended
for relative degree 2 systems, and thus the FCGS must be expanded with an integrator, taking the ancillary input = u(t)

as the control action for the design (see


Fig. 4.12).
4.7.1.1 Twisting Algorithm
As it was previously established, the Twisting controller is given by
= u = r1 sign(s) r2 sign(s )

(4.44)

with sufcient conditions for nite-time convergence


r1 > r2 > 0
(r1 + r2 )m  > (r1 r2 )M + 

(r1 r2 ) > 
m

(4.45)

94

Assessment of SOSM Techniques Applied to Fuel Cells

Fig. 4.12 Schematic block diagram proposed for Twisting and Sub-Optimal algorithms in relative
degree one systems

Note that the control law (4.44) requires the knowledge of s and s and, in the relative
degree 1 fuel cell system, synthesizes a continuous control voltage Vcp , due to the
integrator.
Recall that in this particular case, in which the relative degree 2 system is articially obtained from expanding a relative degree 1 plant, the bounds of the former
are equal to those of the latter (see Chap. 3, Sect. 3.5.2.1). Therefore, the Twisting controller gains can be directly computed using the same bounds calculated in
(4.35) and (4.36) (i.e., m = m = 0.5, M = M = 0.9 and  = = 0.9). Then,
the selected parameters for the controller were r1 = 2.25 and r2 = 0.75.

4.7.1.2 Sub-Optimal Algorithm


The Sub-Optimal control law, detailed in Chap. 3, is given by
= u = (t)U sign(s sM )

1
if (s sM )sM 0
(t) =

if (s sM )sM < 0

(4.46)

It requires the knowledge of s and of the last extremal value of the sliding variable,
sM , that is the value of s at the last local maximum, minimum or horizontal inexion point, which has to be updated on-line (see Chap. 3, Sect. 3.5.2.3 for details).
To ensure the nite-time convergence, the minimum control magnitude U > 0, the
modulation factor > 1 and the anticipation factor 0 < 1 must be selected in
accordance with
U>


m


 + (1 )M U
; +
[1; +)
m U

(4.47)

Then, the Sub-Optimal controller parameters for the electric vehicle fuel cell
were chosen as U = 3, = 0.3 and = 5.

4.7 Comparison with Other Control Strategies

95

Fig. 4.13 Time evolution of the sliding variable under the action of different SOSM controllers

4.7.1.3 SOSM Comparison Simulation Tests


To assess and compare the performance of the aforementioned SOSM controllers,
simulation tests start considering the undisturbed FCGS, and, later, several sources
of disturbances are incorporated. Similarly to the Super-Twisting case, in Sect. 4.6,
the complete ninth-order nonlinear model developed in [8] is used to model the plant
in the simulations. Additionally, the same load prole utilized in Sect. 4.6 is used
for these tests (see stack drained current in Fig. 4.5).
Results of the three SOSM algorithms (namely Super-Twisting, Twisting and
Sub-Optimal) controlling the FCGS are presented in Figs. 4.13, 4.14, 4.15 and 4.16.
Numerical values of the controllers parameters are summarized in Appendix A, Table A.3. The former corresponds to the undisturbed system and depicts a magnied
view of the time response of the sliding variable after the rst step of current demand occurred (a brief lapse between 2 and 2.5 sec is displayed). Figure 4.14 shows
the associate trajectories in the phase plane. The typical reaching behaviors of the
three algorithms are manifested in this gure. It is observed that the three SOSM
controllers display the expected nite-time convergence during this rst part of the
tests, with neither disturbances nor uncertainties present in the system.
Next, prior to the incorporation of the whole model uncertainties and external
disturbances, it is of interest to appreciate the behavior of the control algorithms
under the effect of measurement noise. To this end, between t = 4 sec and t = 11 sec
important measurement noise is added to Wcp .
After that, from t = 15 to t = 30 sec, the FCGS is strongly perturbed by incorporating the compressor disturbance torque (given in (4.43), Fig. 4.4), together with
all the parameter uncertainties of Table 4.1.
Figure 4.15 presents the actual and measured (noise included) Wcp time evolution
of the FCGS controlled with the SOSM algorithms (for the sake of clarity, in this

96

Assessment of SOSM Techniques Applied to Fuel Cells

Fig. 4.14 s(x) vs. s (x)

Fig. 4.15 Compressor air mass ow

gure only the Super-Twisting and Sub-Optimal algorithms are depicted. In fact, the
response of the Twisting controller would almost overlap the Super-Twisting one).
Their performance can be assessed from the actual Wcp curves, proving all of them
to have an excellent behavior. This is specially manifested after t = 15 sec, with the
perturbations disturbing the system simultaneously.
Note that there is an output that can be considerably inuenced by the measurement noise in certain cases. This is the net power of the generation system. Even
though the latter has not been dened as an output from the strict control viewpoint,

4.7 Comparison with Other Control Strategies

97

Fig. 4.16 FCGS net power

it is still an important physical output. Its sensitivity to high-frequency variations in


the control input is because Pnet has relative degree 0 with respect to u. In Fig. 4.16,
it can be seen that the Super-Twisting algorithm may produce a more marked noise
transmission to Pnet than the other algorithms. This is clearly due to the fact that
this algorithm contains a term related to s, not integrated, which directly injects the
uctuations of s into u.
Nevertheless, this is not necessarily a serious drawback for the algorithm application, since that negative effect can be considerably ameliorated with a proper tuning
of . To illustrate this feature, the value of has been changed from 3 to 1.5 in the
interval 8 sec < t < 10 sec. In the zoom window in Fig. 4.16, a substantial reduction
of the noise effect can be appreciated.
In the other two algorithms, measurement noise in s has been attenuated more in
the Pnet value since, in the relative degree 1 FCGS under consideration, the control
terms that directly depend on s are integrated before entering to the plant.

4.7.2 LQR Controller


It is also interesting to establish comparisons between the proposed SOSM solutions
and other more widely accepted techniques, such as state feedback linear control.
To this end, a Linear Quadratic Regulator (LQR) controller proposed in [7, 8] is
considered and analyzed in this subsection. As it can be appreciated in the cited bibliography, satisfactory results controlling the FCGS under study have been attained
with this approach. In particular, this linear controller combines feedforward with

98

Assessment of SOSM Techniques Applied to Fuel Cells

Fig. 4.17 Compressor air mass ow

feedback (state estimated plus integral) control. The form of the proposed controller
is [8]
Vcpi = Vcpff + VcpLQR = Vcpff + Vcpp k x kI q

(4.48)

The implemented control input Vcpi is the compressor voltage command, while
Vcpff and Vcpp are feedforward and pre-compensation terms, respectively, which
only depend on the measurable stack current Ist . The third term is a state feedback
control action, where x is the estimate of the linearized system state x  = x  x  o .
The constant vector x  o represents the state value at the nominal operating point of
linearization, in this case computed at optimal stoichiometry and a net power of
40 kW, with Ist = 191 A and Vcp = 164 V. Note that in this controller, the system
state x  has order eight, due to the additional presence of mH2 ,an and mw,an , the
mass of hydrogen and water inside the anode, respectively:
x  = [mO2 ,ca

mH2 ,an

mN2 ,ca

cp

Psm

msm

mw,an

Prm ]T

(4.49)

The state estimation x is obtained using the Kalman-based observer presented


in [7]. The fourth term in (4.48) provides the integral control action, where the state
equation of the integrator is given by the compressor ow error:
q = Wair,ref Wcp

(4.50)

Finally, the gains of the feedback control terms have been designed using LQR
optimal control techniques, resulting in [7]

k = 28.59 1.6 1013 60.57 7.57 579.74 2.55

(4.51)
3.6 1014 189.97
kI = 0.18

4.7 Comparison with Other Control Strategies

99

Fig. 4.18 Oxygen excess ratio (oxygen stoichiometry)

Fig. 4.19 Control action (in terms of the input voltage)

4.7.2.1 LQR Comparison Simulation Tests


For comparisons, the simulation tests presented in Figs. 4.17, 4.18 and 4.19 have
been conducted under operating conditions similar to the ones in the previous subsections. Numerical values of the LQG controller parameters are summarized in
Appendix A, Table A.4. They show the time evolution of the LQR and the SuperTwisting controlled system variables Wcp (actual and measured, with added noise),
O2 and the control actions (in terms of the input voltage), respectively.

100

Assessment of SOSM Techniques Applied to Fuel Cells

Fig. 4.20 Oxygen excess ratio (small excursion)

In the former gure, it can be appreciated that, during undisturbed operation, both
controllers display very good results, although the Super-Twisting better regulates
the actual Wcp at its desired value, given that it maintains s = 0 (recall Fig. 4.9),
except for the abrupt steps of current Ist .
Respectively, with regard to the undisturbed behavior of the oxygen stoichiometry O2 , it is observed in Fig. 4.18 that the LQR presents an initially faster dynamic
response after the steeps, but in the end it requires longer times to reach the immediate vicinity of the desired value. On the other hand, the Super-Twisting controller
attains less than 1% error in 3 sec.
It is worthwhile to mention that the linear LQR improves considerably its performance in the neighborhood of the linearization point. Moreover, in that region and
considering small excursions, it could even surpass the performance of the SuperTwisting regarding the control of O2 (see Fig. 4.20). However, as it can be observed
in Figs. 4.17 and 4.18, the SOSM Super-Twisting excels the linear controller when
working in a wide operation range.
Nevertheless, in this fuel cell application, it is in the presence of perturbations and
uncertainties that the SOSM controller demonstrates its superiority over the LQR.
This can be attested by comparing the response of both controllers under disturbed
conditions (see t > 15 sec in Figs. 4.17 and 4.18).
To emphasize this feature, showing that the robustness of the Super-Twisting
algorithm surpasses that of the LQR approach, a test conducted at constant Ist =
191 A (i.e., the value used for the linearized design of the LQR) is presented in
Fig. 4.21. In this magnied view, the excellent robustness of the SOSM proposed
controller can be clearly appraised.
To farther prove its advantageous robustness, a more exacting condition is introduced in Figs. 4.22 and 4.23, where the controllers are facing an extreme torque

4.7 Comparison with Other Control Strategies

101

Fig. 4.21 Oxygen excess ratio (oxygen stoichiometry)

Fig. 4.22 Torque disturbance

disturbance (four times greater than the original). Even when subjected to this
bulk disturbance, the Super-Twisting controller exhibits a remarkable robustness.
Comparable satisfactory results have been obtained with the other SOSM controllers.
To conclude, it is interesting to stress that the Super-Twisting controller does
not require of a state observer depending only on measurements, and its on-line
computational burden is considerably lower.

102

Assessment of SOSM Techniques Applied to Fuel Cells

Fig. 4.23 Oxygen excess ratio (oxygen stoichiometry)

4.8 Conclusions
In this chapter the feasibility of SOSM techniques to control PEM fuel cells has
been assessed. The evaluation has been conducted using a benchmark model of a
fuel cell system for an electric vehicle. The analysis has established the viability of
SOSM techniques for oxygen stoichiometry control, aiming to improve the overall
energy efciency.
Taking into account several features, such as the controlled system performance,
robustness and implementation simplicity, the SOSM controllers are shown to be
a highly efcient solution for this challenging problem, proving to be capable to
robustly tracking the optimal air mass ow, even in the presence of severe external
perturbations and model uncertainties. Among them, the Super-Twisting has been
considered as a very suitable algorithm for the FCGS, given that it is specially intended for relative degree 1 systems and only requires the real-time knowledge of
the sliding variable and not of its derivative.
Compared to the standard LQR control, the SOSM controllers demonstrate better
robustness features, in a wide range of operation. Additionally, no state observers
are required, resulting in a simple and low-computational-cost algorithm.
Now that the suitability of SOSM control for FCGS has been conrmed, the
following stage will be the development and implementation of these controllers in
an experimental fuel cell plant. This will be the subject matter of the following two
chapters.

References
1. Arce A (2010) Advanced control for fuel cell systems. PhD thesis, Universidad de Sevilla

References

103

2. Boyce M (1982) Gas turbine engineering handbook. Gulf, Houston


3. Gravdahl J, Egeland O (1999) Compressor surge and rotating stall. Springer, London
4. Kunusch C (2006) Second order sliding mode control of a fuel cell stack using a twisting
algorithm. Master thesis, Electrical Department, National University of La Plata, Argentina
(in Spanish)
5. Kunusch C, Puleston P, Mayosky M, Riera J (2009) Sliding mode strategy for PEM fuel cells
stacks breathing control using a super-twisting algorithm. IEEE Trans Control Syst Technol
17:167174
6. Moraal P, Kolmanovsky I (1999) Turbocharger modeling for automotive control applications.
SAE Technical Paper 1999-01-0908
7. Pukrushpan J, Stefanopoulou A, Peng H (2004) Control of fuel cell breathing. IEEE Control
Syst Mag 24(2):3046
8. Pukrushpan J, Stefanopoulou A, Peng H (2004) Control of fuel cell power systems. Springer,
Berlin
9. Rodatz P (2003) Dynamics of the polymer electrolyte fuel cell: experiments and model-based
analysis. PhD thesis, Swiss Federal Institute of Technology Zurich
10. Springer T, Zawodzinski T, Gottesfeld S (1991) Polymer electrolyte fuel cell model. J Electrochem Soc 138(8):23342342

Chapter 5

Control-Oriented Modelling and Experimental


Validation of a PEMFC Generation System

5.1 Introduction
In the previous chapter, the viability of second-order sliding-mode (SOSM) techniques to control fuel cell (FC) stacks has been theoretically investigated and analysed through computer simulations. After an extensive assessment of their features,
their suitability has been well established. Therefore, the natural next step is to face
the challenge of designing and implementing an efcient SOSM controller for an
actual FC stack. Particularly, the practical case under study in this chapter is an
FC-based generation workbench used for laboratory experimentation. Like in most
control design procedures, the rst and decisive stage is to obtain a reliable and
adequate mathematical description of the system. Specically, the control-oriented
nonlinear model is a key requirement for the development of the sliding-mode control algorithms, capable to robustly improve the efciency of the generation system,
avoid irreversible damage in cell membranes and extend the device lifetime.
Then, the main goal of the current chapter is to present an experimentally validated, fully analytical model of the laboratory FC ow dynamics, specially developed for SOSM control design, testing and implementation. The modelling process
is conducted following a modular methodology combining a theoretical approach,
analogous to the one outlined in Chap. 4, together with an empirical analysis based
on experimental data. The proposed semi-empirical model is capable to adequately
describe the interaction between the different subsystems, while retaining parameters that have physical signicance. The systematic procedure developed in this
chapter is organised in a way such that it can be used as a general modelling guideline, being straightforward to adapt to different fuel cell systems with few modications.
Concisely, the laboratory test plant under consideration mainly comprises a central fuel cell stack and ancillary units: air compressor, hydrogen storage tank, humidiers and line heaters (see schematic representation in Fig. 5.1). In addition, to
measure the required experimental data, different sensors are incorporated into the
system: an air mass ow meter (range 015 slpm) at the end of the supply manifold,
a piezoresistive pressure transducer (range 06 bar) to measure the manifold pressure, a piezoresistive differential pressure transducer (range 0250 mbar) to measure
C. Kunusch et al., Sliding-Mode Control of PEM Fuel Cells,
Advances in Industrial Control,
DOI 10.1007/978-1-4471-2431-3_5, Springer-Verlag London Limited 2012

105

106

Control-Oriented Modelling and Experimental Validation

Fig. 5.1 Fuel cell system diagram

the stack pressure drop, a tachometer (range 03000 rpm) on the motor shaft, a current clamp (range 03 A) and a voltage meter (range 015 V) to measure the motor
stator current and voltage, respectively. Besides, temperature sensors are arranged
in order to register the different operation conditions.
In the sequel the following modelling assumptions have been considered:
A mass ow control device ensures a constant hydrogen supply.
An auxiliary control system efciently regulates gas temperatures at ve points
of the plant (cathode and anode humidiers, cathode and anode line heaters and
stack).
A humidity control loop regulates moisture to a relative level close 100%.
The fuel cell model is one-dimensional, so the gases and reactions are considered
uniformly distributed in the cell.
The electrochemical properties are evaluated at the average stack temperature
(60 C), so temperature variations across the stack are neglected.
The water entering to the cathode and anode is only in vapour phase.
The effects of liquid water creation are negligible at the gas ow model level.
The water activity is uniform across the membrane and is in equilibrium with the
water activity at the cathode and anode catalyst layers.
General physics constants used in calculations and typical operating conditions
are summarised in Appendix B, Tables B.1 and B.5.

5.2 Air Compressor Subsystem


The air compressor is a 12-V DC oil-free KNF diaphragm vacuum pump, which
is based on a simple principle: an elastic diaphragm, xed on its perimeter, moves
up and down its central point by means of an eccentric. On the down-stroke it draws

5.2 Air Compressor Subsystem

107

Fig. 5.2 Schematic diagram of the compressor subsystem

the air or gas being handled through the inlet valve. On the up-stroke the diaphragm
forces the gas through the exhaust valve and out of the head. The compression
chamber is hermetically separated from the drive mechanism by the diaphragm.
The pump transfers, evacuates and compresses completely oil-free gas and is driven
by a 15-W DC motor.
The equations that describe the behaviour of the system are obtained by analysing
the air compressor as two coupled subsystems. The rst one embodies the permanent magnet DC motor dynamics, and the second one represents the compressing
diaphragm nonlinear characteristics (Fig. 5.2).

5.2.1 Air Compressor Motor Dynamic Equations


The following equations summarise the dynamic model of the compressor DC motor
[7]:
dia (t)
+ Ria (t) + k cp (t)
Vcp (t) = L
dt

dcp
= Te (t) Tl cp (t), Pcp (t)
J
dt

(5.1)
(5.2)

with
Te (t) = k ia (t)

(5.3)

where Vcp is the armature voltage, ia the armature current, L and R the electrical
inductance and resistance of the stator winding, k the motor constant, cp the shaft
angular speed, Pcp the absolute pressure at the compressor output, J the inertia,
Te the electrical torque, and Tl is a nonlinear function that groups together the effects of the motor and diaphragm friction and the pneumatic load. The computations
of the electrical and mechanical parameters of the compressor and the load torque
function Tl are systematically developed in the sequel.
To start with, the electrical resistance and inductance of the stator winding can
be directly measured through an electronic impedance meter. Then, the k value is
determined using the electrical equation of the motor (5.1) in steady-state operation:
L

dia (t)
= 0 = Vcp Ria k cp
dt

V R.i

(5.4)

From (5.4) k = cpcp a can be computed by measuring ia and cp at different


equilibria. Figure 5.3 shows that for various compressor pressures (Pcp ), the value
of k remains constant.

108

Control-Oriented Modelling and Experimental Validation

Fig. 5.3 k vs. cp (experimental data)

The next step deals with the characterisation of the load torque function Tl that
lumps the friction and the pneumatic loads. In a rst approach, it can be modelled as
a nonlinear static function of cp and Pcp . For modelling purposes, this load torque
expression was divided into two terms:
Tl (cp , Pcp ) = Tl,amb (cp ) + Tl (cp , Pcp )

(5.5)

The former corresponds to the load torque of the system operating at ambient
pressure. The latter takes into account the extra torque that appears when the diaphragm vacuum pump operates at pressures higher than the ambient [7].
The experimental values of the load torque can be computed using data obtained
from steady-state operation tests. Under these conditions, (t)

is zero, and thus


Eq. (5.2) gives
Te (t) = k ia (t) = Tl (cp , Pcp )

(5.6)

and Tl can be readily inferred from direct measurement of the current ia .


Then, with the assistance of (5.2), the values of the rst term of (5.5), Tl,amb ,
are obtained conducting experiments at ambient pressure and different shaft speeds.
Analysing the data (see Fig. 5.4), it can be concluded that Tl,amb can be well approximated by a linear expression, such as
Tl,amb (cp ) = A0 + A1 cp

(5.7)

The tting parameters A0 and A1 can be found in Appendix B, Table B.2.


To nd the expression of the second term of (5.5), Tl , a new set of experiments is
required, with the compressor working at different speeds and compressor pressures,
covering its entire range of operation (60 rad/s < cp < 360 rad/s, 1 bar < Pcp <
2.5 bar). Then, from (5.5), (5.6) and (5.7), Tl can be written in terms of the current
ia and the speed cp , both measurable variables:

5.2 Air Compressor Subsystem

109

Fig. 5.4 Tl,amb vs. cp : measured data and linear approximation

Tl (cp , Pcp ) = Tl (cp , Pcp ) Tl,amb (cp )


= k ia A0 A1 cp

(5.8)

Then using (8) in combination with data gathered in the experiments conducted
with different values of < cp < and < Pcp <, the following bivariate quadratic
function can be obtained by means of polynomial regression modelling tools (e.g.
polytn [8]):
2
2
Tl = A00 + A10 cp + A20 cp
+ A01 Pcp + A11 cp Pcp + A02 Pcp

(5.9)

with constant coefcients Aij (see Appendix B, Table B.3).


In Fig. 5.5, two surfaces are displayed. The rst one (coloured surface) results
from the interpolation of experimental data (dots and crosses). The second one (grey
surface) depicts the quadratic approximation (9) that best ts the obtained data. In
fact, the approximation error of cp remains below 3% in all the tested working
conditions.
Note that with this approach, losses due to deviations from the isentropic compression and unmodelled friction terms are now incorporated into the model [13].
The next parameter to be estimated is the motor inertia. Considering the practical impossibility to directly measure J , its value can be determined by adjusting
the dynamic response of the model variable cp (t) to the experimental data. In the
validation test presented in Fig. 5.6, step variations are applied to the compressor
input voltage. It can be seen that the angular speed matches the experimental values. This result is veried in a broad span of working conditions, indicating that the
developed model allows for predicting the motor behaviour in an extended range of

110

Control-Oriented Modelling and Experimental Validation

Fig. 5.5 Tl surface interpolation from real data (coloured surface) and polynomial approximation
(grey surface)

Fig. 5.6 Validation of the DC motor model dynamics

5.2 Air Compressor Subsystem

111

Fig. 5.7 Wcp surface interpolation from real data (coloured surface) and polynomial approximation (grey surface)

operation. In fact, the approximation error of cp attained with proposed approach


is less than 3% in all tested working conditions.

5.2.2 Diaphragm Vacuum Pump Characteristics


The next modelling stage required to complete the compressor model is the characterisation of the map that relates the output ow Wcp with the internal variable cp
and the compressor pressure Pcp .
To this end, several steady-state tests exploring different operating conditions
have been conducted in order to gather data from the mass ow meter, the tachometer and the pressure transducer, respectively. Then, following a tting procedure
similar to the one performed with Tl , an approximating bivariate function is obtained:
2
2
Wcp = B00 + B10 cp + B20 cp
+ B01 Pcp + B11 cp Pcp + B02 Pcp

(5.10)

with Bij constant values.


In Fig. 5.7 the Wcp approximated surface corresponding to Eq. (5.10) is shown
(grey surface), as well as the spline interpolation of the real data coloured surface.
The parameters and coefcients of the compressor under study can be found in
Appendix B, Table B.3.

112

Control-Oriented Modelling and Experimental Validation

Fig. 5.8 Compressor dynamics

5.3 Air Supply Manifold Subsystem


The air supply manifold of the laboratory fuel cell system under study is relatively
small. From the control viewpoint, a reasonable good model approximation can be
attained by neglecting its dynamics, assuming that the compressor dynamics are
dominant. Thus, the effect of the supply manifold can be taken into account by
merely readjusting the value of J determined in Sect. 5.2.1. The new value of this
equivalent inertia Jg is computed from experimental tests (Jg = 1.2 106 N m).
Figure 5.8 presents the time evolution of the air ow of the compressor (including
manifold). It can be observed that, even during transients, the model provides a very
good approximation of the experimental data. This result is also veried at different
air compressor ows and pressures, covering the whole operating range.
The previous approximated model proved to be sufciently accurate to be used in
the subsequent SOSM control design procedure. Nevertheless, in the case that certain application would require a more precise description, the supply manifold air
dynamics could be fully modelled using the principles of mass and momentum conservation [5]. Similarly to the case study modelled in Chap. 4, the general equations
that govern the supply manifold are:
dmsm
= Wcp Wsm
dt

(5.11)

dPsm
Ra
(Wcp Tcp Wsm Tsm )
=
dt
Vsm

(5.12)

together with the nozzle equation


Wsm = Ksm (Psm Phum )

(5.13)

5.4 Air Humidication Subsystem

113

where msm is the supply manifold mass of air, Wsm the supply manifold output
ow, Ra the atmospheric air gas constant, Vsm the supply manifold volume, Tcp
the compressor air temperature, Tsm the supply manifold air temperature, Psm the
supply manifold air pressure, Phum the humidier pressure, and Ksm is the restriction
constant of the output nozzle.
In the general case, the temperatures Tsm and Tcp are different; however in the
laboratory FC stack this process is practically isothermal, so Eqs. (5.11) and (5.12)
become equivalent, reducing the order by one. Additionally, in this system, due to
the presence of a humidier before the cathode, Phum appears instead of Pca in the
output nozzle equation.
Note that Ra is a known constant and Vsm is easily measurable, so the single
parameter left to be estimated is Ksm . This goal can be achieved by setting the
compressor ow Wcp at different steady-state operating points and collecting data
of the differential pressure Psm Phum . Then, Ksm can be obtained by adjusting the
restriction constant value such that the experimental data satises Eq. (5.13). Note
that this value does not exactly correspond to the actual output nozzle. It can be
interpreted as the restriction constant of an equivalent air supply manifold, which is
taking into account the effect of any possible remaining dynamics, disregarded in
the model of the compressor subsystem developed in the previous subsection.

5.4 Air Humidication Subsystem


Cellkraft P-series humidiers based on exchange membranes are used to maintain
proper humidity conditions inside the cells, which is crucial to ensure the optimal
operation of PEM membranes. Gas humidication at ows rates up to 10 slpm and
pressures close to 10 bar can be achieved with this sort of devices. The line heaters
and stack temperatures are controlled by a power station via decentralised PID controllers, allowing for independent gas conditions to the stack.
The operating principle of the membrane humidication technique is feeding
deionised temperature-controlled water and the gas to be humidied to each side
of a polymeric membrane, respectively. During the process, a humidity gradient is
established across the membrane, allowing the transference of water towards the
circulating gas by diffusion. The degree of humidication is regulated by adjusting
the water temperature within the humidier. The closed-loop control of moisture
can be performed by two different methods, i.e. controlling either the dew point of
the gas or the water temperature [3].
Prior to entering the stack, the humidied gas circulates through a line heater. The
purpose of this device is twofold. On the one hand, by rising the gas temperature,
condensation inside the cathode channels is prevented. On the other hand, regulation
of the differential temperature (between the humidier and line heater) allows one
to control the relative humidity of the cathode inlet gas ow.
To obtain the equations that govern the humidication subsystem, the modelling
process is divided into two steps. First, it is assumed that the subsystem dynamics
is dominated by the air humidier pressure change, without taking into account the

114

Control-Oriented Modelling and Experimental Validation

Fig. 5.9 Schematic diagram of the humidication subsystem

effect of the vapour injected to the gas (see Fig. 5.9). Then, as a second stage, the
effect of the vapour ow injected to the gas, Wv,inj , is taken into consideration by
adding a nonlinear static term, function of the gas temperature, pressure and speed
conditions [7].

5.4.1 Step 1
Under the aforementioned assumptions and considering that the humidier pressure
is similar to the compressor load pressure (Pcp = Phum ), the equations of the cathode
humidier dynamics are given by the following expressions:
dmhum
= Wcp Whum
dt

(5.14)

dPhum
Ra
=
(Wcp Tcp Whum Thum )
dt
Vhum

(5.15)

Whum = f (Thum , Pca , Phum,diff )

(5.16)

Phum,diff = Phum Pca

(5.17)

with

where mhum and Phum are the mass and pressure of air inside the humidier, Whum is
the ow of air that leaves the humidier, Vhum is the volume of the humidier, Thum
is the gas temperature inside the humidier, Phum,diff is the humidication subsystem pressure drop, and Pca is the cathode inlet pressure. Note that in the approximated model proposed in Sect. 5.3 the effect of the manifold has been incorporated
into the compressor model hence, in (5.14), Wcp has been considered as the inlet
ow into the humidier.
The right-hand side of (5.16) corresponds to a nonlinear nozzle function, strongly
dependent on the humidier gas temperature, the cathode pressure and the humidier pressure drop (see Fig. 5.10). It could be approximated by a trivariate function
or a family of bivariate functions (parameterised by Thum ), obtained through experimental data gathered from tests performed at different operating temperatures and
pressures. In particular, the current laboratory test plant is set to operate at constant
stack temperature of 60 C and humidier temperature of 55 C, respectively, regulated through external controllers. The former is a recommended operation temperature for the equipment, whereas the latter is adjusted to obtain high relative

5.4 Air Humidication Subsystem

115

Fig. 5.10 Humidier characteristics (ts and real data)

humidity of the cathode inlet ow (> 79%) while preventing vapour condensation.
Then, considering a constant humidier working temperature, the nozzle function
can be well approximated by the following bivariate function:
Whum = C0 + C1 Phum,diff Pca + C2 Pca

(5.18)

where the coefcients C0 , C1 and C2 are experimentally determined from the tests
(see thick solid lines in Fig. 5.10). Coefcient values are given in Appendix B,
Table B.4.

5.4.2 Step 2
In this step, the vapour injected to the air stream is incorporated to the model. Then,
the total humidied air ow entering the cathode (Wca ) is given by
Wca = Whum + Wv,inj

(5.19)

Assuming that the humidifying closed-loop system of the device efciently regulates the gas relative humidity, the computation of the injected water to the air ow
can be described by
Wv,inj =

Gv RH hum Psat (Thum )Wa,hum


Wv,hum
Ga Pa

(5.20)

116

Control-Oriented Modelling and Experimental Validation

Gv being the vapour molar mass, Ga the dry air molar mass, and Psat (Thum ) the
vapour saturation pressure at the humidier temperature. RH hum is the relative humidity of the gas exiting the humidier, which in normal operating conditions can
be considered a known value, in accordance with the humidier technical specications (usually, close to 100%). The dry air partial pressure (Pa ), the dry air output
ow of the humidier (Wa,hum ) and the ow of vapour due to ambient moisture entering the humidier (Wv,hum ) are variables that depend on the ambient conditions
and can be directly computed using the following relationships:
Wa,hum =

1
Whum,out
1 + hum

Wv,hum = Whum Wa,hum

(5.21)

(5.22)

with
hum =

Psat (Tamb )RH amb


Gv
Ga Pamb Psat (Tamb )RH amb

(5.23)

where hum is the ambient humidity ratio, Tamb the ambient temperature, Pamb the
ambient pressure, and RH amb is the ambient relative humidity.
At this point, there is only one parameter left to be estimated to complete the humidication subsystem model. This is the humidier volume (Vhum ), present in the
dynamic equation (5.15). Instead of using a measured value of the volume, a convenient estimation of this parameter can be attained by adjusting the transient response
of the modelled Phum to match the experimental data (tests varying the compressor
air ow at xed humidier temperature are considered). The estimated value does
not exactly correspond to the actual humidier volume. It rather represents the volume of an equivalent humidication subsystem, which takes into account modelling
errors and unmodelled dynamics. Figure 5.11 shows that highly satisfactory dynamic validation results are achieved (refer to Appendix B, Tale B.4 for estimated
parameters).
A nal remark is pertinent to close this section. It was previously mentioned that
regulation of the line heater temperature allows controlling the relative humidity
of the gas. Then, in accordance with the Dalton law, the effect of the increase of
temperature (from Thum to Tlh ) on the partial pressures and the relative humidity of
the cathode inlet gas ow can be easily computed through
Tlh
Pi,hum
Thum
Pv,lh
RH lh =
Psat (Tlh )
Pi,lh =

(5.24)
(5.25)

where i stands for O2 , N2 and vapour, respectively, and Psat (Tlh ) is the vapour saturation pressure at the line heater temperature Tlh .

5.5 Fuel Cell Stack Flow Subsystem

117

Fig. 5.11 Humidier dynamic validation

5.5 Fuel Cell Stack Flow Subsystem


The stack is an EFC50-ST ElectroChem , which is a laboratory PEM fuel cell
system designed for the study of membrane electrode assemblies and fuel cell operation. The system consists of a seven-cell stack with Naon 115 membranes,
platinum catalyst loading of 1 mg cm2 , Toray carbon bre papers as gas diffusion
layers and 50 cm2 of active area. This unit generates up to 100 W.

5.5.1 Cathode Channels


The dynamic mass balance within the stack channels depends on the gases partial
pressures, on the water transported by the membrane and on the electric current
drained from the stack (Ist ). The cathode state variables are the masses of the circulating elements, i.e. oxygen (mO2 ,ca ), nitrogen (mN2 ,ca ) and vapour (mv,ca ). Then,
the dynamic equations that govern the behaviour of the gases inside the cathode are
given by
dmO2 ,ca
= WO2 ,ca,in WO2 ,ca,out WO2 ,react
dt

(5.26)

dmN2 ,ca
= WN2 ,ca,in WN2 ,ca,out
dt

(5.27)

118

Control-Oriented Modelling and Experimental Validation

dmv,ca
= Wv,ca,in Wv,ca,out + Wv,ca,gen + Wv,mem
(5.28)
dt
while the following relationships hold for the cathode input and output ows (Wca
and Wca,out ):
Wca = WO2 ,ca,in + WN2 ,ca,in + Wv,ca,in

(5.29)

Wca,out = WO2 ,ca,out + WN2 ,ca,out + Wv,ca,out

(5.30)

where WO2 ,ca,in , WN2 ,ca,in , Wv,ca , WO2 ,ca,out , WN2 ,ca,out and Wv,ca,out are the input and output ows of oxygen, nitrogen and vapour, WO2 ,react the ow of oxygen
that reacts in the cathode, Wv,ca,gen the ow of vapour generated in the reaction,
and Wv,mem is the ow of water transferred across the membrane (comprising the
electro-osmotic drag term and the back-diffusion term).
Subsequently, the calculation of the ow terms that constitute the right-hand sides
of Eqs. (5.26)(5.28) must be addressed.
To begin with, the amount of reduced oxygen and generated vapour in the cathode reaction is computed from the stack current, according to the following two
electrochemical principles:
nIst
(5.31)
4F
nIst
Wv,ca,gen = Gv
(5.32)
2F
where n is the number of cells of the stack, and GO2 and Gv are the molar masses
of oxygen and vapour, respectively, and F is the Faraday constant.
Next, the components of the cathode input and output ows are considered. First,
assuming knowledge of Wca from (5.19), the partial input ows WO2 ,ca,in , WN2 ,ca
and Wv,ca are readily calculated using (5.29):
WO2 ,react = GO2

1
Wca
1 + ca
1
WN2 ,ca,in = (1 XO2 ,ca )
Wca
1 + ca
Wv,ca,in = Wca WN2 ,ca,in WO2 ,ca,in

WO2 ,ca,in = XO2 ,ca

(5.33)
(5.34)
(5.35)

where ca is the humidity ratio, and XO2 ,ca is the mass mole fraction of the input
air ow, given by
Gv Pv,lh
Ga (PO2 ,lh + PN2 ,lh )
O2 GO2
XO2 ,ca =
O2 GO2 + (1 O2 )GN2
ca =

(5.36)
(5.37)

O2 being the oxygen mole fraction of the ambient air.


Second, using (5.30), the partial output ows WO2 ,ca,out , WN2 ,ca,out and Wv,ca,out
can be obtained following a similar procedure:

5.5 Fuel Cell Stack Flow Subsystem

119

1
Wca,out
1 + ca,out
1
WN2 ,ca,out = (1 XO2 ,ca,out )
Wca,out
1 + ca,out
Wv,ca,out = Wca,out WO2 ,ca,out WN2 ,ca,out

WO2 ,ca,out = XO2 ,ca,out

(5.38)
(5.39)
(5.40)

with the output humidity ratio and mass mole fraction


ca,out =
XO2 ,ca,out =

Gv Pv,ca
Ga (PO2 ,ca + PN2 ,ca )

O2 ,out GO2
O2 ,out GO2 + (1 O2 ,out )GN2
O2 ,out =

PO2 ,ca
Pca

(5.41)
(5.42)
(5.43)

where O2 ,out is the cathode oxygen mole fraction.


However, for these computations, the cathode output ow Wca,out is not yet available, given that it is not measurable due to its high vapour content. It must be indirectly obtained, making use of the pressure drop measurement. The relationship
between the output ow and the pressure drop can be modelled as a linear nozzle
equation:
Wca,out = Kca,out (Pca Pamb )

(5.44)

where Pamb is the ambient pressure, and Kca,out is the cathode output restriction that
can be also governed through a mechanical back pressure regulator.
Then, to compute Wca,out , it is necessary to determine Kca,out (with a bypass
in the cathode back pressure regulator). To estimate this parameter, experimental
data of the pressure drop and the cathode output ow are required. The former is
available from the differential pressure transducer, but, as it was previously said, no
direct measurement of Wca,out is feasible due to its high relative humidity. However,
under appropriate experimental conditions, its steady-state values can be inferred
from measurements of the compressor ow Wcp . The estimation test conditions are:
(a) steady-state operation, (b) equally humidied reactant gases and (c) nil stack
current. On the one hand, Ist = 0 guarantees that the liquid water (Wl,ca,out ) and the
reaction ows (WO2 ,react and Wv,ca,gen ) remain zero. On the other hand, considering
anode and cathode gases at similar relative humidities ensures that at steady-state
operation there is no water concentration gradient across the membrane, so the effect
of Wv,mem can be neglected. Therefore, under these testing conditions, Wca,out is
equal to Wca (see (5.26)(5.28)). Then, using (5.14), (5.19) and (5.20), the data
of Wcp allow the computation of Wca,out and, consequently, the estimation of the
nozzle restriction [7].
Note that the partial pressures of the gases inside the cathode, required in (5.36),
can be obtained from the stack temperature and the masses of oxygen, nitrogen and
vapour. Using the Dalton law, the cathode partial pressures and relative humidity
are:

120

Control-Oriented Modelling and Experimental Validation

mi,ca Ri Tst
Vca
Pv,ca
RH ca =
Psat (Tst )
Pi,ca =

(5.45)
(5.46)

where subscript i stands for O2 , N2 and v, respectively, and Vca is the cathode
volume.
The last ow term of (5.26)(5.28) to be computed is the water transferred across
the membrane. To this end, the anode relative humidity is required, so the anode ow
model will be rst addressed, and, subsequently, the calculation of Wv,mem will be
resumed.

5.5.2 Anode Channels


In this type of PEMFC systems the input hydrogen ow is independently regulated;
thus the input Wan is assumed to be a known constant. Moreover, due to the fact that
Wan remains constant and the anode humidier and line heaters are controlled, it is
not necessary to model such gas conditioning devices in the anode line. Under this
conditions, the dynamics of the anode channel can be modelled by
dmH2 ,an
= WH2 ,an,in WH2 ,an,out WH2 ,react
dt
dmv,an
= Wv,an,in Wv,an,out Wv,mem
dt
while the following equations hold for the anode input and output ows:

(5.47)
(5.48)

Wan = WH2 ,an,in + Wv,an,in

(5.49)

Wan,out = WH2 ,an,out + Wv,an,out

(5.50)

where WH2 ,an,in , Wv,an,in , WH2 ,ca,out and Wv,an,out are the input and output ows
of hydrogen and vapour, respectively, WH2 ,react is the ow of hydrogen consumed
in the reaction, and Wv,mem is the aforementioned ow of water transferred to the
cathode. In this particular case, no liquid water is supposed to be condensed in the
anode channels, given that in normal working conditions the relative humidity of
the anode is always below 100%. On the other hand, the hydrogen consumed in the
reaction is
nIst
(5.51)
2F
where GH2 stands for the molar mass of hydrogen.
Analogously to the cathode channel, the components of the anode input and output ows must be calculated. The partial input ows WH2 ,an,in and Wv,an,in are obtained through
WH2 ,react = GH2

5.5 Fuel Cell Stack Flow Subsystem

121

WH2 ,an,in =

1
Wan
1 + an

Wv,an,in = Wan WH2 ,an,in


an =

Gv Pv,lh,an
GH2 PH2 ,lh,an

(5.52)
(5.53)
(5.54)

where an is the humidity ratio of the anode input gas, PH2 ,lh,an the anode input H2
partial pressure, and Pv,lh,an is the anode input vapour partial pressure that can be
obtained using the Dalton law.
Besides, the partial output ows WH2 ,an,out and Wv,an,out are computed as follows:
1
WH2 ,an,out =
Wan,out
(5.55)
1 + an,out
Wv,an,out = Wan,out WH2 ,an,out
an,out =

Gv Pv,an
GH2 PH2 ,an

(5.56)
(5.57)

where an,out is the humidity ratio of the gas inside the anode, PH2 ,an the anode H2
partial pressure, and Pv,an is the anode vapour partial pressure.

5.5.3 Membrane Water Transport


Now the calculation of Wv,mem can be taken up again. The ow of water across the
membrane is modelled assuming linear concentration gradients from channels inlet
to outlet and across the membrane thickness. Then, it can be expressed as [15]


i
cv,ca cv,an
Wv,mem = nd + Dw
(5.58)
Gv Afc n
F
tm
where i is the stack current density, Afc the fuel cell active area, tm the membrane dry
thickness, and cv,ca and cv,an are the water concentrations at the membrane surfaces
on the cathode and anode sides, respectively. The term nd is the electro-osmotic
drag coefcient (number of water molecules carried by each proton), and Dw is
the back-diffusion coefcient of the membrane. The water concentration terms are
determined from the membrane water contents on the cathode (ca ) and anode (an )
sides:
m,dry
cv,ca =
ca
(5.59)
Gm,dry
m,dry
an
(5.60)
cv,an =
Gm,dry

122

Control-Oriented Modelling and Experimental Validation

Fig. 5.12 Experimental and estimated membrane water diffusion characterisation

where m,dry is the membrane dry density, and Gm,dry is the membrane dry molecular weight. The water content in the membrane is dened as the ratio of water
molecules to the number of charge sites. When no liquid water is present in the
channels, the ratio can be estimated at both sides using the following equation [15]:
j = a0 + a1 RH j + a2 RH 2j + a3 RH 3j

(5.61)

RH j being the gas relative humidity, and subscript j referring to cathode or anode
(j = ca, an), respectively.
The next step is to estimate the apparent diffusion coefcient Dw of expression (5.58). Two different experiments can be set-up to compute this parameter,
either a cathode or an anode drying test. In both cases, the stack current must be
set to zero (Ist = 0) in order to cancel the stack current density i in (5.58). For the
former, a long-term cathode drying procedure is conducted, decreasing the cathode
humidier temperature from 55 C to 40 C, while setting the temperatures of the
anode humidier, both line heaters and the stack at 60 C. With this test, a water
concentration gradient is established between the channels, and an increasing extra
ow in the cathode output can be detected due to the membrane contribution [7].
The second test is conducted analogously to the rst one, but in this case the
anode channel is dried out, keeping the other variables at similar stationary conditions. Following this simple procedure, the Dw coefcient can be directly determined without using humidity sensors or a more specic equipment. Figure 5.12
displays the data gathered from both tests (the average value obtained for the backdiffusion coefcient is given in Appendix B, Table B.4).

5.6 Electrical Characterisation of the Fuel Cell Stack

123

Finally, the electro-osmotic coefcient nd is characterised through the widely


accepted expression developed in [4] and reported in [11] and [9]:
nd = n0 + n1 m + n2 2m

(5.62)

where m is the average membrane water content, which can be derived from equation (5.61) considering
RH m = (RH ca + RH an )/2

(5.63)

5.6 Electrical Characterisation of the Fuel Cell Stack


Up to this point, the equations and parameters required to build the model of the
laboratory FC ow dynamics have been obtained. To complete the system model, it
is of interest to provide an electrical characterisation of the fuel cell stack behaviour.
To this end, a semi-empirical approach, derived from the physical processes that occur in the MEA and the phenomena that take place in the channels, is used to model
the output electrical properties of the stack under different operating conditions,
covering the whole operation range of air ow rates and pressures [6].
As explained in Sect. 2.2, combining the open circuit voltage E with the effects
of activation, ohmic resistance and concentration, the operation voltage of a PEM
fuel cell (Vfc ) can be described by the following expression:
Vfc = E Vact Vohm Vconc

(5.64)

The parameters that dene expression (5.64) can be computed with the help of
appropriate experimental data. Under these conditions, the term E can be substituted by V0 = E0 = 1.229 V, the theoretical electrochemical potential of a single
hydrogen/oxygen fuel cell. Then, the following particular expression is valid:

 

RTst
PO2 ,ca
i
(nc i)
ln
(5.65)
+ b ln
(R0 R1 m )i me
Vfc = V0
2F
i0
a
where i is the current density, and , i0 , R0 , R1 , m, nc , b and a are empirical parameters obtained from the experiments, which are required to characterise the general
expression (5.64) (see the explicit equation in Chap. 2). Hereinafter the section is
devoted to describe the procedure followed to obtain such parameters. This procedure can be used as a general guideline applicable to different PEM fuel-cell-based
generation systems.
First, considering that the activation losses do not depend on the pressure and
oxygen stoichiometry, the coefcients and i0 are calculated. To this end, a logarithmic plot of the VI curve is a helpful tool to better determine the electrical
behaviour of the MEAs at low current densities (see Fig. 5.13).
In this particular system, = 0.7 and i0 = 4.5 106 A/cm2 are the values that
better adjust the experimental data when the cathode air ow rates (Wca ) is varied
from 1.108 105 kg/s to 7.707 105 kg/s and Pan = Pca = 1 bar. It is important
to stress that in all the experiments presented in this section the air and hydrogen

124

Control-Oriented Modelling and Experimental Validation

Fig. 5.13 VI logarithmic curve

dew points are set at 55 C, while the line heaters and stack temperatures are xed
at 60 C in order to avoid water condensation inside the stack. Considering these
externally regulated temperatures, the relative humidity of the inlet gases is 0.75.
The stack current prole is set to vary from 0 A to 12 A, and the anode hydrogen ow
is xed at 6.483 106 kg/s (2 slpm of dry H2 ). Under these testing conditions, the
hydrogen stoichiometry remains above 5, even at the highest current (12 A), and
thus the losses due to hydrogen concentration can be neglected [6].
At this stage, is important to remark that liquid water occupies the pore volume and reduces the gas diffusion layer surface, resulting in a decrease in voltage
and fuel cell efciency. This phenomenon is unimportant in the cathode and anode
dynamics because it does not interfere with the channels gas transport as the water
excess is dragged by the ows. Nevertheless, it affects the system efciency because
the amount of liquid water inside the diffusion gas layers increases, decreasing the
membrane effective area (Afc ), resulting in an increased current density and hence
a drop in voltage [1, 10, 14]. Consequently, for model purposes, an apparent fuel
cell area (Aapp ) can be dened, to take into account the surface reduction due to the
liquid lm at the interface of the gas diffusion layers and membrane. In the present
test bench there is not a purging schedule, and cathode and anode ows are kept
constant in every operating conditions, so the aforementioned effect can be modelled with the incorporation of a scaling factor for Afc , which takes into account the
membrane effective area decrease due to liquid water generation:
i=

Ist
Ist
=
Aapp Afc (1 ca mH2 O,ca,liq an mH2 O,an,liq )

(5.66)

5.6 Electrical Characterisation of the Fuel Cell Stack

125

Fig. 5.14 Membrane water content

where mH2 O,ca,liq and mH2 O,an,liq are the cathode and anode liquid water masses
in the channels, respectively. This terms were not approximated in this modelling
approach, due to the fact that the model presented in this book is mainly oriented
to be used in gases ow control issues. Besides, note that the parameters ca and
an can be adjusted in order to take into account the modulation effects of cathode
and anode water accumulation in the stack voltage (low-frequency phenomena).
Nevertheless, for steady-state computations done in this section, the approximated
value of the apparent fuel cell area for the current stack was Aapp = 20.74 cm2 for
all working conditions.
Subsequently, the ohmic polarisation is determined in base of the membrane average water content m (Eqs. (5.61)(5.63)). The following gure (Fig. 5.14) depicts
the prole described by m in the validated model of the stack presented in Sect. 5.5,
at different current densities and air ow rates.
Considering that above a certain value of i the relationship between m and i
remain xed, it is natural to assume that the ionic conductivity of the membrane
remains practically unaltered at different air ow rates. This assumption conducts
to select a small value of R1 and R0 close to the average resistance of the MEA,
according to the data. In this particular case, the most representative values for the
stack cells are R1 = 0.005 /cm2 and R0 = 0.22 /cm2 .
The next pair of parameters to be estimated are a and b, which consider the
effect of low oxygen stoichiometry in the cathode. As it can be seen in Fig. 5.15, in
all polarisation curves there is a signicant voltage drop at certain current densities.
Moreover, the values of current densities where voltage drops start vary with the
air mass ow. Then, in the model, the parameter a can be associated to the value

126

Control-Oriented Modelling and Experimental Validation

Fig. 5.15 Polarisation curve of the fuel cell stack at constant pressure and different levels of air
ow

of oxygen partial pressure that produces the steep effect of the concentration loss.
On the other hand, parameter b represents the sensitivity of the polarisation curve at
different air ow rates, so it can be adjusted depending on the data obtained from the
polarisation curves at the minimum and maximum air ow rates (red line and yellow
line in Fig. 5.15). For the laboratory fuel cell stack, the estimated values of these
parameters were a = 0.2 bar and b = 0.06 V. A standard curve tting procedure
was followed to estimate the parameters [6].
Lastly, values m and nc are adjusted in order to verify the general gas diffusion
loss at high stack currents. The obtained values from the tting were m = 8.1
105 V and nc = 15.2 cm2 /A.
Figure 5.15 presents the polarisation curves obtained at constant pressure (1 bar)
and different levels of air ow (1.108 105 kg/s to 7.707 105 kg/s), while
Fig. 5.16 depicts the polarisation curves obtained at constant air ow (7.707
105 kg/s) and increasing levels of working pressures (12.35 bar).
In these gures it can be appreciated that the model based on Eq. (5.65) with
the proposed parameters satisfactorily predicts the fuel cell VI behaviour under
different tested conditions of air ow rates and cathode pressures.

5.7 Conclusions
The control-oriented modelling of an actual PEM fuel cell stack has been approached in this chapter. The proposed procedure tackles the modular modelling of

5.7 Conclusions

127

Fig. 5.16 Polarisation curve of the fuel cell stack at constant air ow and increasing levels of
working pressure

an experimental complex system that combines mechanical, electrical, pneumatic


and electrochemical subsystems. It provides a nonlinear characterisation that satisfactorily describes the steady-state and dynamical behaviour, successfully covering
the entire operation range of the fuel-cell-based system under study.
Due to the fact that the model has been primarily developed for model-based
control studies, a system level approach has been considered and only dynamics in
the range of 102 to 100 seconds has been taken into account.
The semi-empirical methodology followed in this chapter is not an example
on identication nor a theoretical exercise. Guided by the knowledge of the processes and reactions that take place in the real fuel cell, the different components
were modelled using available general information and particular experimental data,
gathered from simple tests. Therefore, the proposed procedure can be used as a
guide for control-oriented modelling of PEM fuel cell systems with similar features.
Important control problems found in PEM fuel cells such as the ones presented
in [2, 11, 13] and [12] (H2 /O2 stoichiometry regulation, total and partial pressures
control, H2 consumption minimisation, etc.) can be approached using the developed
control model. In particular, taking advantage of the continuous, smooth dynamic
equations presented here, in the beginning of next chapter a seventh-order nonlinear state space model, primarily focused on the fuel cell uid dynamics, will be
formulated as the rst stage for SOSM control design and practical implementation.

128

Control-Oriented Modelling and Experimental Validation

References
1. Arce A (2010) Advanced control for fuel cell systems. PhD thesis, Universidad de Sevilla,
Spain
2. Arce A, del Real AJ, Bordons C, Ramrez DR (2010) Real-time implementation of a constrained MPC for efcient airow control in a PEM fuel cell. IEEE Trans Ind Electron
57(6):18921905
3. Cellkraft (2007) P-10 humidier manual: v 2.0
4. Dutta S, Shimpalee S, Van Zee JW (2001) Numerical prediction of mass-exchange between
cathode and anode channels in a PEM fuel cell. Int J Heat Mass Transf 44:20292042
5. Helvoirt J, Jager B, Steinbuch M, Smeulers J (2005) Modeling and identication of centrifugal
compressor dynamics with approximate realizations. In: IEEE conference on control applications, Toronto, Canada
6. Kunusch C, Puleston PF, Mayosky MA, More J (2010) Characterization and experimental
results in PEM fuel cell electrical behaviour. Int J Hydrog Energy 35:58765881
7. Kunusch C, Puleston PF, Mayosky MA, Husar A (2011) Control-oriented modelling and experimental validation of a PEMFC generation system. IEEE Trans Energy Convers 26(3):851
861
8. DErrico J (2006) Polytn (N-d polynomial regression model). http://www.mathworks.com/
matlabcentral/leexchange/10065
9. McKay DA, Ott WT, Stefanopoulou AG (2005) Modeling, parameter identication, and validation of reactant and water dynamics for a fuel cell stack. In: ASME international mechanical
engineering congress & exposition
10. McKay DA, Siegel JB, Ott WT, Stefanopoulou AG (2008) Parameterization and prediction of
temporal fuel cell voltage behavior during ooding and drying conditions. J Power Sources
178:207222
11. Pukrushpan JT, Stefanopoulou AG, Peng H (2004) Control of fuel cell power systems.
Springer, Berlin
12. Ramos-Paja C, Bordons C, Romero A, Giral A, Martinez-Salamero L (2009) Minimum fuel
consumption strategy for PEM fuel cells. IEEE Trans Ind Electron 56(3):685696
13. Rodatz PH (2003) Dynamics of the polymer electrolyte fuel cell: experiments and modelbased analysis. PhD thesis, Swiss Federal Institute of Technology Zurich
14. Siegel JB, McKay DA, Stefanopoulou AG (2008) Modeling and validation of fuel cell water
dynamics using neutron imaging. In: American control conference
15. Springer TE, Zawodzinski TA, Gottesfeld S (1991) Polymer electrolyte fuel cell model. J Electrochem Soc 138(8):23342342

Chapter 6

SOSM Controller for the PEMFC-Based


Generation System. Design and Implementation

6.1 Introduction
The main objective of this chapter is to present the actual development and experimental validation of SOSM controllers, previously reviewed and analysed in this
book. As anticipated, the controllers were designed and implemented in the real
FC-based generation workbench introduced and modelled in Chap. 5. Three control
set-ups based on Super-Twisting, Twisting and Sub-Optimal algorithms were developed. As detailed in Chap. 4, power conversion optimisation of the laboratory PEM
fuel cell system is sought via oxygen stoichiometry regulation. In the implementation, the effect of practical problems such as saturation and possible wind-up have
been also taken into account and counteracted.
There is also a second intention for this chapter. Taking advantage of the consecutive steps required to present the development of the aforementioned actual FC
controllers, it is aimed to provide in parallel a condensed recapitulation of the concepts and design procedures reviewed and elaborated along the previous chapters.
The objective is twofold. On the one hand, this endows the chapter with a certain
degree of self-containment, improving its readability by sparing the reader from
awkward goings and comings. On the other hand, the goal is to summarise the successive stages of the FC-SOSM control design process presented in this book into
one concise format. This would offer a unied abridged design guideline to develop
procient practical SOSM controllers for efciency optimisation of FCGS.

6.2 State Space Model of the Experimental Fuel Cell System for
Control Design
A necessary phase, previous to the controller design stage, deals with the rearrangement of the equations presented in Sects. 5.25.5, in order to obtain the state space
model suitable for nonlinear control design purposes. This procedure involves coupling all the presented differential equations with their auxiliary equations in order
C. Kunusch et al., Sliding-Mode Control of PEM Fuel Cells,
Advances in Industrial Control,
DOI 10.1007/978-1-4471-2431-3_6, Springer-Verlag London Limited 2012

129

130

6 SOSM Control Design and Implementation

to represent the system only in terms of the space states, external inputs (Ist , Wan,in
and Vcp ) and constants, as well as taking into account any possible issue leading
to order reduction. In this particular case, the latter involves taking Tcp = Thum and
dia /dt = 0 [4]. In fact, experimental tests revealed that Eqs. (5.14)(5.15) are linearly coupled within the PEMFC operation range. Besides, the time constant of the
variable ia can be neglected with respect to the rest of the system dynamics. It must
be taken into account that the system is operating under the assumptions stated in
Sect. 5.1.
Taking the state x R7 , the control input u = Vcp R1 and the operating conditions given in Table B.5 of Appendix B, the experimental PEMFC generation system
under study can be described as a seventh-order nonlinear system with C 1 vector
elds, accordingly represented by the following state space relations (6.2)(6.8):
x = f (x, t) + gu(t)
xR ;
7

u R;

f : R 7 R7

(6.1)

where g = [K 0 . . . 0]T with K = m1 , and the state variables are as follows:

x1 = cp : motor shaft angular velocity.


x2 = mhum,ca : air mass inside the cathode humidier.
x3 = mO2 ,ca : oxygen mass in the cathode channels.
x4 = mN2 ,ca : nitrogen mass in the cathode channels.
x5 = mv,ca : vapour mass in the cathode channels.
x6 = mH2 ,an : hydrogen mass in the anode channels.
x7 = mv,an : vapour mass in the anode channels.

The physical input to the system is the armature voltage of the compressors DC
motor (Vcp ). The measured outputs selected to evaluate the control performance
are the compressor current, the stack voltage, the stack cathode pressure and the
compressor air mass ow. The stack current is an external perturbation that can also
be measured in real time. In what follows, each differential equation is analysed,
reviewing the corresponding physical relations that support them [3].
The rst expression comes from Eq. (5.1) and relates the motor shaft angular
velocity, the inertia, the electric torque delivered by the motor and the compressors
load torque. This last variable value comes from the empirical relation that links the
compressor working pressure and the shaft angular velocity (x1 ) (see Eq. (5.10)),
considering both friction and pneumatic load terms:

x1 = m1 (u m2 x1 ) x1 m3 + A0 + A00 + A10 (x2 m5 + m6 )
+ A20 (x2 m5 + m6 )2 + A01 x1 + A11 (x2 m5 + m6 )x1

+ A02 x12 m4

(6.2)

The dynamics of the second state variable is obtained considering the mass conservation principle inside the humidier. It takes into account the mass ow of air
injected by the compressor, given by the relation obtained from (5.10), which includes the working pressure of the compressor and the shaft angular velocity:

6.2 State Space Model of the Experimental Fuel Cell System

131

x2 = B00 + B10 (x2 m5 + m6 ) + B20 (x2 m5 + m6 )2 + B01 x1


+ B11 (x2 m5 + m6 )x1 + B02 x12 b1 (x)3 C3 b1 (x)2 C2
b1 (x)C1 C0

(6.3)

The remaining states of the model are related to the PEM fuel cell stack variables.
As already stated in Chap. 5, these are given by three states of the cathode channels
and two of the anode. Thus, the third state space equation describes the behaviour
of the oxygen mass within the cathode channels that comes from Eq. (5.26):


1
x3 = XO2 ,ca m9 b1 (x)3 C3 + b1 (x)2 C2 + b1 (x)C1 + C0 G1
a (x2 m5 m10 )

1+

m11
x2 m5 m10

1


+ b1 (x)3 C3 + b1 (x)2 C2 + b1 (x)C1 + C0


1 
1
m11
m14
1+
1+
x2 m5 m10
(x3 RO2 + x4 RN2 + x5 Rv )m8 m12

Kca (x3 RO2 + x4 RN2 + x5 Rv )m8 Pamb x3 RO2 GO2






1
x3 RO2
x3 RO2 GO2
1 + Gv x5 Rv
+ 1
GN2
x 3 RO 2 + x 4 R N 2
x 3 RO 2 + x 4 R N 2
1
(x3 RO2 + x4 RN2 )1
(x3 RO2 + x4 RN2 )1




1
x3 RO2 GO2
x 3 RO 2
GO nIst

+ 1
1/4 2
GN2
x 3 R O 2 + x 4 RN 2
x 3 RO 2 + x 4 RN 2
F
(6.4)
The rst term of Eq. (6.4) corresponds to the incoming oxygen mass fraction, while the second models the cathode output ow using a linear nozzle equation (5.44). The third term takes into account the oxygen mass consumed per second
by the cathodic reaction (5.31).
In accordance with Sect. 5.5.1, the dynamics of the fourth state variable, the
mass of nitrogen in the cathode, is similar to that obtained for x3 , apart from the
fact that there is no nitrogen consumption in the reaction. Then, from the analysis of
Eq. (5.27), there are only two terms involved, relating the nitrogen input and output
ow fractions:


1
x4 = m9 b1 (x)3 C3 + b1 (x)2 C2 + b1 (x)C1 + C0 G1
a (x2 m5 m10 )

1+

1+

m11
x2 m5 m10
m11
x2 m5 m10

1


+ b1 (x)3 C3 + b1 (x)2 C2 + b1 (x)C1 + C0


1 
(1 XO2 ,ca,in ) 1 +



x3 m8 GO2
1 x3 m8 GO2 b3 (x)1
b3 (x)

Gv m12
Ga (b2 (x) m12 )


1 
x3 m8
+ 1
GN 2
b3 (x)

1

132

6 SOSM Control Design and Implementation





1
x3 m8
x3 m8 GO2
1 + Gv x5 Rv m8
+ 1
GN 2
b3 (x)
b3 (x)
1

(x3 RO2 m8 + x4 RN2 m8 )1


Kca,n b2 (x) Pamb

(6.5)

The vapour mass equation in the cathode takes into account four different ows
(5.28), namely the input and output partial vapour ows, the water generated by the
reaction (which depends on the stack current Ist ) and the water ow transported by
the polymeric membrane (Sect. 5.5.3):

1
x5 = Gv m12 b1 (x)3 C3 + b1 (x)2 C2 + b1 (x)C1 + C0 G1
a (x2 m5 m10 )

1
Gv m10
b4 (x)
1+
+
Ga (x2 m5 m10 )
b5 (x)

1

b4 (x)
Gv m12 b4 (x)
Gv m12
+

1+
Ga (x2 m5 m10 )b5 (x) b5 (x)
Ga (b2 (x) m12 )

Kca,out b2 (x) Pamb + Kca,out b2 (x) Pamb






1
1
x3 m8
1
1 x3 m8 GO2
1 + Gv x5 Rv m8 b3 (x)
RO 2
+ 1
GN 2
b3 (x)
b3 (x)


Gv nIst
+ 1/2
+ n0 + n1 a0 + a1 b6 (x) + a2 b6 (x)2 + a3 b6 (x)3
F

+ n2 a0 + a1 b6 (x) + a2 b6 (x)2 + a3 b6 (x)3 Ist /Afc /F



(a0 + a1 x5 m16 + a2 x52 m216 + a3 x53 m316 )m,dry
Dw
Gm,dry
 
(a0 + a1 x7 m15 + a2 x72 m215 + a3 x73 m315 )m,dry 1

(6.6)
tm Gv Afc n
Gm,dry
The rst anode state (Sect. 5.5.2) is derived from the mass conservation principle
applied to the hydrogen in the anode channels. Following Eq. (5.47), its dynamics is
computed in a similar way as x3 . It relates the input and output partial ows, as well
as the hydrogen consumed by the electrochemical reaction on the anodes catalyst
surface:
1


Gv m17
x6 = Wan,in 1 +
Kan,out b7 (x) Pamb
GH2 (b7 (x) m17 )


GH nIst
Gv x7 m19 1
1/2 2
(6.7)
1+
GH2 x6 m20
F
The last state space equation models the vapour mass dynamics in the anode
channels. Regarding Eq. (5.48), it only consists of three ow terms, given the assumption that no water is generated at the anode side:
1


Gv m17
x7 = Wan,in Wan,in 1 +
Kan,out b7 (x) Pamb
GH2 (b7 m17 )

6.2 State Space Model of the Experimental Fuel Cell System

Gv x7 m19
1+
GH2 x6 m20

133

1

+ Kan,out b7 (x) Pamb





n0 + n1 a0 + a1 b6 (x) + a2 b6 (x)2 + a3 b6 (x)3

+ n2 a0 + a1 b6 (x) + a2 b6 (x)2 + a3 b6 (x)3 Ist /Afc /F



(a0 + a1 x5 m16 + a2 x52 m216 + a3 x53 m316 )m,dry
Dw
Gm,dry
 
2
2
(a0 + a1 x7 m15 + a2 x7 m15 + a3 x73 m315 )m,dry 1

tm Gv Afc n
Gm,dry

(6.8)

6.2.1 Control Output


From the model developed in Chap. 5, a wide variety of control objectives can be
tackled. Hence, output variables such as partial pressures of oxygen and hydrogen,
water content in the membranes, relative humidity in the channels, etc., can be chosen. However, to address the classical problem of oxygen regulation, the compressor air mass ow is considered as the control output in what follows. Note that Wref
comes from the Eq. (5.10) and depends on x1 and x2 :
y(t) = Wcp
= B00 + B10 (x2 m5 + m6 ) + B20 (x2 m5 + m6 )2 + B01 x1
+ B11 (x2 m5 + m6 )x1 + B02 x12

(6.9)

6.2.2 Auxiliary Functions


b1 (x) = x2 m5 (x3 RO2 + x4 RN2 + x5 Rv )m8

(6.10)

b2 (x) = (x3 RO2 + x4 RN2 + x5 Rv )m8


x3 RO2 m8 + x4 RN2 m8
b3 (x) =
RO 2

(6.11)

b4 (x) = (x2 m5 b2 )3 C3 + (x2 m5 b2 )2 C2 + (x2 m5 b2 )C1 + C0


Gv m10
b5 (x) = 1 +
Ga (x2 m5 m10 )
b6 (x) = 1/2x7 m15 + 1/2x5 m16

(6.13)

b7 (x) = (x6 RH2 + x7 Rv )m18

(6.16)

(6.12)

(6.14)
(6.15)

Appendix C resumes the values of the parameters used in the model equations
presented above. Most of them come directly from physical constants of the experimental system. The proposed model that can be straightforwardly adjusted or scaled
to other PEM fuel-cell-based systems.

134

6 SOSM Control Design and Implementation

Fig. 6.1 Pnet vs. Wcp . Cathode pressure Pca = 1 bar

6.3 Control Objective


As previously mentioned, the main objective of the proposed control strategy is the
optimisation of the energy conversion of the fuel cell, maximising the net power
generated by the system under different load conditions. This is specially important, for instance, in automotive applications, where the extra energy generated by
the conversion system (not used by the vehicle electric power train), can be stored
in a battery or super-capacitor bank. This improves autonomy and fuel efciency.
Considering that the compressor is also driven by the fuel cell (in fact, it can be regarded as a parasitic load), the output net power (Pnet ) can be dened as the electrical power delivered by the stack (Pst = Vst Ist ) minus the electrical power consumed
by the compression subsystem (Pcomp = Vcp ia ). Then, the optimisation of the system efciency can be achieved by regulating the air mass ow entering to the stack
cathode at different load conditions. Let us recall that accomplishing such optimal
comburent ow is equivalent to maintaining the cathode line oxygen stoichiometry
(or oxygen excess ratio) in an optimal value [5].
In the particular case of the laboratory FC system, the most efcient operating
zones were determined from the experimental analysis of the conversion system
under different working conditions (see as examples Fig. 6.1 at Pca = 1 bar and
Fig. 6.2 at Pca = 1.6 bar). It can be seen that, for different cathodes working pressures, the system presents variations on the net power generated for different air
ows and cells currents. This indicates that, for a given load current, the maximisation of the conversion performance can be achieved through the regulation of the
input air mass ow as a function of Ist . In this way, the proper comburent ow can
be ensured in the membranes in order to satisfy load requirements.

6.3 Control Objective

135

Fig. 6.2 Pnet vs. Wcp . Cathode pressure Pca = 1.6 bar

As stated in Chaps. 4 and 5, a direct relation exists between Pnet and the system conversion efciency, given that for a constant current demand, the amount of
hydrogen consumed by the anodic reaction remains constant.
In Figs. 6.3 and 6.4 the experimental static characteristics of the plant are displayed as functions of the cathodes oxygen stoichiometry (also for Pca = 1 bar and
Pca = 1.6 bar, respectively). It becomes evident that setting the air ow at its optimum value for each case is equivalent to maintaining an optimal value of oxygen
stoichiometry. This value can be obtained off-line or using on-line trajectory planning.
From Chap. 4, oxygen stoichiometry or oxygen excess ratio is dened as follows:
O2 =

WO2 ,in
WO2 ,react

(6.17)

where WO2 ,in is the partial oxygen ow that inputs the cathode, and WO2 ,react is the
amount of oxygen consumed by reduction at the cathode:
nIst
(6.18)
4F
From these gures it becomes evident that starting at low oxygen stoichiometry
values, as O2 increases (and so does the partial oxygen pressure), the stack power
(and therefore Pnet ) is also increased. However, after reaching a region of optimal
values, further increments on O2 cause a decrement on Pnet due to the excess power
drained by the compressor. This effect is veried over the entire stack working range
(1 bar < Pca < 2.5 bar). Analysing the optimal values for O2 , Figs. 6.3 and 6.4
suggest that the use of oxygen excess ratio values should be within a range from 2
to 3. It is important to stress that values lower than one (O2 < 1) must be avoided
because, if WO2 ,in < WO2 ,react , there is the risk of cathode starvation.
WO2 ,react = GO2

136

6 SOSM Control Design and Implementation

Fig. 6.3 Pnet vs. O2 characteristic. Cathode pressure Pca = 1 bar

Once the desired values for the oxygen excess ratio for each load demand are determined, the control strategy can be formulated from concepts already presented in
Chap. 4. The immediate way to attain the control goal of O2 regulation would be to
directly control WO2 ,in . However, this is impossible to do in practice, given that it is
an inaccessible internal variable. To circumvent this problem, an alternative formulation in terms of control of the compressor air mass ow previously proposed. The
compressor output mass ow (Wcp ), which is easily measurable, is directly related
with WO2 ,in through the humidier dynamics. Besides, once this dynamics is extinguished, a constant relationship between Wcp and WO2 ,in exists. Therefore, setting
proper Wcp reference values, an operation in the neighbourhood of the maximum efciency can be achieved, and, after the humidier transient is elapsed, O2 = O2 ,opt
can be ensured for each load condition [5].
In the sliding mode framework, this objective can be attained using the following
restriction or sliding surface:
s(x, t) = Wcp Wair,ref

(6.19)

where s is the sliding variable that must be steered to zero. The reference value
(Wair,ref ) can be directly obtained from the desired dry air ow. Given that the oxygen mass fraction (O2 ) is known for the current ambient conditions, the required
dry air ow (Wdry,ref ) can be found from the following expression:
Wdry,ref =

nIst
1
O2 ,opt GO2
O2
4F

(6.20)

Then, considering that the relative humidity of the air (amb ) is also known, the
nal expression of the reference to be followed by the control system is
Wair,ref = (1 + amb )

nIst
1
O ,opt GO2
O2 2
4F

(6.21)

6.4 Controller Synthesis

137

Fig. 6.4 Pnet vs. O2 characteristic. Cathode pressure Pca = 1.6 bar

It is worth noting that for stable ambient conditions (amb and O2 constant), this
value depends on a single measurable variable, Ist .
Therefore, according to Chap. 3, the control problem can be mathematically
stated as follows:

x = f (x, t) + gu(t)
(6.22)
s = s(x, t) R

u = U (x, t) R
where x R7 , t is time, u is a bounded control action, and f , s are smooth functions.
The control task is to keep s 0.

6.4 Controller Synthesis


Analysing the sliding variable and its Lie derivative along the vector eld g (Lg s),
it becomes clear that s(x, t) has relative degree one (i.e. s(x, t) does not explicitly
depend on the control input):
s(x, t) = B00 + B10 (x2 m5 + m6 ) + B20 (x2 m5 + m6 )2 + B01 x1
+ B11 (x2 m5 + m6 )x1 + B02 x12 Wair,ref
while u explicitly appears in its derivative s (x, t) = s
t + Lf s + Lg su:


Lg s(x, t) = B01 + B11 (x2 m5 + m6 ) + 2B02 x1 m1 m4
= 0

(6.23)

(6.24)

Given that s(x, t) has relative degree 1 with respect to the control action u, it
could have been possible to implement a classic (rst-order) sliding-mode controller. However, as it was discussed in Chaps. 3 and 4, this option does not represent

138

6 SOSM Control Design and Implementation

an attractive alternative for this particular type of plant. A switching control action
may provoke undesirable oscillations in the compressor DC motor and also affect
the net output power, which has a relative degree 0 with respect to the compressor
voltage (Vcp ). In this context a second-order sliding-mode controller proves to be
an interesting choice to robustly control the system, achieving smoothness on Vcp ,
thus avoiding deterioration of the quality of Pnet . Differentiating twice the sliding
variable with respect to time, we obtain the following expressions:

s(x, t) +
s(x, t). f (x, t) + gu(t)
t
x

= s(x, t) + B01 + B11 (x2 m5 + m6 ) + 2B02 x1 m1 u(t) m2 x1


t
x1 m3 + B0 + A00 + A10 (x2 m5 + m6 ) + A20 (x2 m5 + m6 )2 + A01 x1

+ A11 (x2 m5 + m6 )x1 + A02 x12 m4 + B10 m5 + 2B20 (x2 m5 + m6 )m5


+ B11 m5 x1 B00 + B10 (x2 m5 + m6 ) + B20 (x2 m5 + m6 )2 + B01 x1 + B02 x12

3
+ B11 (x2 m5 + m6 )x1 x2 m5 (x3 RO2 + x4 RN2 + x5 Rv )m8 C3

2


x2 m5 (x3 RO2 + x4 RN2 + x5 Rv )m8 C2 x2 m5

(6.25)
(x3 RO2 + x4 RN2 + x5 Rv )m8 C1 C0

s =

s (x, t) +
s (x, t). f (x, t) + g +
s (x, t).u(t)

t
x
u
= (x, u, t) + (x, u, t)u(t)

s =

(6.26)

with
(x, u, t) = Lg s(x, t)

(6.27)

(x, u, t) = Lf s (x, t) + Lg s (x, t)

(6.28)

See Appendix C for the complete expression of (x, u, t). The functions
(x, u, t) and (x, u, t) can be globally bounded as follows:
0 < m (x, u, t) M

(6.29)



(x, u, t)

(6.30)

Particularly, for the PEMFC system under study, the bounding values were computed by means of a numerical study of the nonlinear system and rened through an
experimental analysis. Additionally, uncertainties were included in representative
parameters such as the motor inertia, torque friction, humidier volume and cathode air constant. Under the effect of an appropriate control action (a feedforward,
as in Chap. 4, proved to be suitable), to lead the system to the vicinity of the desired
air mass ow, the following values were obtained:
= 2.3 105 ;

m = 0.002;

M = 0.0083

(6.31)

6.4 Controller Synthesis

139

Once the bounds have been determined, the stabilisation problem of system
(6.22) with inputoutput dynamics (6.26) can be solved through the solutions of
the following equivalent differential inclusion by applying SOSM:

s 2.3 105 ,


2.3 105 + [0.002, 0.0083]u

(6.32)

It is worth noting that this study is performed off-line and does not burden the
real-time controller operation.
As mentioned in Chap. 3, a variety of SOSM algorithms can be found in the literature to solve the control problem, each of them with its distinctive features. In
this chapter, the SOSM controllers analysed and detailed in Chaps. 3 and 4, namely
Super-Twisting, Twisting and Sub-Optimal, have been designed and implemented
in the laboratory fuel-cell-based system. Previous to the experimental tests, the controllers have been assessed by thorough simulations.
The gains of the SOSM algorithms were calculated from , m and M guaranteeing that, once the system is steered to the region where Eq. (6.32) holds, the
trajectories do not escape and converge to s = s = 0 in nite time.
It is worthwhile to emphasise that the considered algorithms depend only on
few parameters, which were computed during the off-line tuning procedure. Thus,
the on-line operation of the control algorithms is very simple and consumes scarce
computational resources. In particular, the Twisting and Sub-Optimal algorithms
are intended for a sliding variable of relative degree 2. Given that the considered
output of the PEMFC-based system is of relative degree 1, for these last two cases,
the system has been expanded with an integrator, considering = u(t)
as the control
action for the design. The algorithms structures and the chosen gains for the PEMFC
controllers are succinctly recalled below.

6.4.1 Super-Twisting Algorithm


As discussed in Chap. 3, this algorithm is specically intended for systems with
relative degree 1. One interesting feature is that, during on-line operation, it does
not require information of s . The output converge to the origin of the sliding plane
following a characteristic looping trajectory. The control law comprises two terms.
One is the integral of a discontinuous control action, and the other is a continuous
function of s, contributing only during the reaching phase [6]:
u(t) = u1 (t) + u2 (t)
u 1 (t) = sign(s)
u2 (t) = |s|

1/2

(6.33)

sign(s)

where and are design parameters that were derived from the corresponding
sufcient conditions for nite-time convergence of the algorithm [6]:

140

6 SOSM Control Design and Implementation

m

2 (m + )2
>
m2 (m )

>

(6.34)

Among the set of gains that full (6.34), the best control performance has been
achieved with
= 1;

= 35

(6.35)

6.4.2 Twisting Algorithm


Let us recall that this algorithm is characterised by the way in which its trajectories
converge to the origin in the sliding plane (s, s ). Knowledge of the signs of s and s
is needed, and the control law is given by [2]
(t) = u(t)
= r1 sign(s) r2 sign(s )

(6.36)

Sufcient conditions for nite-time convergence can be summarised as


r1 > r2 > 0
(r1 + r2 )m > (r1 r2 )M +

(r1 r2 ) >
m

(6.37)

Therefore, the chosen gains resulted in


r1 = 0.7;

r2 = 0.1

(6.38)

6.4.3 Sub-Optimal Algorithm


Trajectories in the (s, s ) plane are conned within parabolic arcs which include
the origin, so the convergence behaviour may exhibit twisting around the origin,
bouncing on the s axis or a combination of both. Its control law is given by [1]
(t) = u(t)
= (t)U sign(s sM )

1
if (s sM )sM 0
(t) =

if (s sM )sM < 0

(6.39)

where U > 0 is the minimum value of u(t),

> 1 is the modulation parameter, 0


< 1 is an anticipation factor, and sM is a piece-wise constant function representing
the last extremal value of the sliding variable s(t).

6.5 Simulation Results

141

The convergence in nite time is guaranteed as long as


U>

[1; +)


+ (1 )M U
; +
m U

(6.40)

This algorithm requires the detection of s becoming zero and the corresponding
values of s at those instants, i.e. sM . The nal choice of the controller parameters
for the PEMFC system under consideration was
U = 1;

= 1.2;

= 0.5

(6.41)

6.4.4 Feedforward Term


It was previously stated that it is necessary to dene an extra control action that
steers the sliding variable within a region such that the bounds on the sliding dynamics given by (6.32) are satised [2]. With this purpose, the aforementioned feedforward (FF) action uff has been included. It provides the control effort required to
approach surface neighbourhood, reaching the validity region. Therefore, the twoterm control action (ui ) proposed in (4.41) is implemented as
ui (t) = u(t) + uff (t)

(6.42)

where u corresponds to the SOSM control action particularised above. The expression of uff is computed via a polynomial obtained from an off-line test covering the
entire operation range of the PEMFC system. The complete polynomial expression
is given in Appendix C.

6.5 Simulation Results


Previous to the implementation and experimental assessment, the controllers were
tested through numerical simulations.
The following two gures show representative simulation results of the controlled system using the algorithms designed above. Figure 6.5 conrms that the
Super-Twisting, Twisting and Sub-Optimal controllers present a satisfactory dynamic response when controlling the air mass ow.
In Fig. 6.6, the response of the nonlinear system using the proposed algorithms
is shown through an ss diagram. It is important to stress that, after an adequate
parameters tuning, the three controllers present comparable satisfactory dynamic
behaviours, conrming that the suitability of the SOSM approach for the breathing
control of this PEMFC system is veried.

142

6 SOSM Control Design and Implementation

Fig. 6.5 Air ow and control input (simulation results)

Fig. 6.6 s vs. s (simulation results)

6.6 Experimental Set-up of the PEM Fuel Cell System

143

Fig. 6.7 Experimental PEM


fuel cells laboratory at IRI
(UPC-CSIC)

6.6 Experimental Set-up of the PEM Fuel Cell System


A detailed description of the laboratory test station used for the controllers development and testing was presented in Chap. 5 (a view is shown in Fig. 6.7). A schematic
diagram of the system interconnection is depicted in Fig. 6.8, where sensors and actuators are also displayed.
The main subsystems can be summarised as follows:
Air Compressor: 12-V DC oil-free diaphragm vacuum pump. The input voltage
to this device is a continuous variable used as the main control action.
Hydrogen and Oxygen humidiers and line heaters. These are used to maintain
proper humidity and temperature conditions inside the cell stack, an important
issue for PEM membranes. Cellkraft membrane exchange humidiers are used
in the current set-up, and decentralised PID controllers ensure adequate operation
values.
Fuel cell stack: an ElectroChem seven-cell stack with Naon 115 membrane
electrodes assemblies (MEAs) is used, with a catalyst loading of 1 mg/cm2 of
platinum, 50 cm2 of active area.
Different sensors were incorporated to measure specic variables, suited for
modelling and control. Regarding Fig. 6.8, these are: motor shaft angular velocity
(cp ), compressor air mass ow (Wcp ), hydrogen mass ow (WH2 ), cathode and an-

144

6 SOSM Control Design and Implementation

Fig. 6.8 PEM Fuel cell system schematics

ode humidiers pressures (Phum,ca and Phum,an ), stack pressure drops (Pca and Pan ),
motor stator current (ia ) and voltage (Vcp ), stack voltage (Vst ) and current (Ist ). Besides, a number of sensors were included to monitor and register relevant temperatures (Tst , Thum,ca , Tlh,ca , Thum,an and Tlh,an ). It must be noted that in a typical fuel
cell application many of these measurements are not necessary [5]. For instance,
the proposed controllers only require the measurement of the stack current and the
compressor air mass ow. The operating conditions for all fuel cell test bench experiments are summarised in Appendix C.

6.7 Experimental Tests


6.7.1 Regulation Tests
The main objective of this section is to present the experimental performance of the
proposed SOSM controllers, considering actual external disturbances and different
working conditions. Comparative results are provided for Super-Twisting, Twisting
and Sub-Optimal controllers.
To assess the rst controller (6.33) performance in real operation, Fig. 6.9
presents the behaviour of the Super-Twisting Algorithm controller at different regulation conditions (without FF action). It depicts the evolution of the compressor air
ow versus a series of steps in the reference. Simulation and experimental results
are superimposed. Close matching of simulation and experimental results shown in
Fig. 6.9 conrm the reliability and accuracy of the design methodology.
It can be seen that, although the closed loop succeeds in driving the air ow
to its desired value in nite time, the reaching dynamics can be improved. As it
was stated in Chap. 4, the control action on the reaching phase of the second-order

6.7 Experimental Tests

145

Fig. 6.9 Control of Wcp using a Super-Twisting algorithm: simulation and experimental results

Fig. 6.10 Control of Wcp with Super-Twisting + FF: experimental results

sliding-mode controllers can be freely designed if it is ensured that u(t) is continuous and drives the sliding variable to a neighbourhood of the surface s = 0 in nite
time.
Then, Fig. 6.10 shows the improvement in the dynamic response obtained when
the SOSM controller is combined with the static FF action given in Appendix C.
In Fig. 6.11, the behaviour of each component of the control action (u and uff ) is
presented, as well as the armature voltage of the compressor motor (Vcp = ui ).
Analogous tests were conducted for the Twisting and Sub-Optimal algorithms.
The results are presented in Figs. 6.11 and 6.12, showing that the three controllers
present similar dynamic performance when the parameters tuning is performed following similar criteria.

146

6 SOSM Control Design and Implementation

Fig. 6.11 Control of Wcp with Twisting + FF: experimental results

Fig. 6.12 Control of Wcp with Sub-Optimal + FF: experimental results

In the following analysis, a test of oxygen stoichiometry regulation was conducted to verify its performance. This is shown in Fig. 6.13, where O2 is displayed
versus time, for the same reference signal. This test was performed at constant load
conditions.
In Fig. 6.14 the dynamical behaviour of Vst , Ist , ia and Pnet for the same test is
presented.
Afterwards, an experiment under variable load conditions was performed to verify the controlled system response in more demanding (and realistic) situations. Figures 6.15 and 6.16 show that the system exhibits good regulation characteristics for

6.7 Experimental Tests

147

Fig. 6.13 O2 regulation, Ist = 3 A: experimental results

Fig. 6.14 O2 regulation, Ist = 3 A: experimental results

changing loads. Figures 6.17 and 6.18 show a similar test, but in this case performed
at a standard regulation point of the oxygen stoichiometry (O2 = 2.5).
An important issue from the practical point of view is related with actuator saturation. Given that the compressors input cannot exceed the power source voltage (0 V < ui < 12 V), a wind-up effect may appear, due to the integral part of
the controller. To correct this, an anti-wind-up algorithm was included which disconnects the integrators input when such saturation is detected. The effect of the
anti-windup action is shown in Fig. 6.19. Note that the ow value recovers quickly,
converging to the desired values once saturation ends and nominal operation is restored.

148

6 SOSM Control Design and Implementation

Fig. 6.15 O2 regulation, variable Ist : experimental results

Fig. 6.16 O2 regulation, variable Ist : experimental results

6.7.2 Perturbation Tests


Having veried the control operation in nominal regime, the system was then tested
under the inuence of external perturbations. In particular, an increment in the
cathode pressure was forced using a mechanical back pressure regulator. This effect can be appreciated in Figs. 6.206.22, where similar tests were induced to the
three algorithms implementations (ST, Twisting and Sub-Optimal). For instance, in
Fig. 6.21 it is shown that during the 0190 sec interval, as the valve is increasingly throttled, reference tracking is successfully preserved. Then, when the valve

6.7 Experimental Tests

149

Fig. 6.17 O2 regulation: experimental results

Fig. 6.18 O2 regulation: experimental results

is suddenly bypassed (t = 190 sec), the system abandons the sliding surface, but the
SOSM controller provides a quick recovery. It is worth noting that during this experiment, the feedforward action remains constant, because it only depends on the
reference.
Finally, in Fig. 6.23 a test analogous to the one presented in Fig. 6.20 is depicted.
Nevertheless, this last perturbation test is done at O2 = 3 that is an operating point
in the area of maximum power generation.

150

6 SOSM Control Design and Implementation

Fig. 6.19 Actuator saturation and anti wind-up effect

Fig. 6.20 Regulation of O2 under external perturbations: experimental results using Super-Twisting

6.8 Conclusions
After theoretically and experimentally evaluating different second-order slidingmode controllers that globally solve the oxygen stoichiometry problem of a PEMFC
generation system, a set of SOSM controllers have been developed and implemented
in a laboratory test station. Its suitability was successfully veried through extensive
computer simulations, based on the plant model previously developed in Chap. 5,
which was especially built for nonlinear control purposes, taking into account external disturbances and uncertainties in the system parameters [4]. Subsequently,
highly satisfactory experimental results using Super-Twisting, Twisting and Sub-

6.8 Conclusions

151

Fig. 6.21 Regulation of O2 under external perturbations: experimental results using Twisting

Fig. 6.22 Regulation of O2 under external perturbations: experimental results using Sub-Optimal

Optimal topologies with feedforward action conrm the feasibility, simplicity and
robustness of the solutions. The main advantages of the proposed SOSM control for
PEMFC systems can be summarised as follows:
robust stabilisation of the oxygen stoichiometry problem avoiding chattering effects;
enhanced dynamic characteristics;
robustness to parameter uncertainties and external disturbances;
guaranteed extended range of operation, in spite of the highly nonlinear nature of
plant;

152

6 SOSM Control Design and Implementation

Fig. 6.23 Regulation of O2 under external perturbations: experimental results

the control law only depends on two measurable variables, namely the stack current and the compressor air ow, and therefore no state observer is required;
simple controller structure, resulting in low real-time computational burden.
The resulting controllers are relatively simple to design and require only a few
measurements of easy and low-cost implementation. This represents a major advantage for fuel cells in industrial applications, where the cost of the plant instrumentation is critical.

References
1. Bartolini G, Ferrara A, Usai E (1997) Applications of a sub-optimal discontinuous control
algorithm for uncertain second order systems. Int J Robust Nonlinear Control 7(4):299310
2. Fridman L, Levant A (2002) Higher order sliding modes. In: Sliding mode control in engineering, Dekker, New York, pp 53101 (Chap 3)
3. Kunusch C (2009) Modelling and nonlinear control of PEM fuel cell systems. Phd thesis, Electrical Department, National University of La Plata, Argentina (in Spanish)
4. Kunusch C, Puleston PF, Mayosky MA, Husar A (2011) Control-oriented modelling and experimental validation of a PEMFC generation system. IEEE Trans Energy Convers 26(3):851861
5. Kunusch C, Puleston PF, Mayosky MA, Dvila A (2010) Efciency optimisation of an experimental PEM fuel cell system via super twisting control. In: Proceedings of IEEE 11th international workshop on variable structure systems, Mexico City, June 2628
6. Levant A (1993) Sliding order and sliding accuracy in sliding mode control. Int J Control
58(6):12471263

Chapter 7

Conclusions, Open Lines and Further Reading

7.1 Main Results


This book addressed the application of Second-Order Sliding-Mode (SOSM) algorithms for robust control of autonomous PEM fuel cells. These techniques allow
the optimisation of the energy conversion performance of the system and are robust
to external disturbances and model uncertainty. SOSM controllers retain many of
the desirable properties of First-Order Sliding-Mode controllers (robustness, nite
reaching time to the sliding surface, implementation simplicity, etc.) and exhibit a
better behaviour with respect to the chattering effect. The design procedure involved
the following steps:
Development of a control-oriented, smooth dynamic model of the plant. This was
accomplished using a semi-empirical approach, combining the knowledge of the
system and numerical approximations, in order to avoid discontinuities and lookup tables.
Design of the proper sliding surface to meet the performance requirements. The
sliding manifold proposed ensures the optimisation of the energy conversion performance.
Calculation of a set of norms for the model.
Choice of the specic algorithm best suited for the application. Twisting, SuperTwisting and Sub-Optimal controllers were tested, and a performance comparison
was made. For the specic devised objective, namely the oxygen stoichiometry
control, the stack current Ist and the air ux provided by the compressor Wcp are
the only measurements required.
Simulation and experimental results were presented, showing that excellent performance can be obtained on a wide range of operating conditions. These results
encourage the application of SOSM controllers to the remaining subsystems of the
PEM fuel cell assembly, in particular the thermal and water subsystems. Several
improvements to the existing control scheme and many open directions of research
can be proposed. In the following sections, the most promising issues are outlined.
C. Kunusch et al., Sliding-Mode Control of PEM Fuel Cells,
Advances in Industrial Control,
DOI 10.1007/978-1-4471-2431-3_7, Springer-Verlag London Limited 2012

153

154

7 Conclusions, Open Lines and Further Reading

7.2 Further Issues Related to PEMFC Control


7.2.1 Adaptive Super-Twisting Algorithms
The SOSM approaches analysed in Chaps. 4 and 6 of this book allow for nitetime convergence to zero of not only the sliding variable but its rst derivative too.
Among them, the Super-Twisting algorithm has the additional advantage of a simple
implementation, requiring only the knowledge of the sliding variable. The resulting continuous controller ensures all the main properties of rst-order sliding-mode
control for a system with smooth matched bounded uncertainties/disturbances, allowing also reduction of the chattering effect. The classic Super-Twisting algorithm
equations (already presented in Chap. 3 and repeated here for convenience) are:
u(t) = u1 (t) + u2 (t)

(7.1)

where
u 1 (t) = sign(s)

|s0 | sign(s)
u2 (t) =
|s| sign(s)

if |s| |s0 |

(7.2)

if |s| |s0 |

where > 0, > 0 and (0, 12 ) are parameters dened from suitable bounds on
the system model.
The homogeneous nature of the classic Super-Twisting algorithm does not allow
the compensation of uncertainties/disturbances that change with the state variables.
This means that it cannot ensure the sliding motion for systems with, for instance,
an unknown linear part. Using Lyapunov-based techniques, a non-homogeneous extension of the Super-Twisting algorithm can be proposed, allowing the exact compensation of smooth uncertainties/disturbances with derivatives bounded by known
functions [8]. These are called Variable-Gain Super-Twisting (VGST) algorithms.
The general form of VGST controllers is
 t
v = k1 (t, x)1 (s)
k2 (t, x)2 (s) dt
(7.3)
0

where
1

(7.4)
1 = |s| 2 sign(s) + k3 s, k3 > 0
1
1
3
2 = sign(s) + k3 |s| 2 sign(s) + k32 s
(7.5)
2
2
Note that when k3 = 0 and the gains k1 and k2 are constant, the classic SuperTwisting controller results. If k3 > 0, the algorithm can reject perturbations growing
linearly in s, i.e. outside of the sliding surface, and the variable gains k1 and k2
can be used to make the sliding surface insensitive to perturbations growing with
bounds given by known functions. In these cases, the VGST algorithm exhibits a
better behaviour with respect to chattering reduction.

7.2 Further Issues Related to PEMFC Control

155

7.2.2 HOSM MIMO Control


Although the techniques analysed in this book are general and can be applied to
MIMO systems, the reported results are shown for a nonlinear SISO model of the
oxygen subsystem. To this end, in Chap. 6 it was supposed that auxiliary controllers
regulate temperature, humidity levels and hydrogen supply at critical points of the
system. However, these control goals could be part of a more global strategy to
be addressed, for instance, with sliding mode controllers. In fact, a MIMO control
would allow fullment of complex, combined objectives. For instance, if an electromechanical valve is installed at the cathodes output, a different control strategy
can be proposed, combining regulation of oxygen stoichiometry and control of the
oxygen partial pressure. Additionally, the regulation of the hydrogen circulation to
the cathode can be accomplished if a mass ux controller at the input of the anodes
humidier is available.

7.2.3 Model Predictive Control


Model Predictive Control (MPC), also known as moving horizon control or receding
horizon control, refers to a class of algorithms which make explicit use of a process
model to optimise the future predicted behaviour of a plant. More particularly, at
each sampling time t , a nite-horizon optimal control problem is solved over a prediction horizon N , using the current state x of the process as the initial state. The
online optimisation problem takes account of system dynamics, constraints and control objectives.
The consideration of model predictive control strategies in dynamical systems
such as fuel-cell-based systems is nowadays a growing topic in the literature. However, the use of those control strategies is becoming more widespread due to the particular capabilities of the technique, which allow solving crucial problems related to
the control and management of the fuel cells and ancillary systems dynamics. MPC
can deal with system constraints in a systematic and straightforward way [18]. Additionally, it can be applied to MIMO systems with multiple objectives, providing
also robustness against system disturbances. Nevertheless, MPC is very sensitive to
the model accuracy since the control computation is precisely based on a mathematical model of the plant. This dependence opens several ways on how to design the
MPC depending on the system model: from the purely nonlinear MPC [22] to linear
approaches [2, 5], piecewise afne models [10] and hybrid systems forms [9].
Regarding the multi-objective control nature of MPC, in [6] it is used for optimising the fuel cell power utilisation and the oxygen ow to the stack, so the strategy is
then implemented for global optimisation, improving both objectives. Additionally,
the use of real data for updating online the prediction allows to minimise the modelling mismatch, while the control law avoids the oxygen saturation, enhances the
transient performance and extends the operating life of the whole system. Regarding the incorporation of actuator limitations and state constraints in the controller

156

7 Conclusions, Open Lines and Further Reading

design, [7] discusses those aspects for the prevention of fuel cell starvation within
the MPC framework.
Given the advantages mentioned above, model predictive control of fuel cell systems has been analysed not only for the control of the fuel cell stacks in itself, but
also in its interaction to other electrical generation system. Research works such
as [9, 22] and [14] discuss aspects such as fuel cell management within micro cogeneration systems, decentralised MPC schemes for mixed topologies of fuel cells
with ultra capacitors and with Photovoltaic Panels. In particular, [21] propose an
adaptive MPC for the network control of fuel cells. All these approaches try to improve the global system performance through MPC controllers (in global and local
mode) in order to full the control objectives and reduce the computation burden.

7.2.4 Observers for Internal Variables


Observers are of great interest for fuel cell operation, both for monitoring purposes
and for feedback control design. They can give information about variables that
are difcult, expensive or even impossible to measure. Several publications of the
last decade deal with the design of PEMFC state observers. Hydrogen and oxygen
partial pressures can be estimated by different kind of observers such as adaptive
observers [1] and sliding-mode observers validated using Lyapunov stability analysis methods [16]. The membrane water content can be estimated with a nonlinear observer from its close relation with the resistive voltage drop [12], or by using
open-loop and Luenberger observers [20]. However, in the presence of disturbances,
Luenberger observers are not accurate, as they can only ensure the convergence to a
bounded region near the real value of the state. In these cases, Sliding-Mode-based
observers are regarded as an alternative to the problem of observation of perturbed
systems [19], being suitable for robust state estimation also in the presence of unknown inputs. Results reported in [3] estimate the internal fuel cell temperature
from surface measurements, while [15] propose the design of rst-order slidingmode observers for the estimation of the oxygen outow of the system compressor.
There, the inherent robustness of the sliding-mode techniques is exploited despite
the resultant chattering phenomenon. Finally, the fast response convergence of robust nonlinear unknown-input observers for the estimation of several variables in a
PEMFC-based system is reported in [4].
Higher-Order Sliding-Mode (HOSM) Based Observers represent a useful technique for the state observation of perturbed systems, due to their high precision and
robust behaviour with respect to parametric uncertainties. In particular, robust exact differentiators [17] based on the Super-Twisting algorithm can be applied for
state reconstruction. The methodology involves the transformation of the system to
a regular form (chain of integrators) using a suitable diffeomorphism. Then the state
is estimated applying HOSM differentiators to the system output. Even when differentiators appear as a natural solution to the observation problem, the use of the
system knowledge for the design of an observation strategy results in reduction of

7.3 Further Issues Related to Fuel-Cell-Based Systems

157

gains for the sliding-mode compensation terms, improving accuracy. Moreover, the
complete or partial knowledge of the system model can give place to the application
of techniques for parametric reconstruction or disturbance reconstruction [11].

7.3 Further Issues Related to Fuel-Cell-Based Systems


7.3.1 Hybrid Standalone Systems
Hybrid alternative energy power generation systems are expected to be an important
part of the power generation paradigm of the future [13]. In fact, none of the currently available technologies is cost efcient or reliable enough to operate as independent power sources. Hybrid systems have better potential to provide dependable
power than a system comprising a single resource. For this reason, hybrid energy
systems have caught worldwide research attention. Of course, each hybrid system
needs a proper control strategy to manage and prioritise power generation among the
different sources in the system. Hybrid energy systems can be made of many combinations of different sources and storage devices. For instance, a typical standalone
hybrid power system can be made of a wind energy conversion system (WECS),
a photovoltaic array and a PEMFC-based systems as energy sources. Due to the
intermittent nature of wind and solar energy, standalone wind and/or PV energy
systems normally require energy storage devices or some other generation sources.
The storage devices can be a battery bank, a super-capacitor bank (to handle fast
transient power demands), compressed air system, electrolyser (or a regenerative
FC) system, etc., or a combination thereof. For standalone applications, the system
needs to have sufcient storage capacity to handle the power variations of the alternative energy sources involved. Due to their different operation principles, each of
the components of the system should be controlled with different objectives, and a
supervisory strategy must be devised to optimise the energy ow, in order to full
load demands and ensure reliable operation.

7.3.2 Distributed Generation Systems


The work of this book analysed the use of PEMFC systems in standalone applications, that is, systems where the fuel cell is the only energy source available. However, the operational advantages of PEM technology, as well as its cost reduction,
suggest that it can be also applied to stationary systems. Control architectures for
these systems are of vital importance and regard PEMFC systems as part of a larger,
distributed generation systems (DGS). Steady progress in power deregulation and
tight constraints over the construction of new transmission lines for long-distance
power transmission have created increased interest in DGS. Of particular importance are renewable DGS with free energy resources, such as wind and photovoltaic,

158

7 Conclusions, Open Lines and Further Reading

and alternative energy conversion devices such as fuel cells and microturbines. DGS
are modular in structure, and they are normally placed at the distribution level or
near load centres. They can be strategically placed in distribution systems for grid
reinforcement, reducing power losses and on-peak operating costs. Additionally,
they can be used as control devices to improve voltage proles and load factors. As
a result, their installation can defer or eliminate the need for system upgrades and
can improve system integrity, reliability and efciency. However, important control
problems arise because many DG devices are intermittent and usually not synchronised with power consumption. Among others, grid balancing, smart power injection, distributed power supervision/monitoring and fault detection/recovery arise as
fundamental open issues to be resolved.

References
1. Arcak M, Gorgun H, Pedersen LM, Varigonda S (2004) A nonlinear observer design for fuel
cell hydrogen estimation. IEEE Trans Control Syst Technol 12(1):101110
2. Arce A, Ramirez DR, del Real AJ, Bordons C (2007) Constrained explicit predictive control
strategies for PEM fuel cell systems. In: 46th IEEE conference on decision and control, New
Orleans, LA
3. Begot S, Harel F, Kauffmann JM (2008) Design and validation of a 2 kW-fuel cell test bench
for subfreezing studies. Fuel Cells 8(1):2332
4. Benallouch M, Outbib R, Boutayeb M, Laroche E (2009) A new scheme on robust unknown
input nonlinear observer for PEM fuel cell stack system. In: 18th IEEE international conference on control applications. Part of 2009 IEEE multi-conference on systems and control,
Saint Petersburg, Russia
5. Bordons C et al (2006) Constrained predictive control. Strategies for PEM fuel cells. In: Proceedings of the 2006 ACC, Minnesota, USA, June 1416
6. Chen Q, Gaob L, Dougal RA, Quan S, (2009) Multiple model predictive control for a hybrid
proton exchange membrane fuel cell system. J Power Sources 191(2):473482
7. Danzer MA, Wittmanna SJ, Hofera EP (2009) Prevention of fuel cell starvation by model
predictive control of pressure, excess ratio, and current. J Power Sources 190(1):8691
8. Davila A, Moreno FA, Fridman L (2010) Variable gains super-twisting algorithm: a Lyapunov
based design. In: 2010 American control conference Marriott Waterfront, Baltimore, MD,
USA, 30 June02 July
9. Del Real AJ, Bordons Arcea AC (2007) Development and experimental validation of a PEM
fuel cell dynamic model. J Power Sources 173(1):310324
10. Fiacchini M, Alamo T, Alvarado I, Camacho EF (2008) Safety verication and adaptive model
predictive control of the hybrid dynamics of a fuel cell system. Int J Adapt Control Signal
Process 22:142160
11. Fridman L, Shtessel Y, Edwards C, Yan X (2008) Higher-order sliding-mode observer for
state estimation and input reconstruction in nonlinear systems. Int J Robust Nonlinear Control
18(45):399412
12. Gorgun H, Arcak M, Barbir F (2006) An algorithm for estimation of membrane water content
in PEM fuel cells. J Power Sources 157(1):389394
13. Hashem Nehrir M, Wang C (2009) Modeling and control of fuel cells. Distributed generation
applications. IEEE Press, New York
14. Houwing M, Negenborn RR, Ilic MD, De Schutter B (2009) Model predictive control of micro
cogeneration fuel cell systems. In: Proceedings of the 2009 IEEE international conference on
networking, sensing, and control (ICNSC 2009), Okayama City, Japan

References

159

15. Kazmi IH, Bhatti AI, Iqbal S (2009) A nonlinear observer for PEM fuel cell system. In: IEEE
13th international multitopic conference
16. Park I, Kim S (2006) A sliding mode observer design for fuel cell electric vehicles. J Power
Electron 6(2):172177
17. Levant A (1998) Robust exact differentiation via sliding mode technique. Automatica
34(3):379384
18. Maciejowski JM (2002) Predictive control with constraints. Prentice Hall, Harlow
19. Spurgeon SK (2008) Sliding mode observers: a survey. Int J Syst Sci 39(8):751764
20. Thawornkuno C, Panjapornpon C (2008) Estimation of water content in PEM fuel cell. Chiang
Mai J Sci 35(1):212220
21. Tong S, Liu G (2007) Networked control of PEM fuel cells using an adaptive predictive control
approach. In: Proceedings of the 2007 IEEE international conference on networking, sensing
and control, London, UK, pp 1517
22. Vahidi A, Stefanopoulou A, Peng H (2006) Current management in a hybrid fuel cell power
system: a model predictive control approach. IEEE Trans Control Syst Technol 14(6):1047
1057

Appendix A

Electric Vehicle PEM Fuel Cell Stack


Parameters

A.1 Return Manifold Polynomial Fitting


Table A.1 Return manifold
polynomial tting

Parameter

Value

Return manifold parameter (p0 )

0.001248 kg/s

Return manifold parameter (p1 )

0.001967 kg/s/KPa

Return manifold parameter (p2 )

0.001524 kg/s/KPa2

Return manifold parameter (p3 )

0.002122 kg/s/KPa3

Return manifold parameter (p4 )

0.02772 kg/s/KPa4

Return manifold parameter (p5 )

0.07804 kg/s/KPa5

A.2 Differential Equations Parameters


For further information about the model parameters, structure and reduction, refer
to [3], [2] and [1].
Table A.2 State equations parameters
Parameter

Expression

Value

B1

ncm /Rcm /Jcp

1.6667 104

B2

Kv ncm /Rcm /Jcp

255

B3

Cp Tatm max a /4dc2 KU c /(ncp 1/2 efmec Jcp )

7.4154 104

B4

( 1)/

0.2857

B5

2Cp Tcp,in KU2c

4.7393 107

B6

1/4max a dc2 KUc / 1/2 Tatm Ra /Vsm

51.7305

C. Kunusch et al., Sliding-Mode Control of PEM Fuel Cells,


Advances in Industrial Control,
DOI 10.1007/978-1-4471-2431-3, Springer-Verlag London Limited 2012

161

162

A Electric Vehicle PEM Fuel Cell Stack Parameters

Table A.2 (continued)


Parameter

Expression

Value

B7

Vsm Ksm,out /Ra Ra /Vsm

5.081 106

B8

B7 pv,ca

0.2391

B9

B7 RN2 Tst /Vca

53.2352

B10

B7 RO2 Tst /Vca

46.5988

B11

1/4max a dc2 KU c / 1/2

8.6437 106

B12

Ksm,out

3.6294 106

B13

Ksm,out pv,ca

170.815 103

B14

Ksm,out RN2 Tst /Vca

38.025

B15

Ksm,out RO2 Tst /Vca

33.284

B16

Gv des psat,Tcl Ksm,out /Ga

106.734 103

B17

B16 pv,ca

5.0234 103

B18

B16 RN2 Tst /Vca

1.1183 106

B19

B16 RO2 Tst /Vca

978.85 103

B20

Gv atm psat,Tatm /Ga /patm

9.6827 103

B21

atm psat,Tatm /patm

15.496 103
)1

B22

Gv ca,in psat,Tcl (O2 ,amb GO2 + (1 O2 ,amb )GN2

B23

des psat,Tcl ca,in psat,Tcl

B24

RO2 Tst GO2 kca,out /Vca

639.069 103

B25

B24 pv,ca

30.077 103

B26

B24 RN2 Tst /Vca

6.6956 106

B27

B24 RO2 Tst /Vca

5.8609 106

B28

RN2 Tst /Vca

10.4770 106

B29

RO2 Tst /Vca

9.1709 106

B30

B24 /kca,out

293.470 103

B31

RO2 Tst GN2 /Vca

256.786 103

B32

1/4GO2 n/F

31.590 106

B33

1 O2 ,amb GO2 /(O2 ,amb GO2 + (1 O2 ,amb )GN2 )

766.990 103

B34

kca,out

2.177 106

B35

kca,out pv,ca

102.489 103

B36

kca,out RN2 Tst /Vca

22.815

B37

kca,out RO2 Tst /Vca

19.970

B38

Ra Tst /Vrm

20.255 106

B39

5.18 1021

B42

B38 pa1 /std5a


B38 pa2 /std4a
B38 pa3 /std3a
B38 pa4 /std2a

B43

B38 pa5 /stda

6.482

B44

B38 pa6

1.5807 106

B45

meana

250 103

B40
B41

29.4084 103

708.05 1018
47.513 1012
5.729 106

A.3 Controller Parameters

163

Table A.2 (continued)


Parameter

Expression

Value

B46

B34 B38

44.108

B47

B35 B38

2.075 106

B48

B36 B38

462.123 106

B49

B37 B38

404.514 106

B50

B33 B16

81.864 103

B51

B50 pv,ca

3.852 103

B52

B50 B28

857.694 103

B53

B50 B29

750.771 103

B54

B33 B12

2.783 106

B55

B33 B13

131.013 103

B56

B33 B14

29.164

B57

B33 B15

25.529

B58

Gv pv,ca

848.103

B59

XO2 ,ca,in B16

24.870 103

B60

XO2 ,ca,in B17

1.1705 103

B61

XO2 ,ca,in B18

260.565 103

B62

XO2 ,ca,in B19

228.082 103

B63

XO2 ,ca,in B12

845.68 109

B64

XO2 ,ca,in B13

39.801 103

B65

XO2 ,ca,in B14

8.8602

B66

XO2 ,ca,in B15

7.7557

B67

B30 B31

36.683 103

B68

B29 GO2 B31

36.683 103

A.3 Controller Parameters


Table A.3 Super-Twisting,
Twisting and Sub-Optimal

Parameter

Value

Super-Twisting algorithm parameter ()

Super-Twisting algorithm parameter ()

Super-Twisting algorithm parameter ()

0.5

Twisting algorithm parameter (r1 )

2.25

Twisting algorithm parameter (r2 )

0.75

Sub-Optimal algorithm parameter (U )

Sub-Optimal algorithm parameter ()

0.3

Sub-Optimal algorithm parameter

( )

164
Table A.4 LQR controller

A Electric Vehicle PEM Fuel Cell Stack Parameters


Parameter

Value

LQR controller parameter (k1 )

28.59

LQR controller parameter (k2 )

1.6 1013

LQR controller parameter (k3 )

60.57

LQR controller parameter (k4 )

7.57

LQR controller parameter (k5 )

579.74

LQR controller parameter (k6 )

2.55

LQR controller parameter (k7 )

3.6 1014

LQR controller parameter (k7 )

189.97

LQR controller parameter (kI )

0.18

References
1. Kunusch C (2006) Second order sliding mode control of a fuel cell stack using a twisting algorithm. MSc thesis, University of La Plata, Argentina (in Spanish)
2. Pukrushpan JT, Stefanopoulou AG, Peng H (2004) Control of fuel cell power systems. Springer,
Berlin
3. Pukrushpan JT, Stefanopoulou AG, Peng H (2004) Control of fuel cell breathing. IEEE Control
Syst Mag 24:3046

Appendix B

Laboratory FC Generation System Parameters

Table B.1 General physics


constants

Table B.2 Compressors


parameters

Parameter

Value

Dry air molar mass (Ga )

0.029 kg/mol

Oxygen molar mass (GO2 )

32 103 kg/mol

Nitrogen molar mass (GN2 )

28 103 kg/mol

Vapour molar mass (Gv )

0.01802 kg/mol

Hydrogen molar mass (GH2 )

2.01 103 kg/mol

Air specic constant (Ra )

286.9 N m/kg/K

Oxygen specic constant (RO2 )

259.8 N m/kg/K

Nitrogen specic constant (RN2 )

296.8 N m/kg/K

Vapour specic constant (Rv )

461.5 N m/kg/K

Hydrogen specic constant (RH2 )

4.124 103 N m/kg/K

Faradays constant (F )

96485 C/mol

Parameter

Value

Electric inductance (L)

2.12 mH

Electric resistance (R)

2.03 

Torque constant (k )

0.0031 N m/A

Motor inertia (J )

0.2 106 N m

Compressors equivalent inertia (Jeq )

1.2 106 N m

Load torque coefcient (A0 )

4.10 104 N m

Load torque coefcient (A1 )

3.92 106 N m s

C. Kunusch et al., Sliding-Mode Control of PEM Fuel Cells,


Advances in Industrial Control,
DOI 10.1007/978-1-4471-2431-3, Springer-Verlag London Limited 2012

165

166

B Laboratory FC Generation System Parameters

Table B.3 Polynomial tting coefcients


Parameter

Value

Parameter

Value

A00

B00

4.83 105 kg/s

A10

0.0058 N m s

B10

5.42 105 kg/s2

A20

0.0013

B20

8.79 106 kg/s3

A01

3.25 106

B01

3.49 107 kg/s2 /bar

A11

2.80 106

B11

3.55 1013 kg/s

A02

1.37 109 N m s/bar2

B02

4.11 1010 kg/s/bar

N m s2
N m/bar
N m s/bar

Table B.4 Humidier and FC stack parameters


Parameter

Value

Air manifold volume (Vsm )


Air manifold constraint (Ksm,out )
Humidier volume (Vhum )
Humidier constraint coefcient (C0 )
Humidier constraint coefcient (C1 )
Humidier constraint coefcient (C2 )
Number of cells (n)
Cathode constraint (Kca,out )
Cathode volume (Vca )
Membrane effective area (Afc )
Dry membrane thickness (tm )
Dry membrane density (m,dry )
Dry membrane molecular weight (Gm,dry )
Membrane diffusion coefcient (Dw )
Water content coefcient (a0 )
Water content coefcient (a1 )
Water content coefcient (a2 )
Water content coefcient (a3 )
Electro-osmotic coefcient (n0 )
Electro-osmotic coefcient (n1 )
Electro-osmotic coefcient (n2 )
Charge transfer coefcient ()
Exchange current density (i0 )
Apparent fuel cell area (Aapp )
Resistance coefcient (R0 )
Resistance coefcient (R1 )
Polarisation curve coefcient (a)
Polarisation curve coefcient (b)
Polarisation curve coefcient (m)
Polarisation curve coefcient (nc )

8 106 m3
0.0486 kg/s/bar
2 104 m3
1.048 107 kg/s
2.109 104 kg/s/bar2
1.562 105 kg/s/bar
7
0.0094 kg/s/bar
4 104 m3
50 cm2
0.0127 cm
0.002 kg/cm3
1.1 kg/mol
5.43 106 cm2 /s
0.043 [H2 O/SO3 ]
17.81 [H2 O/SO3 ]
39.85 [H2 O/SO3 ]
36.0 [H2 O/SO3 ]
3.4 1019 [H2 O/H+ ]
0.05 [H2 O/H+ ]
0.0029 [H2 O/H+ ]
0.7
4.5 106 A/cm2
20.74 cm2
0.22 /cm2
0.005 /cm2
0.2 bar
0.06 V
8.1 105 V
15.2 cm2 /A

B Laboratory FC Generation System Parameters


Table B.5 Operating
conditions

167

Parameter

Value

Humidier temperature (Thum )

55 C

Heating line temperature (Tlh )

60 C

Fuel cell stack temperature (Tst )

60 C

Humidier relative humidity (RH hum )

0.95

Ambient relative humidity (RH amb )

0.5

Ambient pressure (Pamb )

1 bar

Ambient temperature (Tamb )

25 C

Ambient oxygen molar fraction (O2 ,amb )

0.21

Hydrogen input ow (WH2 ,an )

2 slpm

Appendix C

Laboratory FCGS State Space Functions


and Coefcients

C.1 Expression of (x, u, t)



(x, u, t) = 2B02 m1 u(t) m2 x1 x1 m3 + A0 + A00 + A10 (x2 m5 + m6 )

+ A20 (x2 m5 + m6 )2 + A01 x1 + A11 (x2 m5 + m6 )x1 + A02 x12 m4



+ B01 + B11 (x2 m5 + m6 ) + 2B02 x1 m1 m2 m3 + A01


+ A11 (x2 m5 + m6 ) + 2A02 x1 m4 + B11 m5 B00 + B10 (x2 m5 + m6 )
+ B20 (x2 m5 + m6 )2 + B01 x1 + B11 (x2 m5 + m6 )x1 + B02 x12

3


x2 m5 (x3 RO2 + x4 RN2 + x5 Rv )m8 C3 x2 m5 (x3 RO2


+ x4 RN2 + x5 Rv )m8 C2 x2 m5 (x3 RO2 + x4 RN2 + x5 Rv )m8 C1


C0 + B10 m5 + 2B20 (x2 m5 + m6 )m5 + B11 m5 x1 B01

+ B11 (x2 m5 + m6 ) + 2B02 x1 m1 u(t) m2 x1 x1 m3 + A0


+ A00 + A10 (x2 m5 + m6 ) + A20 (x2 m5 + m6 )2 + A01 x1


+ A11 (x2 m5 + m6 )x1 + A02 x12 m4 + B11 m5 m1 u(t) m2 x1
x1 m3 + A0 + A00 + A10 (x2 m5 + m6 ) + A20 (x2 m5 + m6 )2

+ A01 x1 + A11 (x2 m5 + m6 )x1 + A02 x12 m4



+ B01 + B11 (x2 m5 + m6 ) + 2B02 x1 A10 m5 + 2A20 (x2 m5 + m6 )m5


+ A11 m5 x1 m4 + 2B20 m25 B00 + B10 (x2 m5 + m6 )
+ B20 (x2 m5 + m6 )2 + B01 x1 + B11 (x2 m5 + m6 )x1 + B02 x12

3

x2 m5 (x3 RO2 + x4 RN2 + x5 Rv )m8 C3

2

x2 m5 (x3 RO2 + x4 RN2 + x5 Rv )m8 C2


x2 m5 (x3 RO2 + x4 RN2 + x5 Rv )m8 C1 C0



+ B10 m5 + 2B20 (x2 m5 + m6 )m5 + B11 m5 x1 B10 m5
C. Kunusch et al., Sliding-Mode Control of PEM Fuel Cells,
Advances in Industrial Control,
DOI 10.1007/978-1-4471-2431-3, Springer-Verlag London Limited 2012

169

170

C Laboratory FCGS State Space Functions and Coefcients

+ 2B20 (x2 m5 + m6 )m5 + B11 m5 x1

2

3 x2 m5 (x3 RO2 + x4 RN2 + x5 Rv )m8 C3 m5


2 x2 m5 (x3 RO2 + x4 RN2 + x5 Rv )m8 C2 m5 m5 C1

B00 + B10 (x2 m5 + m6 ) + B20 (x2 m5 + m6 )2 + B01 x1
+ B11 (x2 m5 + m6 )x1 + B02 x12

3

x2 m5 (x3 RO2 + x4 RN2 + x5 Rv )m8 C3

2

x2 m5 (x3 RO2 + x4 RN2 + x5 Rv )m8 C2


x2 m5 (x3 RO2 + x4 RN2 + x5 Rv )m8 C1 C0


+ B10 m5 + 2B20 (x2 m5 + m6 )m5 + B11 m5 x1

2
3 x2 m5 (x3 RO2 + x4 RN2 + x5 Rv )m8 C3 RO2 m8


+ 2 x2 m5 (x3 RO2 + x4 RN2 + x5 Rv )m8 C2 RO2 m8 + RO2 m8 C1


m9 x2 m5 (x3 RO2 + x4 RN2 + x5 Rv )m8 C3

2

+ x2 m5 (x3 RO2 + x4 RN2 + x5 Rv )m8 C2


+ x2 m5 (x3 RO2 + x4 RN2 + x5 Rv )m8 C1 + C0 G1
a

1
m11
(x2 m5 m10 )1 1 +
x2 m5 m10

3

+ x2 m5 (x3 RO2 + x4 RN2 + x5 Rv )m8 C3

2

+ x2 m5 (x3 RO2 + x4 RN2 + x5 Rv )m8 C2


+ x2 m5 (x3 RO2 + x4 RN2 + x5 Rv )m8 C1 + C0
1 

m11
XO2 ,ca,in
1+
x2 m5 m10
1

m14
1+
(x3 RO2 + x4 RN2 + x5 Rv )m8 m12

Kca (x3 RO2 + x4 RN2 + x5 Rv )m8 Pamb x3 RO2 GO2




1


x3 RO2 GO2
x 3 RO 2
GN2
1 + Gv x5 Rv
+ 1
x 3 RO 2 + x 4 RN 2
x 3 RO 2 + x 4 RN 2
1

x3 RO2 GO2
1
1
(x3 RO2 + x4 RN2 )
(x3 RO2 + x4 RN2 )
x 3 R O 2 + x 4 RN 2

1


x 3 RO 2
GO2 nIst
GN2
+ 1
1/4
x 3 R O 2 + x 4 RN 2
F

+ B10 m5 + 2B20 (x2 m5 + m6 )m5 + B11 m5 x1


2
3 x2 m5 (x3 RO2 + x4 RN2 + x5 Rv )m8 C3 RN2 m8

C.1 Expression of (x, u, t)

171

+ 2 x2 m5 (x3 RO2 + x4 RN2 + x5 Rv )m8 C2 RN2 m8




3
m9 x2 m5 (x3 RO2 + x4 RN2 + x5 Rv )m8 C3
+ RN2 m8 C1

2

+ x2 m5 (x3 RO2 + x4 RN2 + x5 Rv )m8 C2


+ x2 m5 (x3 RO2 + x4 RN2 + x5 Rv )m8 C1 + C0

1
m11
1
1
+
(x
m

m
)
G1
2 5
10
a
x2 m5 m10

3

+ x2 m5 (x3 RO2 + x4 RN2 + x5 Rv )m8 C3

2

+ x2 m5 (x3 RO2 + x4 RN2 + x5 Rv )m8 C2


+ x2 m5 (x3 RO2 + x4 RN2 + x5 Rv )m8 C1 + C0
1 

m11
1+
x2 m5 m10

1
Gv m12
1+
Gac ai n((x3 RO2 + x4 RN2 + x5 Rv )m8 m12 )

(1 XO2 ,ca,in ) 1 x3 RO2 m8 GO2 (x3 RO2 m8 + x4 RN2 m8 )1

1 

x3 RO2 m8 GO2
x3 RO2 m8
GN2
+ 1
x3 RO2 m8 + x4 RN2 m8
x3 RO2 m8 + x4 RN2 m8

Kca (x3 RO2 + x4 RN2 + x5 Rv )m8 Pamb




x3 RO2 m8 GO2
1 + Gv x5 Rv m8 (x3 RO2 m8 + x4 RN2 m8 )1
x3 RO2 m8 + x4 RN2 m8


1 1 
x3 RO2 m8
+ 1
GN2
x3 RO2 m8 + x4 RN2 m8

+ B10 m5 + 2B20 (x2 m5 + m6 )m5 + B11 m5 x1


2
3 x2 m5 (x3 RO2 + x4 RN2 + x5 Rv )m8 C3 Rv m8

+ 2 x2 m5 (x3 RO2 + x4 RN2 + x5 Rv )m8 C2 Rv m8 + Rv m8 C1




3
Gv m12 x2 m5 (x3 RO2 + x4 RN2 + x5 Rv )m8 C3


2

+ x2 m5 (x3 RO2 + x4 RN2 + x5 Rv )m8 C2


+ x2 m5 (x3 RO2 + x4 RN2 + x5 Rv )m8 C1 + C0 G1
a

1
Gv m10
(x2 m5 m10 )1 1 +
Ga (x2 m5 m10 )


+ (x2 m5 b2 )3 C3 + (x2 m5 b2 )2 C2 + (x2 m5 b2 )C1 + C0
1 


Gv m10
Gv m12 (x2 m5 b2 )3 C3
1+
Ga (x2 m5 m10 )

172

C Laboratory FCGS State Space Functions and Coefcients

+ (x2 m5 b2 )2 C2 + (x2 m5 b2 )C1 + C0 G1


a

1
G
m
v 10
(x2 m5 m10 )1 1 +
Ga (x2 m5 m10 )


+ (x2 m5 b2 )3 C3 + (x2 m5 b2 )2 C2 + (x2 m5 b2 )C1 + C0
1 

Gv m10
1+
Ga (x2 m5 m10 )

1
Gv m12
1+
Ga((x3 RO2 + x4 RN2 + x5 Rv )m8 m12 )

Kca (x3 RO2 + x4 RN2 + x5 Rv )m8 Pamb


+ Kca (x3 RO2 + x4 RN2 + x5 Rv )m8 Pamb




x3 RO2 m8 GO2
1 + Gv x5 Rv m8 (x3 RO2 m8 + x4 RN2 m8 )1
x3 RO2 m8 + x4 RN2 m8

1 1

x3 RO2 m8
Gv nIst
+ 1
+ 1/2
GN2
x3 RO2 m8 + x4 RN2 m8
F



1
a1
n0 + n1 a0 + (x7 m15 + x5 m16 )
+
Af c F
2

a2
a
3
+ (x7 m15 + x5 m16 )2 + (x7 m15 + x5 m16 )3
4
8

a1
a2
+ n2 a0 + (x7 m15 + x5 m16 ) + (x7 m15 + x5 m16 )2
2
4
2 
a3
Ist
+ (x7 m15 + x5 m16 )3
8

(a0 + a1 x5 m16 + a2 x52 m216 + a3 x53 m316 )m,dry

Gm,dry



(a0 + a1 x7 m15 + a2 x72 m215 + a3 x73 m315 )m,dry Dw

Gv Af c n
Gm,dry
tm

C.2 Model Coefcients


m1 = K /R,

m2 = K 30/,

m5 = Tsm Ra /Vhum ,
m8 = Tst /Vca ,

m4 = /30/Jeq

m6 = Psat (Tsm )RH amb + RH hum,ca Psat (Thum,ca )


m9 = Gv RH hum,ca Psat (Thum,ca ),

m11 = Gv Psat (Tsm )RH amb /Ga ,


m13 = RO2 Tst GO2 ,

m3 = A1 30/,

m10 = Psat (Tsm )RH amb

m12 = RH hum,ca Psat (Thum,ca )

m14 = Gv RH hum,ca Psat (Thum,ca )/Ga

m15 = Tst Rv /Van /Psat (Tlh,an ),

m16 = Rv Tst /Vca /Psat (Tlh,ca )

C.3 Feedforward Control Action

m17 = RH an,in Psat (Tlh,an ),

173

m18 = Tst /Van ,

m19 = Tst Rv /Van

m20 = Tst RH2 /Van

C.3 Feedforward Control Action


6
5
4
3
uff = 0.1014Wair,ref
1.1412 Wair,ref
+ 4.8303Wair,ref
9.3370 Wair,ref
2
+ 8.1430Wair,ref
0.6129Wair,ref 0.1974

Index

A
Air
compressor, 75, 106
humidier, 77, 113
supply, 6, 24
supply manifold, 76, 112
Algorithm
sub-optimal, 56, 67, 94, 139, 140
super twisting, 56, 63, 86, 88, 139, 144
twisting, 56, 93, 139, 140
Alkaline fuel cell, 4
Anode
channels, 78, 120
reaction, 3
Armature
current, 107
voltage, 107
B
Back-diffusion, 121
Bipolar plate, 22
C
Catalyst, 13, 20
Catalytic
oxidation, 3, 13
reduction, 3, 13
Cathode
channels, 77, 117
reaction, 3
starvation, 30
Charge transfer coefcient, 16
Chattering, 36, 47
Compressor, 24
Control afne system, 37
Current collector, 16

D
DC motor, 75
DC/AC converter, 25
DC/DC converter, 25
Degradation, 30
Diaphragm vacuum pump, 106, 111
Diffeomorphism, 37
Differential inclusion, 49, 56, 86, 139
Direct methanol fuel cell, 5
Discontinuous control action, 39, 47
E
Efciency, 17
Electro-osmotic drag, 19, 79, 121
Electrochemical potential, 123
Electrode, 20
Electron, 14
Energy conversion, 17
Enthalpy, 17
Entropy change, 15
Equivalent control, 40, 42
Exchange current density, 16
F
Faradays constant, 14, 79
Feedforward control, 86, 87, 141
Filippov
differential inclusion, 50
method, 45
sense, 51
solution, 51
Finite time convergence, 64
Fuel cell, 3
active area, 79
apparent area, 124
stack, 5, 117

C. Kunusch et al., Sliding-Mode Control of PEM Fuel Cells,


Advances in Industrial Control,
DOI 10.1007/978-1-4471-2431-3, Springer-Verlag London Limited 2012

175

176
G
Gas diffusion layer, 21
Gibbs free energy, 14
H
Heat management, 24
Higher order sliding mode, 48
Humidication, 26
Humidier, 26
Humidity ratio, 118
Hydrogen, 1
Hydrogen supply, 24
I
Ideal sliding, 40
Invariance conditions, 41
K
Kalman observer, 98
L
Lie derivative, 37, 41, 137
Line heater, 27
Load torque, 75, 107
Losses
activation, 15, 123
concentration, see diffusion
diffusion, 16, 123
ohmic, 16, 123
LQR, 30, 97, 100
Lyapunov function, 40

Index
excess ratio, see stoichiometry
starvation, 73, 83, 84, 135
stoichiometry, 73, 83, 90, 135, 145
supply, 24
P
Parasitic load, 24
Peak detector, 68
PEM fuel cell, 5, 10, 13, 17, 30
Phosphoric fuel cell, 4
PID, 30
Platinum, 13, 20
Polarisation curve, 17, 123, 126
Polymer electrolyte membrane fuel cell, see
PEM fuel cell
Polymeric membrane, 19, 79
Potential difference, 14
Power conversion, 17
Proton, 14
Proton exchange membrane fuel cell, see PEM
fuel cell
Protonic conductivity, 14
R
Regularity condition, 51
Relative degree, 37, 52, 54, 55
Return manifold, 80
Reversible voltage, 14

M
Majorant curve, 62, 65
Mass mole fraction, 118
Matching condition, 44
MEA, 21, 26, 125
Membrane
conductivity, 16
dry thickness, 16
electrode assembly, see MEA
water content, 16, 125
Modulation factor, 68
Molten carbonate fuel cell, 4
Motor torque, 75, 107

S
Scalar eld, 37
Sealing gasket, 22
Second order sliding mode, 54
Shaft angular speed, 107
Sliding
manifold, see surface
mode control, 35
surface, 37, 39, 82, 136
variable, 137
Solid oxide fuel cell, 5
Stack
current, 89
voltage, 90
State space model, 130

N
Nernst voltage, 15
Net power, 82, 90, 134
Normal form, 52
Nozzle, 112

T
Tafel equation, 16
Thermal management, 6
Torque disturbance, 88
Transversality condition, 42

O
Oxygen, 3
control, 30

V
Vector eld, 37
Voltage drop, 3

Index
W
Water
diffusion, 19

177
management, 6, 24
transport, 121
Windup, 147