Вы находитесь на странице: 1из 6

2nd IFAC Workshop on

2nd IFAC Workshop on


Thermodynamic
2nd IFAC
2nd Foundations
IFAC Workshop
Workshop on
on of Mathematical Systems Theory
Thermodynamic Foundations of Mathematical Systems Theory
September 28-30,Foundations
Thermodynamic
Thermodynamic 2016. Vigo, Spain
of AvailableSystems
of Mathematical online atTheory
www.sciencedirect.com
September 28-30,Foundations Mathematical
2016. Vigo, Spain Systems Theory
September 28-30, 2016. Vigo, Spain
September 28-30, 2016. Vigo, Spain

ScienceDirect
IFAC-PapersOnLine 49-24 (2016) 087092
Distributed
Distributed Nonlinear Control of a
Distributed Nonlinear
Nonlinear Control
Control of
of a
a
Plug-flow
Plug-flow Reactor
Reactor Under
Under Saturation
Saturation
Plug-flow Reactor Under Saturation

J.L.
Pitarch
J.L.
Pitarch M. Rakhshan M.M. Mardani
M. Rakhshan M.M. Mardani

J.L.
Pitarch
J.L. M.S.
Pitarch M.
M. Rakhshan
Sadeghi C. de
Rakhshan M.M.
PradaMardani
M.S.
M.S. Sadeghi
Sadeghi C. deM.M.
Prada
C. de Prada
Mardani


M.S. Sadeghi C. de Prada

Department of Systems Engineering and Control, EII, Universidad de
Department
Department of
of Systems Engineering and Control, EII, Universidad de
Valladolid.
Department
Valladolid. of Systems
C/Real de
Systems
C/Real de
Engineering
Burgos S/N,
Engineering
Burgos S/N,
and
and Control,
47011,
47011, Control, EII,
Valladolid,
Valladolid,
Universidad
EII,SPAIN
SPAIN
de
(e-mail:
Universidad de
(e-mail:
Valladolid. C/Real
Valladolid. C/Real de Burgos
Burgos S/N,
{jose.pitarch,
de S/N, 47011, Valladolid,
Valladolid, SPAIN
prada}@autom.uva.es).
47011, SPAIN (e-mail:
(e-mail:
{jose.pitarch,
{jose.pitarch, prada}@autom.uva.es).
prada}@autom.uva.es).
Department of Electrical
{jose.pitarch,Engineering, University
prada}@autom.uva.es). of Notre Dame,
Department of
Department of Electrical
Electrical Engineering,
Engineering, University
University of
of Notre
Notre Dame,
Dame,
Notre
Notre Dame,
Department of
Dame, IN
IN 46556,
Electrical
46556, USA
USA (e-mail:
Engineering, mrakhsha@nd.edu)
University
(e-mail: of Notre
mrakhsha@nd.edu) Dame,

Notre Dame, IN 46556, USA (e-mail: mrakhsha@nd.edu)
Department
Notre Dame,
Department of
IN
of Control
46556,
Control Engineering,
USA (e-mail:
Engineering, Shiraz University
mrakhsha@nd.edu)
Shiraz University of

Department of Shiraz,
Technology,
Department of ControlIRAN
Control Engineering,
(e-mail:
Engineering, Shiraz
Shiraz University of
{m.mardani,
University of
of
Technology,
Technology, Shiraz,
Shiraz, IRAN
IRAN (e-mail:
(e-mail: {m.mardani,
{m.mardani,
Technology,shasadeghi}@sutech.ac.ir)
Shiraz, IRAN (e-mail:
shasadeghi}@sutech.ac.ir) {m.mardani,
shasadeghi}@sutech.ac.ir)
shasadeghi}@sutech.ac.ir)
Abstract: This
Abstract: This paper
paper dealsdeals with
with the the saturated
saturated control
control problem
problem of of aa class
class of of distributed
distributed systems
systems
Abstract:
which
Abstract: can This
be
This paper
modelled
paper deals
by
deals with
with the
first-order
the saturated
hyperbolic
saturated control problem
partial
control problem of aa class
differential
of class of distributed
equations
of (PDE).
distributed systems
systemsThe
which
which can
canis be be modelled
modelled by
by first-order
first-order hyperbolic
hyperbolic partial
partial differential
differential equations
equations (PDE).
(PDE). The The
objective
which
objective can is designing
be modelled
designing a
a distributed-parameter
by first-order
distributed-parameter state
hyperbolic
state feedback
partial
feedback with guaranteed
differential
with equations
guaranteed performance
(PDE).
performance for
The
for
objective
this class
objective is
of
is designing
systems,
designing a
a distributed-parameter
using the Lyapunov
distributed-parameter state
stability
state feedback
theory
feedback and with guaranteed
polynomial
with guaranteed performance
sum-of-squares
performance (SOS) for
for
this
this class
class of systems, using the Lyapunov stability theory and polynomial sum-of-squares (SOS)
this class of
programming.
programming. of systems,
For this,
systems,
For
using
this,
using
a
the
the Lyapunov
a polynomial
polynomial
Lyapunov stability
parameter theory
theory and
varying
stabilityvarying
parameter (PPV)
and
(PPV)
polynomial
model is
polynomial
model issum-of-squares
employed to
sum-of-squares
employed
(SOS)
to exactly
exactly
(SOS)
programming.
represent
programming. the For
nonlinearthis,
For this, PDE a
PDE polynomial
system
a polynomial inparameter
a local
parameter varying
region of
varying the (PPV)
state
(PPV) model
space
model is
and employed
then,
is employed basedto exactly
on
to exactly it, a
represent
represent the nonlinear
the nonlinear
nonlinear PDE system in a
system guaranteeing local
in aa local region
local region
region of the
of the state
the state
state space
space and and then,
and actuator based
then, based
based on
on it,it, aaa
PPV
PPV state-feedback
represent the
state-feedback law
law is
PDE designed
system
is designed in
guaranteeing exponential
of
exponential stability
space
stability and
and then, saturation
actuator saturation on it,
PPV
in such
PPVsuch state-feedback
region. The
state-feedback law
law is
The approach
approach is designed guaranteeing
is illustrated
illustrated
designed exponential
here through
guaranteeing through the standard
exponential stability
standard example
stability and
and actuator saturation
of aa nonisothermal
actuator nonisothermal
saturation
in
in such region.
region. The approach is
is illustrated here
here through the
the standard example
example of
of a nonisothermal
plug-flow
in such
plug-flow reactor.
region. The
reactor. approach is illustrated here through the standard example of a nonisothermal
plug-flow
plug-flow reactor.
reactor.
2016, IFAC (International Federation of Automatic Control) Hosting by Elsevier Ltd. All rights reserved.
Keywords:
Keywords: Hyperbolic PDE,
Hyperbolic PDE, Polynomial
Polynomial parameterparameter varying,
varying, SOS, SOS, Saturation,
Saturation, Plug-flowPlug-flow reactor
reactor
Keywords: Hyperbolic
Keywords: Hyperbolic PDE, PDE, Polynomial
Polynomial parameterparameter varying,
varying, SOS, SOS, Saturation,
Saturation, Plug-flowPlug-flow reactor
reactor
1. INTRODUCTION
1. troller (Ray, 1981), with its inherent difficulties, and it is
1. INTRODUCTION
1. INTRODUCTION
INTRODUCTION
troller
then
troller
(Ray,
(Ray, 1981),
trollerlumped
(Ray, 1981),
for real
1981),
with
with its
its inherent
inherent difficulties,
implementation.
with its inherent difficulties,
Direct
difficulties, methods
and
and it
and it
is
itap-
is
is
then
then lumped
lumped for
for real
real implementation.
implementation. Direct
Direct methods
methods ap-
ap-
Numerous processes in industry including fluid heat ex- ply
then
ply spatial
lumped
spatial discretization
for real
discretization methods
implementation.
methods (e.g., finite
Direct
(e.g., finite differences,
methods
differences,ap-
Numerous
Numerous processes
processes in
in industry
industry including
including fluid
fluidare heat
heat ex- ply
ex- finite spatial discretization methods (e.g., finite differences,
changers, fiber
Numerous fiber spin lines
processes lines
in or fixed-bed
fixed-bed
industry reactors,
including fluid heatessen-
ex- ply
finite volume,
spatial
volume, orthogonal
discretization
orthogonal collocation
methods
collocation or
(e.g.,
or Galerkins
finite
Galerkins meth-
differences,
meth-
changers,
changers, fiber spin
spin lines or
or fixed-bed reactors,
reactors, are
are essen-
essen- finite volume, orthogonal collocation or Galerkins meth-
tially distributed
changers, fiber spinin space,
lines or i.e. their
fixed-bed behavior
reactors, is determined ods)
finite to the
volume, PDE system
orthogonal in order to
collocation
are essen- ods) to the PDE system in order to obtain an approximate obtain
or an approximate
Galerkins meth-
tially
tially distributed
distributed in
in space,
space, i.e. their
i.e.also
their behavior
behavior is determined
is determined
determined ods) totothat
the PDEPDE system inoforder
order to obtain
obtain an approximate
approximate
not
tially only by
distributed the intime
space, but i.e. their by the
behavior spatial
is position. model
ods)
model that the contains
containssystema set
aa set in ordinary
of ordinaryto differential
an
differential equations
equations
not only
only byby the time but also by
by the spatial position.
position. model that contains
not
Mathematical
not only
Mathematical by the
the time for
models
time
models
butsuch
but
for
also systems
also
such by the can
the
systems
spatial
spatial
can be obtained
be obtained
position. (ODEs)
model
(ODEs) thatin
in time
contains
time a set
(Dochain
(Dochain set of ordinary
of et
et al.,
ordinary
al.,
differential
1992). The
differential
1992). The
equations
subsequent
equations
subsequent
Mathematical
by applying
Mathematical models
the for
fundamental
models for such
such systems
thermodynamic
systems can
can be
be obtained
principles
obtained (ODEs)
ODE
(ODEs) model in
in time
is
timethen (Dochain
used
(Dochain as et
the
et al.,
basis
al., 1992).
for
1992). the The
design
The subsequent
of finite
subsequent
by applying the fundamental thermodynamic principles ODE
ODE model
model is
is then
then used
used as
as the
the basis
basis for
for the
the design
design of
of finite
finite
by
by applying
(balances
applying of the fundamental
momentum,
the fundamental energy thermodynamic
and material),
thermodynamic principles
resulting
principles dimensional
ODE model controllers.
is then used This
as the approach
basis for benefits
the design from
of the
finite
(balances of momentum, energy and material), resulting dimensional
dimensional controllers.
controllers. This
This approach
approach benefits
benefits from
from the
the
(balances
in a set
(balances of of momentum,
semilinear
of momentum, energy PDEs
hyperbolic
energy and material),
and material),
(Aksikas, resulting
2005).
resulting direct application
dimensional of
controllers. finite-dimensional
This approach control
benefits theory
from and
the
in a set of semilinear hyperbolic PDEs (Aksikas, 2005). direct
direct application
application of
of finite-dimensional
finite-dimensional control
control theory
theory and
and
in
in a
a set
set of
of semilinear
semilinear hyperbolic
hyperbolic PDEs
PDEs (Aksikas,
(Aksikas, 2005).
2005). methodologies
direct application but ofit has the important
finite-dimensional drawback
control theoryof that
and
These systems
systems usually
usually do do not work work in in isolation
isolation in the the methodologies
methodologies but
but it has the important drawback of that
These do not isolationas in
the discretized
the methodologies
the discretized but it
ODE
ODE
has
itsize
hasmay
size
the
the be
may
important
very
important
be very
drawback
significant
drawback
significant in
in
of
of that
order
that
order to
These
process
These
process
systems
and
systems
and
usually
chemical
usually
chemical industry,
do
industry,
notbut
not but
work
workwork
work
intogether
in isolation
together as
in
pieces
in the
pieces the discretized
reach
the discretized
the desired ODE
ODE degreesize may
size may
of be very
be very significant
approximation. significantThis in order to
drawback
in order to
to
process
of
andequipment
of aa larger
processlarger
and chemical industry,
chemical
equipment
industry,
directly but
directly
but work in
involved
work
involved
together
production
together
in production as ob- reach
as pieces
pieces
ob- reach
causes
the
thethe
desired
desired degree
degree
controller
of
of
design
approximation.
approximation.
to become high
This
This drawback
drawback
dimensional
of reach
ob- causes thethe desired degree
controller design of approximation.
to become high This drawback
dimensional
of aa larger
larger
jectives.
jectives.
equipment
Therefore,
equipment
Therefore,
directlyanalysis
stability
directly
stability
involved
involved
analysis and in control
in production
production of such
ob- causes
in the
the controller
structure and design
design to
computationally to become
complex. high
high dimensional
jectives.
systems
jectives. Therefore,
with stability
guaranteed
Therefore, stability analysis and
performance
analysis and
is
and of
control
control
both
control
of
of such
such causes
theoreti-
of such in
in structure
structure
controller
and
and computationally
computationally
become
complex.
complex.
dimensional
systems
systems with guaranteed
with guaranteed
guaranteed performance
performance is of
issystemsboth
of both
bothare theoreti-
theoreti- in structure and computationally complex.
cal and
and practical
systems
cal practical
with importance. As PDE
performance PDE is of inher- The
theoreti- The stabilization
stabilization for
for tubular
tubular reactors
reactors in
in particular
particular has
has
cal
ently
cal and practical importance.
importance.
infinite-dimensional,
and practical importance. the
As
As
existent
As PDE
systems
PDE control
systems
are
are inher-
systemsapproaches
are inher- been
inher- The
The stabilization
done
stabilization typically for
for tubular
using
tubular PIDsreactors
(despite
reactors in
in particular
the
particularfact has
that
has
ently infinite-dimensional, the existent control approaches been
been done
done typically
typically using
using PIDs
PIDs (despite
(despite the
the fact
fact that
that
ently
for
ently infinite-dimensional,
lumped-parameter
infinite-dimensional, the
systems
the existent
(LPS)
existent control
are hard
control approaches
to be
approachesused such
been systems
done are nonlinear),
typically using provided
PIDs suitable
(despite the locations
fact thatof
for lumped-parameter systems (LPS) are such systems are nonlinear), provided suitable locations of
for lumped-parameter
directly
for (Wang
lumped-parameter et al., systems
2011):
systems actual
(LPS) are hard
(LPS)controller
are hard
to
to be
hard implemen-
to be used
be used such
used such systems
sensorssystems
and are nonlinear),
are nonlinear),
actuators to ensureprovided
passivity
provided suitable
(Alonso
suitable locations
and
locations Yd-of
of
directly (Wang et al., 2011): actual controller implemen- sensors
sensors and
and actuators
actuators to
to ensure
ensure passivity
passivity (Alonso
(Alonso and
and Yd-
Yd-
directly
tations
directly (Wangbeet
should
(Wang etdone
al., 2011):
al., 2011):
within actual
a finite
actual controller
number
controller of implemen-
actuators
implemen- stie,
sensors2001).and PID designs
actuators to are
ensurebased on
passivity local linear
(Alonso models,
and Yd-
tations should be done
done within a finite number
number of actuators stie, stie, 2001).
2001). PID
PID designs are based on local linear models,
tations
and sensors
tations
and
should
sensors
should in
be
in be
practice.
done
practice.
within
Thus,aa aafinite
within
Thus,
finite
guaranteed
number
guaranteed
of actuators
actuators obtained
distributed-
of
distributed- stie, 2001).
obtained in
in PID designs
different
designs
different
are
are based
operation
operation regions
based
regions
on
on bylocal
local
by
linear
linear models,
input-output
input-output lin-
models,
lin-
and sensors
parameter
and in practice.
controller
sensorscontroller
in practice. design Thus,
Thus,becomesa guaranteed
a challenging
a guaranteed distributed-
task.
distributed- obtained
earization
obtained in different
(Aguilar
in(Aguilar
differentet etoperation
al.,
operation 2002; regions
regions by
Mikhalevich input-output
et
by input-output al., lin-
2015).
lin-
parameter design becomes a challenging task. earization
earization (Aguilar al.,
et al.,
al., 2002; Mikhalevich
2002; Mikhalevich
Mikhalevich et al.,
et al.,
al., 2015).
2015).
parameter controller
parameter controller design
design becomes
becomes aa challenging
challenging task. task. Though
earization the use
(Aguilarof PIDet controllers
2002; is quite simple,et the main
2015).
Many research works have been proposed for the control Though
Though the
the use
use of PID controllers is quite simple, the main
Many
Many
of PDE
PDE
research
research works
works
systemsworks during
have
have been
been
thebeen
proposed
proposed
last proposed
decade. These
for
for the
These
control
control Though
themethods
drawback
drawback the use of
arises
arises in
of
in
PID
finding
PID
finding
controllers
the
controllers
the right
right
is quite
quite simple,
is tuning
tuning in
simple,
in order
order
the
theto
to
main
pro-
main
pro-
Many
of research
systems during havethe last decade. for the control
methods drawback
vide good
drawback arises
arises in
in finding
robustness/performance
finding the
the right
right tuning
trade-offs.
tuning in
in order
In
order to pro-
addition,
to pro-
of
can
of PDE
be
PDE systemsinto
divided
systems during
two
during the last
well-known
the last decade.
types:
decade. These
indirect
These methods
methodsand vide
vide good
good robustness/performance
robustness/performance trade-offs.
trade-offs. In
In addition,
addition,
can be divided into two well-known types: indirect and even
vide
even if
good
if a
a good trade-off
robustness/performance
good trade-off has
has been
been achieved
trade-offs.
achieved in
In
in practice,
addition,
practice,
can
can be
direct divided
be divided into
(Christofides,
into twotwo well-known
2012).
well-known types:
Indirecttypes:methods indirect
employ and
indirect and
the thereeven if ifis aanogood
good trade-off has been achieved
achieved in practice,
practice,
direct (Christofides, 2012). Indirect methods guarantee of constraints satisfaction in the
direct
original
direct (Christofides,
PDE model
(Christofides, to 2012).
design
2012). Indirect
an methods employ
infinite-dimensional
Indirect methods employcon-
employ
the
the even
the there
there is
is no
no
trade-off
guarantee
guarantee of
of
has been
constraints
constraints satisfaction
satisfaction
in in
in the
the
original PDE model to design an infinite-dimensional con- entire
there
entire operating
is no
operating region.
guarantee
region. ofRecently,
constraints
Recently, some
some works
satisfaction
works have
have ingiven
the
given
original
original
 PDE model
PDE leading to
modeltotothese design
design an infinite-dimensional
an has
infinite-dimensional con-
con- entire operating region. Recently, some works have given
The research results received
 The research leading to these results has received funding from the funding from the alternatives
entire operating
alternatives to
to the PID
region. control
Recently, of PDE
some systems.
works In
have partic-
given

European
 The Unionleading
The research
research (FP7/2007-2013,
to
to these
these resultsgrant agreement
has received 604068)from
received funding and the alternatives
ular, Aksikas
alternatives toetthe
to the
theal.
PID
PID
PID (2007)
control
control
control
of
of PDE
proposed
of PDEansystems.
PDE systems.
LQR
systems.
In
In partic-
In partic-
control de-
partic-
European Unionleading
(FP7/2007-2013, results
grant has
agreement funding
604068)from the
and the ular, Aksikas et al. (2007) proposed an LQR control de-
Spanish
European
European Government
Union
Union (MINECO/FEDER
(FP7/2007-2013,
(FP7/2007-2013, grant
grant
Spanish Government (MINECO/FEDER DPI2015-70975-P).
DPI2015-70975-P).
agreement
agreement 604068)
604068) and
and the
the ular,
ular, Aksikas
Aksikas et
et al.
al. (2007)
(2007) proposed
proposed an
an LQR
LQR control
control de-
de-
Spanish
Spanish Government
Government (MINECO/FEDER
(MINECO/FEDER DPI2015-70975-P).
DPI2015-70975-P).
Copyright
2405-8963 2016 IFAC 102 Hosting by Elsevier Ltd. All rights reserved.
Copyright 2016, IFAC (International Federation of Automatic Control)
2016 IFAC 102
Copyright
Peer review
Copyright 2016
under IFAC
2016 responsibility
IFAC 102
of International Federation of Automatic
102Control.
10.1016/j.ifacol.2016.10.760
2016 IFAC TFMST
88
September 28-30, 2016. Vigo, Spain J.L. Pitarch et al. / IFAC-PapersOnLine 49-24 (2016) 087092

sign for nonisothermal PFRs using spectral factorization, form A bB takes place, being b > 0 the stoichiometric
and Wang et al. (2011) presented a fuzzy Takagi-Sugeno coefficient. Thus, the composition of the reaction mixture
(TS) approach (quasi-LPV) for such systems based on changes along the length x of the reactor, as represented
linear matrix inequalities (LMI) and the well-known sector in Figure 1. The reaction is endothermic and a jacket is
nonlinearity modelling (Tanaka and Wang, 2001). Nev- used to heat the reactor, so that the system is dissipative,
ertheless, there is still a common drawback for practical therefore open-loop stable.
implementation of the above designs: the actuators physi-
cal limitations are sometimes neglected or their treatment
leads to very conservative controllers. However, treating
them appropriately throughout the control design is key.
There are, of course, other control strategies based in op-
timization (e.g., model predictive control) which explicitly
consider input/state constraints, not discussed here.
Recently, TS/LPV modelling techniques have been ex-
tended to employ polynomial vertex models (Tanaka et al.,
2009b) instead of linear ones. This class of polynomial
parameter-varying (PPV) representation allows asymptot- Fig. 1. Nonisothermal plug-flow reactor.
ically reducing conservativeness of the design approach if
the Taylor series decomposition is used for the sector mod- In Figure 1, CA and CB are the reactant and product
elling (Sala and Arino, 2009). Then, the sum-of-squares concentrations respectively, T denotes the reactor temper-
(SOS) programming developed for pure polynomial sys- ature, Tin/out and CA,in/out are defined as the temperature
tems (Papachristodoulou and Prajna, 2005; Pitarch et al., and concentration of the inlet/outlet streams respectively,
2016b) is used to design PPV-based control systems with FB is the partial flow of product B, and L denotes the
guaranteed performance (Tanaka et al., 2009a; Sala, 2009; total length of the reactor. Under assumptions of perfect
Pitarch, 2013). Numerical solutions for these designs can radial mixing, constant density and heat capacity of the
be computed via the Gram-matrix decomposition and reacting liquid, and negligible diffusive phenomena, a dy-
semidefinite programming (SDP) (Seiler, 2013). namic model of the process can be derived from material
The objective of this paper is extending the existent and energy balances. Note that CB is known if CA and T
LPV control designs for a nonisothermal plug-flow reactor are known, so its mass balance has been omitted and only
(PFR) to a PPV approach, including explicit consideration states T and CA will be considered henceforth:
of the actuator limits from the design phase. First, a T T k0 H E 4h
= v CA e RT + (TJ T ) (1)
PPV-PDE model based on the Taylor series is proposed t l p C p p C p d
to accurately represent the nonlinear hyperbolic PDE CA CA E

system. Then, based on this model, a PPV state-feedback = v k0 CA e RT (2)


t l
with antiwindup is proposed. In this way, the problem of Where E, R, k0 , H, h and d are the activation energy, the
finding suitable controller gains fulfilling input-saturation ideal gas, the pre-exponential factor, the enthalpy of the
and guaranteeing local exponential stability of the closed- reaction, the wall heat transfer coefficient and the reactor
loop system is derived in terms of a set of spatially- diameter, respectively. The control input is chosen to be
dependent polynomial constraints, to be checked for SOS. the spatially distributed jacket temperature TJ and t, l
Briefly, the rest of the paper organizes as follows: Section 2 denote the independent time and space variables.
describes the PFR, its thermodynamic model and gives a The process is subject to the boundary conditions
PPV local representation for it; Section 3 states the control
problem and its formulation into a SOS programming T (0, t) = Tin , CA (0, t) = CA,in , CB (0, t) = 0 (3)
problem; Section 4 shows the effectiveness of the proposed and the initial state-conditions profiles are:
approach with some results in simulation and, finally, a T (l, 0) = T0 (l), CA (l, 0) = CA0 (l) (4)
conclusion is drawn in the last section.
Notation: I stands for the identity. D[v] denotes a diag- 2.1 Nonlinear dimensionless PDE model
onal matrix formed by the elements of v. M [k] will denote
the k-th row of the matrix M . A symmetric matrix P (x) A scaled model will be obtained in order to compensate
in the spatial variable x is positive definite (semidefinite) large differences in the magnitude orders of states, avoid-
in an interval l1 x l2 if P (x)  0 (P (x)  0) for all ing thus numerical problems in the SOS design phase.
x [l1 , l2 ]. The symbol () denotes the symmetric element Hence, the following dimensionless states and input are
in matrix expressions, e.g., [M (x) + N (x) + ()] [M (x) + introduced:
N (x) + M T (x) + N T (x)]. A SOS polynomial p(y) in vari- T Tin CA,in CA TJ Tin
1 := , 2 := , J := (5)
ables y is denoted by p(y) y . Similarly, an n m SOS Tin CA,in Tin
polynomial matrix L(y) will be denoted by L(y) nm y . Define also x := l/L. Then an equivalent representation
of (1)-(2) in variables (5) can be obtained (omitted for
2. PLUG-FLOW REACTOR MODELLING brevity 1 ). Note that the dimensionless equilibrium profile
(t ) in one variable can be computed given a prefixed
A nonisothermal PFR is an ideal flow reactor in which 1 The reader is referred to Aksikas (2005, Chap. 5) for the system

no back mixing occurs while a chemical reaction of the description with more details.

103
2016 IFAC TFMST
September 28-30, 2016. Vigo, Spain J.L. Pitarch et al. / IFAC-PapersOnLine 49-24 (2016) 087092 89

one for the other variable. In this way, setting T (x) =


Tin along the reactor in equilibrium for simplicity, the
remaining profiles read:
2 L 1 2 L
1e (x) = 0, 2e (x) = 1 e v x , Je (x) = e v x
b
with 1  2 , 2  k0 e , and

E (H)CA,in 4h
 ,  , b .
RTin p Cp Tin p C p d
Numerical values for the above parameters are in Table 1.
Table 1. Model parameters for simulation.
Parameters Numerical values
v 0.025 m/s
L 1m Fig. 2. Bounding with polynomial vertex models.
E 11250 cal/mol
3. CONTROL SYNTHESIS
k0 106 s1
4h/p Cp d 0.2 s1
CA,in 0.02 mol/L
Definition. The system (7) is exponentially stable with
R 1.986 cal/(mol.K)
decay rate > 0, if there exists > 0 such that:
Tin 340 K y(x, t)22 e2t y(x, 0)22 t 0. (10)
0.25
Consider the following candidate Lyapunov functional
 1
Now, let us consider a new input vector u(x, t) = J (x, t)
Je (x) and the state transformation: V (t) = y(x, t)T P 1 (y2 , x)y(x, t)dx (11)
  0
1 (x, t) 1e (x) where P (y2 , x) = P T (y2 , x) R22
y(x, t) = (6) y2 ,x is a positive-definite
2 (x, t) 2e (x) polynomial matrix.
Finally, the system (1)-(2) can be rewritten as: Lemma 1. (Wang et al. (2011)). If there exists a Lya-
y(x, t) y(x, t) punov functional (11) and a scalar > 0 satisfying
= + f (y, x) + B(x)u(x, t) (7) dV (t)
t x + 2V (t) 0 (12)
v dt
= I, B(x) = [b, 0]T , f (y, x) = [1 f0 (y, x) by1 , the system (7) is exponentially stable with decay rate .
L
y1 y1
T
2 f0 (y, x)] , f0 (y, x) = (12e (x))[e 1+y1 1]y2 e 1+y1 Note that condition (12) is indeed a dissipation inequality
when the supply rate is zero but initial conditions are not.
2.2 PPV modelling So, if the storage function (11) is proven convex (ensured
by (12)), the above Lyapunov requirements are analogous
The only non-polynomial term appearing in (7) is (y1 ) = to the ones in passivity theory (Alonso and Ydstie, 2001).
y1
e 1+y1 . Consider a modelling region = {y1 : 2 y12 0}. SOS Polynomials. Checking a polynomial for SOS, i.e.,
Then, using the Taylor series (Sala and Arino, 2009) up to positive, is being able to express it as the sum of squares
degree 3, we can exactly represent (y1 ) in as of simpler polynomials. An even-degree polynomial p(y)
(y1 ) = 122.13y12 + 16.661y1 + 1+ is SOS iff there exist a vector of monomials m(y) and a
matrix H  0 such that p(y) = m(z)T Hm(y). In this way,
1 (y1 ) sup T3 (y1 ) + 2 (y1 ) inf T3 (y1 ) (8)
if the polynomial is affine in decision variables (typically
being 1 , 2 the nonlinear membership functions and T3 its coefficients), it can be checked for SOS via SDP solvers
the Taylors remainder of order 3, computed as: searching for such an H. See Seiler (2013) for details.
(y1 ) 122.13y12 16.661y1 1 Local positivity of polynomials can be checked via the well-
T3 (y1 ) = known Positivstellensatz theorem (Stengle, 1974), imple-
y13
mentations in (Jarvis-Wloszek et al., 2005). The following
T3 (y1 ) inf T3 (y1 ) lemma is reduced version of such theorem. Consider a
1 (y1 ) = , 2 (y1 ) = 1 1 (y1 )
sup T3 (y1 ) inf T3 (y1 ) region defined by polynomial boundaries as := {y :
For example, Figure 2 shows a comparison between the g1 (y) > 0, ..., gng (y) > 0, h1 (y) = 0, ..., hnh (y) = 0}.
sector bounding of in 0.2 y1 0.2 using linear vertex Lemma 2. (Pitarch (2013)). If polynomials si (y) y
models (LPV/TS) and polynomial ones up to degree 3 and rj (y) Ry can be found fulfilling
ng nh
(Taylor O3).  
p(y) (y) si (y)gi (y) + rj (y)hj (y) y (13)
Assuming y1 is measurable, this technique allows to i=1 j=1
rewrite (7) in PPV form:
2
then p(y) is locally greater or equal than (y) in .
y y 
= + i (y1 )Ai (y, x)y + B(x)u(x, t) (9) Based, also in Positivstellensatz results, we can set SOS
t x i=1
conditions for a semialgebraic set Z = {zk (y) > 0, k =

104
2016 IFAC TFMST
90
September 28-30, 2016. Vigo, Spain J.L. Pitarch et al. / IFAC-PapersOnLine 49-24 (2016) 087092

 
1, . . . , kM } to be included in . For instance, the existence Pq (y2 ) s6qi (y)(2 y12 ) s7qi (y)( 2 y22 ) ()
of multipliers (si , sik ) y , rij Ry such that si gi + Mqi (y) Lqi (y) 2
 nh  kM
j rij hj k=1 sik zk y , i : 1, . . . , ng . 33
y q : 1, . . . , N i : 1, 2 (22)
Pq (y2 (xq , 0)) y(xq , 0) y(xq , 0)T  0 q : 1, . . . , N (23)
3.1 Problem statement  
Pq (y2 ) qi I ()
33 q : 1, . . . , N ; i : 1, 2 (24)
Consider the following modelling region for the plug-flow Mqi (y) 2 /2 y

reactor 2

  Pq (y2 )
:= (y, x) : 2 y12 0, 2 y22 0, x [0, 1] (14) qi = Aqi (y)[k] y
yk
k=1
Based on the PPV model (9), consider the subsequent qi = 2wT rqi (y)BqT + wT rqi (y)w wT Lqi (y)
infinite-dimensional state-feedback law with saturation for
controlling the nonlinear PDE system (7): qi = s8qi (y)(2 y12 ) + s9qi (y)( 2 y22 )
2
 being x = N 1 , then the closed-loop system (17) is
u(x, t) = sat((x, t)), (x, t) = j (y1 )Kj (y, x)y (15) guaranteed to be exponentially stable with decay rate in
j=1 the region (14), with controller (15) fulfilling bounds (16)
and the actuators speed is bounded by   +
where Kj (y, x) = M (y, x)P 1 (y2 , x), M (y, x) R12
y,x are , (, ) > 0. The controller gains can be obtained as
the controller gains to be found. The control law (15) must Kqi (y) = Mqi (y)Pq1 (y2 ).
fulfill a saturation constraint:
Proof. Conditions in this theorem asymptotically ap-
|u(x, t)| , R+ (y, x) proach 2 those in Lemma 1 for large N , i.e., Kqj (y)
sat((x, t)) := sign((x, t)) min (|(x, t)| , ) (16) and Pq1 (y2 ) are good approximations to Kj (y, x) and
P 1 (y2 , x) on the interval [xq x/2, xq + x/2]. In this
Now, defining := sat(), from (15) and (9) the overall way, conditions (18) ensure V > 0 locally in each q by
closed-loop system can be written as: Lemma 2. Conditions (23) ensure the initial conditions
y(xq , 0) belong to the level set {(y, xq ) : y T P 1 (y2 , xq )y
y y 
2  1} by applying Schur complement twice:
= + i (y1 ) Ai (y, x)  
t x i=1 T 1 1 yT
 1 y Pq y 0  0 Pq y y T  0
y Pq
+ B(x)Ki (y, x) y B(x) (17)
Conditions (19) are a particular convex implementation of
The goal of this paper is finding a suitable finite- (12) with a backward difference for the spatial derivative
dimensional approximation to (15), fulfilling (16), which, P/x, details omitted 3 . On the one hand, multipliers
in closed loop with (7), makes the overall system locally s2 , s3 ensure that the spatial discretization of (12) holds
exponentially stable in region (14) with decay rate > 0, in each q . On the other hand, multipliers r together with
starting from initial conditions y(x, 0) . conditions (22) make (12) hold also with controller (15),
as follows. We need to ensure the bound (16) inside each
3.2 Finite-dimensional convex SOS design level set {(y, xq ) : yT Pq1 (y)y 1}. This, by the convex-
sum property of i (y1 ) and Lemma 2, is:
The following SOS approach follows a direct schema. So,  
2 ()T (Kqi (y)y (q )) sq 1 y T Pq1 (y2 )y y
the spatial length x [0, 1] is gridded in N equally-
distributed points in order to apply finite differences 
Now, assume that () = i i Gi (y)y when () =
(Ames, 1992) to (12). Now, define q := (y, xq ) with 0, being G(y) R12 slack decision variables. Then,
q N , xq = Nq , i.e., q is the cutting hyperplane of in y

x = xq . Similarly, Aq (y) will stand for A(y, xq ). choosing sq = 2 (classical S-procedure for quadratic
LMIs) and doing the changes of variables y = Pq (y2 ),
Theorem 3. Given scalars (, , , ) R+ , if there exist Lqi (y) = Gqi (y)Pq (y2 ), the SOS conditions before lead to
polynomial matrices Pq (y), Mqi (y), Lqi (y), plus multipli-  
ers s(y) y and r(y) Ry , fulfilling: T Pq (y2 ) 2 ()T (Mqi (y) Lqi (y))
  which becomes (22) by Schur complement plus multipliers
Pq (y2 ) 1 +s1q (y2 )( 2 y22 ) I 22
y2 q : 0, . . . , N (18)
s6 , s7 to add local information of q . So, what remains
    is checking the dissipativity condition (12) together with

T Pq1 (y2 )Pq (y2 ) Aqi (y)Pq (y2 )+Bq Mqi (y) ensuring () = i i (y1 )Gi (y)y when () = 0. By
 x  Lemma 2, this is ensured adding the term
() + qi 2Pq (y2 ) 2 + s2qi (y)(2 y12 ) + s3qi (y)  
  T 1

( 2 y22 ) I + qi y,,w q : 1, . . . , N ; i : 1, 2 (19) (q ) rq (y) (q ) i (y1 )Gqi (y)y
i
D[2 , 3 ] Pq (y2 ) s4q (y2 )( 2 y22 )I 22
y2 q : 1, . . . , N 2 It is well established (Balas, 1986) that as N increases, the closed-
(20)
loop system that results from the PDE model plus an approximate
D[4 , 2 ] Pq (y2 ) s5q (y2 )( 2 y22 )I 22
y2 q : 1, . . . , N finite-dimensional controller converges to (17).
(21) 3 The extended proof will be available in Pitarch et al. (2016a).

105
2016 IFAC TFMST
September 28-30, 2016. Vigo, Spain J.L. Pitarch et al. / IFAC-PapersOnLine 49-24 (2016) 087092 91

to the spatially-discretized version of (12). By the convex- and decreasing exponentially up to 0.014 mol/L at x = 1
sum property and the change of variable () = w, the (see Figures 3, 4). To ensure stability and performance
above can be expressed as guarantees starting from these initial conditions, we have
   modelled the plug-flow reactor as a PPV model (8) in the
qi (y) ()
[ w] 1 1 y,,w region (14) with = 0.1, = 1.
Bq + rqi (y)Lqi (y)/2 rqi (y) w
(25) For numerical computation we have decided to set N =
being qi (y) the left-hand side of (19) without the 30 distributed points along the reactors length, fourth
terms. Henceforth, (25) becomes (19) by matrix congru- degree for the Lyapunov matrices Pq (y2 ) and third degree
ence with [I rqi (y)]. for polynomial matrices Lqi (y), Mqi (y). The user-defined
scalars for the theorem were set to = 0.008, = 0.1,
Enforcing {(y, xq ) : y T P 1 (y2 , xq )y 1} q is required = 0.02 and 1,2,3,4 = 0.0001. The evolution of the
to keep the validity of the PPV model in the rest of reactor states in closed loop with the computed controller
conditions. In this way, conditions (20)-(21) are proving is depicted in Figures 3 and 4.
that y T P 1 (y2 , xq )y 1 in the boundary of , resorting
to similar argumentations to the ones discussed for (22).
Finally, the application of Lemma 1 with conditions (18)-
(19) ensures y 1 et y0 , 1 > 0. By differentiating
both sides with respect to time, one can obtain the bound
y y 2 y, 2 > 1. Then, thanks to the
Cauchy-Schwarz inequality, the control input derivative
   Kqi (y) [k] 
q = i Kqi (y)y + k yk y y can be bounded by
qi  Kqi (y)y + Kqi (y) y +
2 Kqi (y) y + 2 3 Kqi (y)y + (26)
with > 0, 3 1 and the fact that Kqi (y)/yk are
bounded in q . Hence, ensuring T 2 ()T (Kqi (y)y)
0 when y T Pq1 y 1 allows bounding   + ,
= 2 3 . This set inclusion is checked by (24) after
the usual change of variables, Positivstellensatz and Schur Fig. 3. Evolution of the reactor temperature.
complement, analogous to the development followed to
obtain (22). 
The synthesis methodology using Theorem 3 is fixing a
priori particular structures (usually by setting a maximum
degree) for polynomial decision variables Pq (y), Mqi (y),
Lqi (y), s(y) and r(y), and then check (18)-(24) for SOS.
Remark 4. Notice that the controller gains computed by
Theorem 3 are neither constant nor polynomial, but ra-
tional in the state variables. This fact, together with the
nonlinear nature of the membership functions i (y1 ), gives
great freedom to get complex nonlinear controllers.
Remark 5. Note also that no limit cycles or other attrac-
tors than the considered equilibrium profile can appear in
if Theorem 3 renders feasible, because strict positivity
of Lyapunov inequalities is ensured by the  tolerances.
Fig. 4. Evolution of the reactant concentration.
Despite conditions (24) do not state a rigorous bound like Figure 5 shows the evolution with time of the control
the ones to ensure input saturation, Theorem 3 addresses input, the reactor jacket temperature. The actuator satu-
also an important limitation in practice: the actuators rates at its limits during the first 2-3 seconds, converging
speed, often forgotten in most of theoretical LMI/SOS then to the equilibrium profiles in an smooth way.
designs. In fact, can be understood as a qualitative design
parameter to tune the aggressiveness of the controller. For completeness, the obtained Lyapunov surface for the
point x7 = 0.23 is depicted in Figure 6. The estimated
4. SIMULATION RESULTS domain of attraction, i.e., the level set y T P71 (y2 )y = 1, is
also shown at the right-hand side of the picture, together
The controller computed with the proposed method has with its corresponding initial conditions (red point).
been tested in closed loop with the plug-flow reactors
model. The initial conditions for simulation are
2 xL
chosen 5. CONCLUSION
y1 (x, 0) = 0.07 sin(2xL), y2 (x, 0) = 0.5(e v 1).
This means an initial reactor temperature fluctuating The proposed systematic methodology for distributed-
25 K around the inlet temperature Tin and an initial controller synthesis with input constraints has succeeded
concentration of reactant A starting at Ca,in in x = 0 in finding a nonlinear state feedback which guarantees

106
2016 IFAC TFMST
92
September 28-30, 2016. Vigo, Spain J.L. Pitarch et al. / IFAC-PapersOnLine 49-24 (2016) 087092

computing. Academic Press, Boston, 3rd ed edition.


Balas, M.J. (1986). Finite-dimensional control of dis-
tributed parameter systems by galerkin approximation
of infinite dimensional controllers. Journal of Mathe-
matical Analysis and Applications, 114(1), 17 36.
Christofides, P.D. (2012). Nonlinear and robust control
of PDE systems: Methods and applications to transport-
reaction processes. Springer Science & Business Media.
Dochain, D., Babary, J., and Tali-Maamar, N. (1992).
Modelling and adaptive control of nonlinear distributed
parameter bioreactors via orthogonal collocation. Auto-
matica, 28(5), 873 883.
Jarvis-Wloszek, Z., Feeley, R., Tan, W., Sun, K., and
Packard, A. (2005). Control applications of sum of
squares programming. In D. Henrion and A. Garulli
Fig. 5. Evolution of the controlled jacket temperature. (eds.), Positive Polynomials in Control, 322. Springer.
Mikhalevich, S., Rossi, F., Manenti, F., and Baydali, S.
(2015). Robust PI/PID controller design for the reliable
control of plug flow reactor, volume 43 of Chemical
Engineering Transactions, 15251530. AIDIC.
Papachristodoulou, A. and Prajna, S. (2005). Analysis
of non-polynomial systems using the sum of squares
decomposition. In D. Henrion and A. Garulli (eds.),
Positive Polynomials in Control, 2343. Springer.
Pitarch, J.L. (2013). Contributions to Fuzzy Polynomial
Techniques for Stability Analysis and Control. Ph.D.
thesis, Universitat Politecnica de Valencia (UPV).
Pitarch, J.L., Rakhshan, M., Mardani, M.M., and Sadeghi,
M.S. (2016a). Distributed saturated control for a class
of semilinear PDE systems: A SOS approach. Under
review.
Pitarch, J.L., Sala, A., Lauber, J., and Guerra, T.M.
(2016b). Control synthesis for polynomial discrete-
Fig. 6. Obtained Lyapunov function for x7 = 0.23. time systems under input constraints via delayed-state
lyapunov functions. International Journal of Systems
exponential stability of the PFR. It relies on PPV mod-
Science, 47(5), 11761184.
elling and SOS programming. The obtained controller with
Ray, W. (1981). Advanced Process Control. McGraw-Hill.
antiwindup is able to make the PFR reach the desired state
Sala, A. (2009). On the conservativeness of fuzzy and
profiles with a fast but smooth response.
fuzzy-polynomial control of nonlinear systems. Annual
Further work might consider a more realistic problem: Reviews in Control, 33, 4858.
output feedback design, disturbance analysis (practical Sala, A. and Arino, C.V. (2009). Polynomial fuzzy models
stability), etc. Reducing conservativeness (iterative SOS for nonlinear control: a Taylor-series approach. IEEE
approaches for nonconvex problems) and generalizing to a Transactions on Fuzzy Systems, 17, 284295.
wider class of systems is within the scope. Seiler, P. (2013). SOSOPT: A Toolbox for Polynomial
Optimization. ArXiv e-prints.
REFERENCES Stengle, G. (1974). A Nullstellensatz and a Positivstel-
lensatz in semialgebraic geometry. Mathematische An-
Aguilar, R., Poznyak, A., Martnez-Guerra, R., and Maya- nalen, 207(2), 8797.
Yescas, R. (2002). Temperature control in catalytic Tanaka, K. and Wang, H.O. (2001). Fuzzy Control Sys-
cracking reactors via a robust PID controller. Journal tems Design and Analysis: A Linear Matrix Inequality
of Process Control, 12(6), 695 705. Approach. Wiley.
Aksikas, I., Winkin, J.J., and Dochain, D. (2007). Optimal Tanaka, K., Yoshida, H., Ohtake, H., and Wang, H.O.
LQ-feedback regulation of a nonisothermal plug flow (2009a). Guaranteed cost control of polynomial fuzzy
reactor model by spectral factorization. IEEE Trans. systems via a sum of squares approach. IEEE Trans.
on Automatic Control, 52, 11791193. Syst., Man, Cybern. B, Cybern., 39, 561567.
Aksikas, I. (2005). Analysis and LQ-optimal control of Tanaka, K., Yoshida, H., Ohtake, H., and Wang, H.O.
infinite-dimensional semilinear systems: application to (2009b). A sum of squares approach to modeling and
a plug flow reactor. Ph.D. thesis, Universite Catholique control of nonlinear dynamical systems with polynomial
De Louvain (UCL). fuzzy systems. IEEE Trans. Fuzzy Syst., 17, 911 922.
Alonso, A.A. and Ydstie, B.E. (2001). Stabilization of Wang, J., Wu, H., and Xiong, H. (2011). Distributed fuzzy
distributed systems using irreversible thermodynamics. control design of nonlinear hyperbolic PDE systems
Automatica, 37(11), 1739 1755. with application to nonisothermal plug-flow reactor.
Ames, W.F. (1992). Numerical methods for partial dif- IEEE Trans. on Fuzzy Systems, 19, 514 526.
ferential equations. Computer science and scientific

107