Академический Документы
Профессиональный Документы
Культура Документы
DISTILLATION COLUMN
by
Master of Science
JANUARY 2014
ACKNOWLEDGEMENT
In the Name of Allah, the Most Beneficent and the Most Merciful. All
Praises is due to Almighty Allah, The Cherisher and The Sustainer of the world.
Blessings and utmost greetings of peace upon the prophet Muhammad (peace be
upon him), his families, companions and followers. Praise be to Allah, whom has
bestow me His blessings, guidance, and strength for me to complete this thesis
successfully.
gratitude to my supervisor, Assoc. Prof Dr. Norashid bin Aziz for his kind support,
guidance and motivation that had assist me throughout my study. I really appreciate
his enthusiasm, patient and advices that had help me greatly during my research and
writing this thesis. I also want to offer my special thanks to the lecturers from the
Process Control Research Group (PCRG); Assoc. Prof. Dr. Zainal bin Ahmad (my
co-supervisor), Dr. Suhairi bin Abdul Sata and Assoc. Prof. Dr. Syamsul Rizal bin
Abd Shukor for their insightful feedbacks and comments on my work. In addition, I
would like to show my greatest appreciation to the Dean of the School of Chemical
Engineering, Prof. Dr. Azlina Bt Harun @ Kamaruddin and to all the staffs in the
School of Chemical Engineering for their support, help, and kindness throughout my
study here. I own my sincere gratitude to Universiti Sains Malaysia for providing me
with the financial aid through the fellowship scheme during my research period.
colleagues, and friends from my research group, PCRG and other schools who had
given me the persistence support and constructive comments all this time in order to
complete my work. They are Imam, Kak Asyura, Husaini (mech), Abdul Ghani (ee),
ii
Pak Fadli (ee), Pak Fakhrony, Pak Sudibiyo, Kak Alwani, Noraini, Amirah, Linda,
Norazwan, Fariz, Kak Lin, Kak Abot, Kak Hana, Muaz, Ihsan (ee), Mior (mech),
Khalis (ee), Zambri (ee), Hizami, Helmi (material), Senthil, Dipesh, and all my dear
friends who had being a part of my life during my stay here. Thank you for those
Awang and Azni binti Che Ngah for their love, caring and never ending support in
my pursue in the academic world. Your unconditional love has always been an
inspiration for me to move forward in this world. And to my dear siblings; Anas,
Nabilah, Nasuha and Izzati; thank you for always brighten up my life.
directly or indirectly towards the completion of this research. Without all these
people help, this research would not have become a reality. May Allah bless us all
iii
TABLE OF CONTENTS
ACKNOWLEDGEMENT ii
TABLE OF CONTENTS iv
LIST OF TABLES x
ABSTRAK xxi
ABSTRACT xxiii
CHAPTER 1: INTRODUCTION 1
iv
2.3 Distillation Control 14
CHAPTER 3: METHODOLOGY 48
3.1 Overview 48
3.2.1 Matlab 49
v
3.7.1 Linear Model Identification 58
3.7.3 Tuning 64
vi
CHAPTER 4: RESULTS AND DISCUSSION 79
vii
4.6.3 (a) Setpoint Tracking Test 95
viii
CHAPTER 5: CONCLUSIONS AND RECOMMENDATIONS 126
REFERENCES 130
APPENDICES 144
APPENDIX A 144
APPENDIX B 146
ix
LIST OF TABLES
Page
Table 3.3 Distillation column operation data (Klemola and Ilme, 1996) 53
Table 4.2 Comparison of current and proposed value for the distillation 81
column steady state
Table 4.3 System identification results for Model XB and Model T68 93
Table 4.4 Summary results for Linear CIMIC, 2DOF IMC and IMC 102
performance
Table 4.5 Summary results for LCIMIC-AS and LCIMIC 106
x
Table 4.14 Neural network disturbance compensator Y validation results 113
Table 4.16 The tuning parameter used in the Linear and Nonlinear 116
CIMIC
Table 4.17 Controllers performance summary 125
xi
LIST OF FIGURES
Page
Figure 2.3 General model based control scheme (Brosilow and Joseph, 19
2002)
Figure 2.4 Schematic diagram of a single neuron (Seborg et al., 2004) 46
Figure 3.11 Neural network inverse using direct method (Norgaard et al., 75
2000)
Figure 3.12 NLCIMIC scheme in Matlab Simulink 76
xii
Figure 4.4 Step test results for bottom product purity (left) and 84
temperature Tray 68 (right) by manipulating the reboiler duty
at 5%, 10% and 15% change from the nominal condition
Figure 4.5 Step test results for bottom product purity (left) and Tray 68 85
temperature (right) by manipulating feed flowrate at 3%,
5% and 10% change from the nominal condition
Figure 4.6 Step test results for bottom composition (left) and Tray 68 86
temperature (right) by manipulating the n-butane feed
composition at 3%, 5% and 10% change from the
nominal condition
Figure 4.7 Effect of reflux ratio and reboiler duty variations towards 87
bottom product purity
Figure 4.8 Effect of reflux ratio variations towards the bottom product 89
(n-butane) purity. The input multiplicity occurrences are
located inside the drawn circle.
Figure 4.9 Effect of reboiler duty variations towards the bottom product 90
(n-butane) purity. The input multiplicity occurrences are
located inside the drawn circle.
Figure 4.10 Effect of reboiler duty using two different operating lines 90
towards the bottom product (n-butane) purity. The output
multiplicity occurrences are located inside the drawn oval
shape.
Figure 4.11 Reboiler duty profile 91
Figure 4.14 Setpoint tracking response results for LCIMIC, 2DOF IMC 96
and IMC
Figure 4.15 Setpoint tracking result for a step up change at t=5 hours 96
Figure 4.16 Setpoint tracking MV profile results for LCIMIC, 2DOF IMC 97
and IMC
Figure 4.17 The MV profile for step up at t=5 hours 97
xiii
Figure 4.18 Input disturbance rejection results for CIMIC, 2DOF IMC and 99
IMC
Figure 4.19 Input disturbance rejection MV profile for CIMIC, 2DOF 99
IMC and IMC
Figure 4.20 Output Disturbance rejection results for Linear CIMIC, 2DOF 101
IMC and IMC
Figure 4.21 MV profile for Linear CIMIC, 2DOF IMC and IMC output 101
disturbance rejection test
Figure 4.22 Setpoint tracking CV profile results for LCIMIC-AS and 104
LCIMIC
Figure 4.23 Setpoint tracking MV profile results for LCIMIC-AS and 104
LCIMIC
Figure 4.24 Disturbance rejection results for LCIMIC-AS and LCIMIC 105
Figure 4.25 Disturbance rejection MV profile results for LCIMIC-AS and 105
LCIMIC
Figure 4.26 Model Y response profile 109
Figure 4.31 Overall setpoint tracking test results for NLCIMIC and 117
LCIMIC-AS
Figure 4.32 Setpoint tracking MV results for NLCIMIC and LCIMIC-AS 115
Figure 4.33 The step up test response for NLCIMIC and LCIMIC-AS 118
Figure 4.34 The MV profile for NLCIMIC and LCIMIC-AS in the step up 120
test
Figure 4.35 The step down test response for NLCIMIC and LCIMIC-AS 120
Figure 4.36 The MV profile for Nonlinear and Linear CIMIC in the step 120
up test
xiv
Figure 4.37 Disturbance rejection test response profile for the NLCIMIC 122
and the LCIMIC
Figure 4.38 MV profile for the disturbance rejection test between the 122
NLCIMIC and the LCIMIC
Figure B.1 Reboiler duty profile for setpoint data generation 145
xv
LIST OF ABBREVIATIONS
CV Control variable
DV Disturbance variable
FB Feedback
FF Feedforward
xvi
GC Gas chromatography
H H-infinity
JIT Just-in-time
LM Levenberg-Marquardt
MV Manipulated variable
NN Neural network
QB Reboiler duty
xvii
QP Quadratic prediction
R2 Coefficient of determination
RR Reflux ratio
SP Setpoint
SS State space
TF Transfer function
xviii
LIST OF SYMBOLS
Disturbance
Effect of disturbance
Bottom flowrate
Distillate flowrate
Hessian matrix
Jacobian matrix
Reflux flowrate
Process
Process model
Nonlinear process
Controller
Nonlinear controller
Setpoint
Controller effort
xix
Vapour boil up flowrate or right singular vector matrix
Greek letters
Learning rate
Momentum rate
Delay ratio
Tuning parameter
Deviation
xx
PENGGABUNGAN KAWALAN MODEL DALAMAN DAN
ABSTRAK
industri adalah masa lengah dalam pengukuran dan analisa komposisi. Kelengahan
ini boleh mengakibatkan masa mati dalam gelung kawalan dan mengurangkan
penyulingan juga boleh menjejaskan kualiti hasil penyulingan dengan teruk. Salah
satu daripada skima kawalan yang mempunyai potensi untuk menyelesaikan semua
masalah yang telah dinyatakan adalah Penggabungan Kawalan Model Dalaman dan
Oleh yang demikian, dalam kajian ini, CIMIC tak lelurus (NLCIMIC) telah
Aspen dan telah berjaya disahkan berdasarkan data loji sebenar yang terdapat dalam
literatur. Selain itu, kajian tahap ketaklelurusan proses penyulingan juga telah dinilai,
xxi
IMC dan IMC, dan hasil yang diperoleh telah membuktikan kelebihan elemen
lelurus dengan model tak lelurus yang dibangunkan dalam Aspen (LCIMIC-AS)
digunakan sebagai model dan model songsang. Model NN yang terbaik telah dipilih
berdasarkan kepada nilai MSE diperoleh. Dalam usaha untuk menilai prestasi
titik set, NLCIMIC telah didapati melaksanakan pengesanan titik set langkah naik
dengan lebih baik (IAE = 0.0349) jika dibandingkan dengan LCIMIC-AS (IAE =
gangguan telah dapat diperhatikan dalam ujian penolakan gangguan (IAE = 0.0107)
xxii
NONLINEAR COMBINED INTERNAL MODEL AND
ABSTRACT
One of the main constraints in industrial distillation control is the time delay
in the composition measurement and analysis. This delay can introduce dead time in
the control loop and deteriorate the controller performance. In addition, the
quality. One of the promising control schemes that can handle all the problems
mentioned earlier is the combined internal model and inferential control (CIMIC) as
based on a linear model which can degrade its performance when dealing with
based on actual plant data available in the literature. In addition, a degree of the
nonlinearity study of the distillation process is also evaluated which then classifies
CIMIC (LCIMIC) has been compared to 2DOF IMC and IMC and the results
achieved proved the advantage of the inferential element embedded in the LCIMIC.
xxiii
Then, the linear based CIMIC integrated with a nonlinear based model developed in
Aspen (LCIMIC-AS) is implemented and the results obtained justify the need of the
model and inverse model. The best NN model chosen is based on the MSE value
process. Based on the setpoint tracking test, it is found that the NLCIMIC performed
better in tracking the step up setpoint (IAE = 0.0349) if compared to the LCIMIC-AS
is also observed in the disturbance rejection test (IAE = 0.0107) as it outperforms the
process.
xxiv
CHAPTER 1
INTRODUCTION
and unique approaches for control system applications compared to other fields such
considerations and equipment designs which make the control approach distinctive
from others. In order to understand these issues, Bequette (1991) and Rhinehart et al.
(2011) listed a number of challenges faced by the CPI which cause the control
Nonlinear process behaviour the process gain and time constant change for
different states.
Non-stationary process the process gain and time constant change with
time.
corresponding CVs.
safety considerations.
1
Individuality CPI have different behaviours even for the same kind of
operational conditions.
unprecedented reactions.
Furthermore, various chemical processes have been reported in the past and
the present that exhibit multiplicity behaviours which are the occurrence of two or
more steady states in the process (Koppel, 1982; Ma et al., 2010). These behaviours
have been categorized as mild and strong degree of nonlinearity by Pearson (2003)
which poses critical modelling and control limitation problems (Chidambaram and
Reddy, 1996). Rhinehart et al. (2011) also listed several control schemes that have
been accepted by the CPI based on real industrial practices. They are divided into
Basic control generally first level single input single output (SISO)
techniques in the control system, thus many of them are based on the analog
2
Advanced Process Control (APC) - Controllers which are developed
Generally, the controller selection for any process depends on many factors
such as the behaviour of the process, the input-output variables pairing, practicality
issues, constraints and costs. Rhinehart et al. (2011) and Hernjak et al. (2004)
provided some guidelines in selecting the controller based on the characteristics and
complexity of the process. The rule of thumb in controller selection is always start
with basic controllers and only move up to the next advanced controller if the
simpler scheme is not robust enough. The APC is known to perform better when
1996).
Furthermore, the implementation of the APC can increase the CPI financial
profits (Bauer and Craig, 2008). From their survey, the APC implementation can
product quality and reduce waste. Thus, the APC implementation in the CPI is
controllers.
plants. It is estimated that 95% of the separation processes in the CPI and refinery
plants in the world use the distillation process (Enagandula and Riggs, 2006). In
3
general, distillation is a method to separate a mixture of components of a liquid
solution based on the distribution of these components between the liquid and vapour
phase. During the distillation process, all the components are present in both phases
and are separated based on their boiling points. The distillation operation demands a
high usage of energy and cost. Dez et al. (2009) reported that 60% of the energy
used in the chemical industry is consumed by the distillation process alone and
Cheremisinoff (2000) stated that the distillation operation cost can contribute up to
Since the distillation process is a widely used and costly separation process in
the CPI thus, it is important that it can be properly controlled and maintained.
distillation process can be caused by many factors such as the non-ideal Vapour
Liquid Equilibrium (VLE) of the mixture, the complex column design, dissimilar
materials and heat balance at each tray and operation problems such as tray wetting
and fouling (Luyben, 1987). Furthermore, acquiring higher product purity (more
than 98%) can cause the system to become highly nonlinear. The distillation column
is also reported to exhibit multiplicity behaviour in its process (Zheng et al., 1998).
Due to such attributes in the distillation process, the controller task in the distillation
One of the conventional and generally used control schemes in the distillation
Rhinehart et al. (2011), it is estimated that about 95% of all control scheme
applications in the CPI are based on the PID controller. The PID controller is
4
commonly preferred due to its simple design, easy implementation and reliability in
system behaviour, the effectiveness of the PID controller is often degraded when
achieving its desired output during certain operating conditions and constraints
handling (Dutta and Rhinehart, 1999). Gokhale et al. (1995) have shown that the
Advanced Process Control (APC) have been proposed for the distillation control.
computer based controller, software sensor (soft sensor) and a process model in the
nonlinearity process and handling disturbance in the distillation column. Among the
APC schemes, the Model Based Control (MBC) strategy has attracted a great deal of
interest from process control practitioners. The MBC is preferred due to its generic
control strategy with a wide selection of process models and controller (or optimizer)
designs.
One of the well known MBC techniques is the Internal Model Control (IMC).
algorithm by utilizing a process model explicitly in the controller design. The IMC
has also gained some degree of prominence due to its straightforward control
strategy, control law, stability property and ability to integrate with existing PID
controllers (Morari and Zafiriou, 1989). In addition, the IMC strategy is practised in
many CPI processes such as CSTR (Varshney et al., 2009), combustion (Awais,
5
al., 1997) and even distillation process (Alina-Simona et al., 2011). However, the
IMC performance can suffer greatly from the distillation process nonlinearity
behaviours and disturbance effects if the process model is not properly developed. In
distillation column with non perfect mixing. Abdullah et al. (2007) reviewed the
models that were previously used in the distillation control scheme and concluded
that the application of a nonlinear model is necessary due to the behaviour of the
is difficult to control. Hernjak et al. (2004) suggested that the controller selection
should be made based on the behaviour of the process such as degree of nonlinearity
and dynamic characteristics. Pearson (2003) summarized some criteria to identify the
guideline in order to select the best controller for the distillation process.
concentration properly in order to obtain the correct feedback on the current stream
condition. The ideal solution would be to use an online composition analyzer that
can determine the product stream purity instantaneously. However, the problem with
the composition analyzer is the considerable sampling and measurement delay which
can affect the controller performance (Kano et al., 2003). An economical and
temperature, the more accurate dynamics of the distillation process can be estimated
6
in order to give the best measurement for product purity (Kumar and Kaistha, 2007;
Luyben, 2006b).
capability is very important since disturbance can occur at anytime and from
anywhere. Generally, the disturbances in the distillation column are measured from
the feed condition in terms of flowrate and temperature. However, there are other
disturbances in the distillation process which are typically hard to measure such as
upstream feed composition, sudden pressure drop in the column, tray conditions such
as sudden flooding and weeping, and surrounding temperature changes. Thus, the
product quality.
distillation process which can cater to the issues mentioned earlier i.e. nonlinearity
behaviour, the existence of measured and unmeasured disturbances and time delay.
In the past, there is one control scheme that could address most of the issues
mentioned earlier that were reported in the literature which is known as the
combined internal model and inferential control (CIMIC). The CIMIC was originally
process is inferred from the secondary variable which is, in this case, the columns
perform well since the tray temperature is more sensitive to changes in the
product composition. The CIMIC advantage lies in its transparent framework in the
control scheme design and the application of the tuning parameter. However, the
7
CIMIC proposed is based on a linear model and thus the control scheme is expected
iii. To develop Neural Network models for IMC and inferential control strategy
distillation column
The main focus of this research is to study the performance of the nonlinear
based CIMIC (NLCIMIC) over the linear based CIMIC. This comparison study is
carried out in order to observe and highlight the significance of the nonlinear model
embedded in the NLCIMIC. The inferential strategy in the CIMIC control scheme is
8
developed implicitly to cater the disturbance problems in the control scheme
process, a SISO control configuration can also be used when only a single product
in this study, the bottom product (n-butane) is favoured over the distillate product (i-
butane) since the bottom product would be sold directly in the market (Neste Oil,
2013). At the same time, the effect of the measurement delay and disturbance also
occurs in the bottom product control loop which needs to be handled properly.
In Chapter 1, the general idea of the current work is presented through the overview
of the applications of the control scheme in CPI and the research background. Next,
the problem statement reveals the related issues regarding the research area and the
ways to solve them. In addition, the purpose of this research is also highlighted via
the research objectives. Finally, the summary of each chapter is provided at the end
of this chapter.
the chapter. Then, the n-butane/i-butane separation process, which is the distillation
control, the IMC development and the IMC implementation in the distillation
column is also presented. The general concept of the CIMIC scheme is also
9
presented and finally a short introduction on the neural network theory and its
In Chapter 3, the methodology of the overall research is explained. The first part of
the methodology is the development of the steady state and dynamic distillation
column model using Aspen. This is followed by the sensitivity analysis and the
introduced to show how the data for the distillation column model is generated.
Then, in the linear CIMIC section, the development of the proposed control scheme
is presented from the IMC and two degree of freedom IMC design. In the Nonlinear
CIMIC section, the development and implementation of the neural network in the
CIMIC scheme is presented. In order to evaluate the distillation column model and
control scheme performance, certain benchmark guidelines are used which are given
In Chapter 4, the research results are presented and discussed. The first results are
from the distillation column Aspen modelling and validation results. Next, the results
from the sensitivity analysis and nonlinearity study are presented. The data generated
for the distillation column modelling is also shown. Furthermore, the results from the
linear based CIMIC are presented along with the comparison with other linear
control schemes. Then, the neural network modelling and control results in the
nonlinear CIMIC are presented and discussed. Finally, the comparison of the linear
discussed.
10
In Chapter 5, a concise conclusion of the current study is presented and some
11
CHAPTER 2
LITERATURE REVIEW
to separate mixtures into their respective components. The basis of the distillation
the liquid and vapour phase must be different for them to be separated (Halle and
is called the Vapour-Liquid Equilibrium (VLE) which can be correlated by using the
Figure 2.1. The distillation process begins from the upstream process which enters
the column as feed product (F) with, zf concentration. Generally, the feed stream is
disturbing the distillation process equilibrium. The section of the column above the
feed tray is called the rectification section and the section below is referred as the
stripping section. At each tray, the counter current of vapour and liquid flow takes
For a multicomponent distillation system, the term key is used to refer to the point
of separation that occurs between the components. The saturated vapour that is rich
with light key component is condensed at the top using the condenser to be produced
as distillate product (D). The saturated liquid that contains most of the heavy key
12
component is collected at the bottom of the column, known as the bottom product
(B).
Inside the tray, the mixture is heated and condensed into the vapour and
liquid phase. During the equilibrium process, the heated vapour entering the tray
(from the bottom) is equal to the saturated vapour leaving the current stage with a
new equilibrium vapour fraction and enthalpy and the liquid entering the stage (from
the top) is equal to the saturated liquid leaving the current stage with a new
each equilibrium stage the mixture is assumed to be perfectly mixed. If the mixture
is not perfectly mixed, which is true in real applications, the Murphy efficiency can
13
2.2 n-butane/i-butane Process
a low relative volatility which is (Tedder and Rudd, 1978) and a large time
constant (Aske et al., 2008). The low relative volatility points out that both
components have near boiling points which make them difficult to be separated. In
fact, the n-butane and i-butane are structure isomers of butane with different
number of trays are needed. The significant number of trays can cause lagging in the
contributes to the process large time constant (Riggs, 2001). Skogestad and Morari
(1987) had proposed a calculation for the dominant time constant for the distillation
process.
In addition, Finco et al. (1989) said that the distillation column that separates
close boiling materials tend to have a large reflux ratio (greater than 10), a large
number of trays (more than 100) and a long time constant (2-10 hours or more)
butane/i-butane has been studied in the past for the isomerisation process (Adeeva
and Sachtler, 1997), catalytic cracking (Roohollahi et al., 2012) and column
instruments and the control part is based on the control strategy implemented. The
14
effective control of the distillation process can be determined by controlling many
the distillation column process in order (Shinskey, 1996). Table 2.1 shows the
control variables (CV) in distillation which are need to be controlled and the reason
to do so.
these variables can be rejected and thus, the desired operation can be achieved. The
stream ( ) and its composition ( ). Generally, the distillation column inventory and
the pressure are considered as a low dynamic system which is handled by a basic
controller such as a PID. Nonetheless, one of the weaknesses of the PID control is its
inability to maintain a stable operation if there is some considerable dead time and/or
lag between the controller valve and the effect of the control variable (Riggs and
Ford, 2010). Although this condition may not be significant for level and pressure
15
There are many options in selecting the proper manipulated variable (MV) in
order to control product purity. Theoretically, the relationship between MVs and
DVs with CVs that is based on the steady state material balance can be used as a
guide. Based on the material balance split equation, the distillate and bottom product
purity is linked to the behaviour of the distillate flowrate ( ), the bottom flowrate
Furthermore, the steady state distillation equation also shows that the reflux flowrate
( ) and vapour boil up flowrate ( ) can affect the final product purity indirectly.
system which is easy to design and to control (Haggblom and Waller, 1990). Thus, a
regular distillation process would generally have 5 SISO control loops which are top
and bottom purity control, reflux drum and column level control and top column
pressure control. Currently, there are a variety of combinations of using the MVs and
its ratio to control CVs (Hurowitz et al., 2003; Sandelin et al., 1991; Skogestad et
al., 1990). According to Luyben (2006b), the selection for the best control structure
product purities, reflux ratio, column pressure, cost of energy, column size and
Furthermore, the pairing of the MVs and CVs for the MIMO process can be
(Seborg et al., 2004). One of the most well-known schemes is the LV configuration
16
where the top composition is controlled by manipulating the reflux flowrate and the
bottom purity is controlled based on the vapour boil-up rate. In many distillation
control practices, tray temperature is used to infer the composition. Thus, instead of
directly controlling the composition, the MVs are used to control the selected tray
temperature which is more practical and cost saving. This control scheme is shown
in Figure 2.2 together with pressure and level control. Here, the top composition is
The bottom composition is controlled by , by adjusting the steam stream inlet based
schemes, some issues on the approach of decomposing the system arise (Haggblom
due to the control loop interaction (Seborg et al., 2004). The control loop interacting
17
behaviour is caused by the process interaction when the controller output response of
a CV is also affecting other CVs. In order to resolve the interaction problem, one can
use the decoupling technique to reduce the interaction before implementing the
Apart from the conventional control of using an analog system, the advances
control into the process control field. Among those that had benefited from this
development is the model based control (MBC) techniques. The MBC is regarded as
model and controller. In comparison to the PID approach, the MBC action is
setpoint tracking which translate into better process safety and economics
(Rhinehart, 2006).
The general form of the MBC scheme is shown in Figure 2.3. The process
model (or model) block is developed to simulate and predict the actual process.
Then, the measurement from the block is compared with the actual process to
estimate the disturbance or model parameter. The mismatch from this comparison is
sent back to the model based controller or the optimizer blocks to compute the
control action needed to drive the process to its desired or optimum setpoint while
complying with the specified constraints. This scheme is in fact similar to the
18
Disturbances
Manipulated
Setpoint Model-Based Inputs Controlled Outputs
Controller or Process
Optimizer
Measurments
Figure 2.3 General model based control scheme (Brosilow and Joseph, 2002)
Although the main idea of the MBC is to develop an intelligent control strategy, the
selection of the model structure, the control objective, the adjusting mechanism and
the adjustable parameters has led to a variation of model based controllers. However,
the choice of the model structure is the primary and important aspect of the MBC
since it affects the control implementation and the accuracy of the control decision.
The MBC has been tested experimentally to prove its superiority over the
conventional control. Gupta and Rhinehart (1995) tested the MBC control strategies
distillation column. Based on their observation, the MBC produced a smooth and
faster response in controlling the setpoint and rejecting the disturbance. Meanwhile,
the ACC produced a noise-like response for the same control tasks. Subawalla et al.
(1996) evaluated experimentally several MBC strategies and the ACC using plasma
reactor system and distillation column. In their study, the nonlinear MBC strategies
showed better performance in the servo test when compared with the ACC. Joshi et
al. (1997) also experimentally tested the MBC and ACC control strategies in a pilot-
scale fluid flow and heat exchanger unit. In their work, the MBC strategies were
found to be more effective compared to the ACC in terms of both setpoint tracking
and disturbance rejection. Riggs (2000) made a comparison of the MBC and PI
19
control for a main fractionator column control. It was found that the MBC
outperformed the conventional control in controlling the light key and heavy key
components.
One of the well known MBC control schemes is known as the internal model
control (IMC). The IMC scheme is based on the internal model principle which
states that in order to achieve good control, the controller strategy must incorporate
Compared with other MBCs, the IMC provides a more straightforward design and
transparent framework for the control scheme development (Saxena and Hote, 2012).
In addition, the advantages of the IMC can be shown by its close loop stability
properties, perfect control and zero offset performance (Garcia and Morari, 1982).
Thus, the IMC control scheme is selected for this work and will be further
elaborated.
strategy that utilizes an internal model with the inverse model control technique to
compute its controller action. The early application of the internal model control
Garcia and Morari (1982) with the proper foundation. Afterwards, the IMC
theoretical framework was further developed and discussed through a series of work
(Economou and Morari, 1986; Economou et al., 1986; Garcia and Morari, 1985a;
Garcia and Morari (1982) first developed the IMC for the discrete single
input single output (SISO) continuous process. In their work, the development of the
20
IMC methodology was further detailed. They evaluated the IMC scheme using a
second order minimum phase and a non-minimum phase system. The results of the
IMC for both cases in their study were acceptable. In addition, the approximation of
the inverse model can be used in cases where the inverse is unstable and produce
excessive response. Thus, the researchers applied the numerical predictive method
Garcia and Morari (1985b) widened the IMC design scheme for the
multivariable (or MIMO) system. In the MIMO system, the usual trade off between
the good control performance of one variable and another is often regarded as a
difficult challenge. Thus, in their work the ability of the IMC scheme to serve as a
transparent controller scheme and achieve close-loop stability in MIMO system was
input constraints and close-loop decoupling effect were also addressed in the MIMO
IMC design procedure. The MIMO IMC scheme was tested using Wood and Berrys
distillation column and it performed well in the servo and regulator problem. In
addition, the application of the Feedforward IMC (which exploits the addition of
feedforward model and compensator in the original IMC scheme) was reported to be
One of the main concerns of the IMC control scheme is to obtain an exact
and stable inverse of the process. Without this, some of the prominent advantages of
the IMC scheme would not be available. Basically, the method of acquiring the
21
rational and stable part of the process through factorization is awkward and is
possibility of the zeros that appear close to the unit circle (region of convergence)
will give an undesired response characteristic (Garcia and Morari, 1985b). Thus, in
their work Garcia and Morari (1985b) used the method of approximation of the
inverse through the model predictive control law. In their simulation work, it was
shown that the approximated inverse model controller can perform well if it was
given the right tuning. Moreover, the IMC controllers using the factorization of time
Another advantage of the IMC scheme is its ability to combine with a PID
Since the PID is notably the most used controller in the process control, this
terms of solving reset windup problem, dead-time and coupling effect. Even if the
IMC is not explicitly combined with the PID, the IMC technique can be used to
develop tuning rules for the PID. Rivera et al. (1986) discussed this issue and
developed the tuning rule for the IMC-PID to be used with the existing PID. The
IMC-PID tuning offers a more simple tuning method (only one tuning parameter)
and compromising the performance and robustness trade off issues. In addition,
Skogestad (2003) also introduced a set of tuning settings for the IMC-PID based on
The application of the IMC was further expanded by Economou et al. (1986)
in a nonlinear system application. In their work, the design of the IMC was extended
to control a stationary MIMO lump parameter model. The Lump parameter (or
22
ordinary differential equations (ODEs). The capability of the IMC design philosophy
advantages of the IMC in a linear system are passed on to the nonlinear system. One
of the main issues here is the availability of an exact and stable inverse of the
nonlinear system. Thus, this issue has been discussed and some methods are
proposed. It should be noted that numerical inversion method i.e. Newtons method
Another design variation of the IMC scheme was proposed by Economou and
Morari (1986) who introduced the multiloop structure. In the former multivariable
scheme, a single IMC controller was designed to handle the entire variables at once.
However, in the multiloop structure, each of the variables was controlled by its own
respective SISO IMC controller. Economou and Morari (1986) first developed the
SISO IMC design strategy in a multiloop environment based on the Small Gain
Theorem. The idea of the multiloop IMC was to treat the interaction as an added
deterioration and interaction measurement was created in order to analyze the control
loop and pairing system. In the results obtained from simulating a packed-bed
reactor, the multivariable IMC performed better than the multiloop IMC in the servo
test. In addition, the integration of the steady state decoupler further improved the
In order to prove the practicality of the IMC theory, Arkun et al. (1986)
tested it experimentally for a heat exchanger and stirred tank control. The heat
exchanger and stirred tank is a typical distributed parameter system which displays
characteristics such as time delay and nonlinearities which are synonym with
23
industrial chemical processes. In the first case, the SISO IMC was used to control the
heat exchanger temperature. The purpose of this test was to show the simplicity of
implementing the IMC techniques in real time and that could be easily adjusted
online for good performance and robustness. In the experiment, the effect of the
filter time constant for the step change and disturbance rejection test was evaluated.
During the step test, the overshoot and the oscillation of the response was minimized
by increasing the filter time constant. However, this produced a much slower
response. The model sensitivity and time delay mismatch were also evaluated in the
work. In the presence of model mismatch, if the gain of the model was smaller than
the gain of the process, it was found that the controller overacted and overshoot
response occurred. If the gain of the model was larger than the process gain, then the
controllers behaviour was similar to the integral action with decay oscillation
response. Arkun et al. (1986) also compared the IMC performance with a decoupling
PI controller in controlling the stirred tank. Based on the overall test results, the IMC
outperformed the PI control for output containing time delay (temperature control).
As for the control output with no time delay (level control), both controllers gave
comparable performances.
Wassick and Tummala (1989) used the IMC for the multivariable control for
a full scale distillation column model. In their work, a reduced order model was used
to develop the controller and a feedforward decoupling was also introduced to handle
the control loop interaction. The feedforward decoupling was selected over
conventional decoupling due to the different sampling times for both the controllers.
The feedforward decoupling was designed based on the Garcia and Morari (1985a)
decoupling technique to solve the multi sampling time control loop. In this
24
technique, the feedforward compensator for the bottom product (slow sampling rate)
was based on the trailing average of the top product (fast sampling rate). The
feedforward compensator for the top product could be used directly since the bottom
product response was much slower than the top product response. This gave a
Henson and Seborg (1991) developed a new approach for the IMC control
strategy based on the work proposed by Garcia and Morari (1982) on linear IMC. In
their work, the global input-output linearization technique was used to compute the
IMC controller for a SISO system with affine input and stable inverse. Based on a
reasonably mild assumption, the developed closed loop system can be considered to
behave as a linear IMC in terms of zero offset and perfect control ability. One
directly on certain regions of state space (on singular manifolds) where the relative
degree is not well defined and not invertible (Patwardhan and Madhavan, 1998).
full state feedback in its control law synthesis. However, based on the new approach
from other researchers, the dependence on the full state feedback can be eliminated.
The inverse model was developed based on the Hirschorn inverse (Hirschorn, 1979)
disturbance was also proposed. For comparison, the performance of the Nonlinear
IMC (NIMC) was compared with the Input-Output Linearizing (IOL) Control with
full state feedback and the PI controller on a continuous fermentor model. In the
setpoint tracking problem, all the controllers showed similar responses with
comparable controller efforts. However, for the disturbance rejection test, the IOL
25
performance was superior than the NIMC and PI control since it used all the three
Hunt and Sbarbaro (1991) used the neural network in their nonlinear model
based control system. The IMC scheme was chosen due to its simplicity and because
the neural network can be incorporated directly. In their work, a neural network
based on the Gaussian activation function was trained to represent the system
response and its inverse. The architecture and learning algorithm for the neural
network model and inverse used were based on the work of Psaltis et al. (1988). A
numerical approach using the recursive method to represent the system inverse was
explored and integrated with the NN for better inverse performance. In the
simulation result, the proposed control scheme managed to track the setpoint and
Psichogios and Ungar (1991) implemented the neural network in their model
based control scheme. In their approach, the neural network application into the IMC
was divided by the direct and indirect approach. In the direct approach, the neural
network was trained to represent the IMC process model and inverse controller
explicitly. In the indirect approach, a rigorous inverse process model was developed
by inverting the process model using Newtons Method. This method required the
controller action to be computed online based on the inversion of the neural network
process model at each sampling time. The performance of the direct approach
seemed to be faster to reject disturbance than the indirect approach but resulting in a
longer settling time. On the other hand, in the study, the linear process model was
also developed using the linear regression model and tested with the nonlinear
process model. In the performance test, the nonlinear process model performed
26
Nahas et al. (1992) applied the NN in the IMC for a nonlinear SISO process.
In their work, the NN model was identified using open-loop data and trained with a
conjugate gradient algorithm. The NN controller for the NIMC (Nonlinear IMC) was
based on the inversion of the NN model to ensure offset free performance. However,
the NIMC approach was restricted to processes with stable inverse only. A filter with
a single tuning parameter was added to the scheme to deal with the plant-model
mismatch. The NIMC control strategy was developed based on two approaches
which was numerical inversion using optimization technique and the Neural
Network direct inverse method. The proposed strategy also included the time delay
compensation based on the Smith predictor for the numerical inversion method. The
neutralization process. The results showed that the NIMC outperformed the
using the IMC scheme for controlling a nonlinear system. In their work, the IMC
inverse controller was developed based on the quadratic model control law. The
start-up operation and pH system. During the control of the CSTR start-up operation,
the quadratic model control law performed better by reducing the target settling time
while maintaining the controller output in operating range when compared to the
was also done in the CSTR control to study the controller robustness. In the test, the
quadratic model control law was able to produce stable response while the pseudo-
Newton based IMC response became oscillatory. For the pH control problem, similar
27
results were observed which showed the superiority of the IMC using quadratic
Berber and Brosilow (1999) proposed a new strategy for an unstable system
called the Algorithmic Internal Model Control (AIMC). The problem of the unstable
system is critical in the IMC design because all the properties of the IMC are based
on the assumptions of the stable system used. Thus, in order to address this problem,
Morari and Zafiriou (1989) proposed a two degree of freedom IMC (2DOF IMC) for
controlling the unstable system. However, the implementation of the 2DOF IMC
could still experience instability unless the filter was increased drastically beyond the
requisite for robust performance. Hence, the aim of the AIMC control structure was
to implement the two degree of freedom IMC for unstable system without increasing
the filter time constant substantially, thus degrading the control robustness. The
AIMC control strategy was presented in terms of the sequence of tasks to compute
the desired controller output that forced the model to track the desired setpoint. The
model state feedback (MSF) was also integrated into the new scheme to produce an
improved controller effort. By using the MSF, the controller effort was calculated at
each sampling time based on the current model state and the desired response. The
application of the AIMC was tested on linear and nonlinear systems. In the linear
system, the AIMC performed better than the 2DOF PID for setpoint tracking and
disturbance rejection. In the nonlinear system, the AIMC was reported to produce an
indistinguishable performance with the MSF and global linearizing control (GLC) in
inferential technique to handle disturbance more effectively in the process. The new
scheme is called the combined internal model and inferential control (CIMIC) and it
28
is a continuity development from the Disturbance Rejection and Decoupling (DRD)
scheme carried out by Haggblom and Waller (1990). In the CIMIC, the disturbance
in the primary variable is measured through the behaviour of the secondary variable
using the inferential loop. Since the secondary variable responds significantly faster
than the primary variable, the disturbance in the primary control loop can be rejected
promptly. In their research, the CIMIC controller parameters are developed via the
the unstable system application in a simpler and plain way. In his work, the
parameterization of the stabilizing controller was integrated into the original IMC
scheme. Thus, due to the stabilizing parameters, the proposed control scheme was
Parameter adaption and additional feedback loop techniques were used in the
IMC to reduce modelling error when dealing with a nonlinear system. Hu and
Rangaiah (1999) developed the adaptive IMC (AdIMC) for a nonlinear process
based on the input-output linearization (IOL) technique. In their work, a new scheme
of the IMC based on IOL techniques was proposed in order to reduce the state
variable needed. Additionally, the IMC-IOL also improved the controller action and
robustness. The unknown parameter in the control law was estimated by the adaptive
law based on the process output, model output, and state variables which were
predicted by the model. Simulation results showed that the performance of the
Hu et al. (2000) added an additional loop to the original IMC scheme where
29
controller. Then, the response from the compensation control was used to update the
IMC controller. The proposed scheme, called the Augmented IMC (AuIMC), was
then tested online using an experimental neutralization process rig along with PI
control and a nonlinear IMC using the input-output linearization technique. The
AuIMC performed faster than the nonlinear IMC in the setpoint tracking test while
the PI control produced a sluggish response. In the regulatory test, the AuIMC had
this work.
Wang et al. (2001) simplified the IMC design by combining the inverse
model and the process model of the original IMC into a new block. Under the
assumption of an ideal case (i.e. a perfect model), the new block became the desired
closed loop transfer function controller in order to shape the process response
according to the setpoint. The closed loop transfer function controller was designed
based on the process characteristic and a tuning method for the new control scheme
was also developed. The new IMC design showed satisfactory performance on
Yang et al. (2002) applied the IMC based single loop controller design
method to develop a new controller for an unstable process with time delay. In their
study, the new controller was developed from the closed loop controller transfer
function using the parameterization technique. The new controller can be reduced to
resemble the classical feedback controller which can be approximated to the ideal
PID controller form or a higher order form. In the ideal PID controller form, the
parameters for the new controller are obtained based on the optimization technique
in order to find the best PID parameters. In the higher order controller form, the
recursive least square (RLS) algorithm is used in the model reduction method to find
30
the approximation to the new controller parameter. Between the PID controller form
and the higher order controller form, the latter shows better performance when tested
Lee et al. (2000) developed a modified PID for an unstable system with time
delay. However, the controller that was developed could only reject disturbance if
the ratio of the time delay to the time constant of the process was less than one. As
an alternative solution, Tan et al. (2003) further modified the IMC scheme to handle
a similar problem. In their work, the researchers proposed three new compensators
inside the IMC scheme. The first compensator was designed to stabilize the unstable
system by neglecting the time delay. The setpoint tracking ability was handled by the
second compensator which was the IMC controller based on the stabilized model.
The final compensator was used to handle disturbance and to stabilize the original
unstable and delayed process. From the test using different unstable systems with
varying time delays, the proposed scheme was observed to perform well.
technique in the Adaptive IMC design framework. The JIT model was selected due
to its simple design in modeling the nonlinear process. In their work, the JIT model
was developed based on a second order ARX structure and was used to update the
process model and to adjust the IMC controller parameter on-line via an adaptive
learning algorithm. The IMC controller was developed based on the inversion of the
JIT model augmented with a dynamic filter which was adjusted at each sampling
instant. The simulation results obtained showed the advantage of the adaptive IMC
process with delay ratio within , where is the ratio of time delay to time
31
constant. In their work, four new controllers were integrated with the original IMC
scheme. The first was the feedback controller which was used to convert the unstable
process to a stable process. The second was the cascaded inner loop controller to
suppress the disturbance. The third controller was for setpoint tracking using the
IMC controller while the last one was a feedforward controller to handle disturbance.
A low-pass filter was also used with the IMC controller. In their assessment test
using the First-Order Delay Unstable Process (FODUP) and Second-Order Delay
Unstable Process (SODUP), the proposed controller scheme performed better than
the controller proposed by Tan et al. (2003) and Lee et al. (2000).
and Zafiriou, 1989) based on the standard 2DOF IMC design to improve the
IMC scheme gave the freedom to design the feedback and feedforward controller
independently. Morari and Zafiriou (1989) pointed out if such perfect disturbance
relevant since the feedforward controller was more sensitive than the feedback
scheme was reported to perform better than the original 2DOF IMC and Skogestad
Chia and Lefkowitz (2010) developed a simple method for controlling the
integrating system using the IMC. In their work, the integrating process was
approximated with a first order lag with a large time constant. In addition, the
original IMC scheme was simplified similar to a feedback control scheme. The
proposed controller design was analytically proven and verified by simulation. Based
on the new IMC design, the numerical instability issues (due to the integrating
32
system) were handled properly, zero steady-state error was achieved and a new
tuning parameter (from the approximation model) was introduced for additional
degree of freedom in tuning the process response. The new IMC design was
implemented successfully in several real industrial case studies such as level control
Liu and Gao (2011) enhanced the IMC design for handling the disturbance
for integrating an unstable system with slow dynamic. In their study, the type of
disturbance was further classified as step and ramp which was often practised in the
2DOF IMC scheme, they formulated a new controller based on different types of
disturbance. In order to illustrate the controller performance, the new scheme was
tested with a modified IMC on integrating an unstable linear model. In the step
disturbance test using the integrating process, the performance of the proposed
controller was similar to the modified IMC. However, for the ramp and slow
dynamic disturbance test, the performance of the proposed controller was superior to
33
Multivariable
Garcia and Numerical
3 1985 control law and TF Simulation
Morari Predictive
tuning
Nonlinear
Economou et Numerical
5 1986 Multivariable ODE Simulation
al. Predictive
IMC
Economou and
6 1986 Multiloop IMC SS Inverse SS Simulation
Morari
Experimental
7 Arkun et al. 1986 TF Inverse TF Experiment
IMC
Input-Output
Henson and Hirschorn Inverse
9 1991 Linearization ODE Simulation
Seborg with IOL
(IOL)
Neural Network NN
12 Nahas et al. 1992 Inverse NN Simulation
(NN) Model
Unmeasured
13 Haggblom 1996 disturbance SS Inverse SS Simulation
rejection
Patwardhan Quadratic QP
14 1998 QP Control Law Simulation
and Madhavan Prediction (QP) Model
34
16 Yamada 1999 Unstable system TF Modified IMC Simulation
Hu and
17 1999 Adaptive IMC TF IOL Simulation
Rangaiah
Cheng and
22 2007 Adaptive IMC JIT Inverse JIT Simulation
Chiu
Modification
24 Juwari et al. 2008 TF FB/FF IMC Simulation
2DOF IMC
Integrating and
unstable system Modified 2DOF
26 Liu and Gao 2011 TF Simulation
with slow IMC
dynamic
Based on Table 2.2, it can be concluded that the IMC has been studied
linear model such as the transfer function model and tested in simulation
environments. Thus, there is still room for improvement in this area using nonlinear
35
can be a good way to evaluate the controller performance since it can practically
resemble real conditions. Based on Mcmillan (2006), process simulators can help
users to improve control system design strategies to reduce control loop errors,
reduce start-up time, and improve on-stream time. Furthermore, one of the aspects
that has caught the interest of researchers is the modification of the IMC design
scheme. Since the original IMC framework is theoretically simple and robust, the
modifications can be done explicitly. It should be noted that researchers had tried to
improve the IMC in handling an unstable system even though theoretically, the
original IMC cannot control unstable systems due to the unstable closed loop
response.
highlighted by many researchers, only one-third of them had mentioned handling the
unmeasured disturbance issues. In the standard IMC, the disturbance is often treated
disturbance can occur in the beginning of process which makes it hard to identify
until it already affects the process. This condition can be worse for a slow response
primary variable as the action taken by the controller will depend on how fast the
control scheme has been identified as a possible solution in handling this problem.
By using the inferential technique, disturbance can be compensated earlier and faster
36
2.5.2 IMC in Continuous Distillation System
Garcia and Morari (1985a) were the first to implement the IMC control
IMC design using the binary distillation model from Wood and Berry (1973). The
IMC design was based on the discrete time transfer function and the IMC
performance was compared with the PI control and the multivariable dead-time
compensator. In their work, the optimal factorization matrix was used to realize the
decoupling and achieved its setpoint in two minutes despite a strong input action. For
the servo response, the SISO PI control alone gave high oscillation and interaction.
When the compensator was used with PI controllers, a less oscillation response was
achieved. However, the variable interaction was still strong and the settling time was
quite long. For the IMC, the result was better from both the PI with or without
input and two output linear model of the industrial distillation column was built
based on experimental data using discrete-time transfer function. Due to the high
order of the system, the controller was designed from an inverse of a reduced-order
model of the process. This reduced the high-frequency control actions and
complexity of the control program. Besides that, the advantage of the FFIMC was
that the measurable disturbance in the process was compensated by the FFIMC to
cancel their effect on the process. The modified part of the FFIMC was to allow the
37
feedback control calculations for both controllers to be made independently and then
their results were processed and added to each other to determine the overall control
signal. Since the controller was based on a multiple rate sampling, a critical
decoupling was introduced and a trailing average of the overhead controller output
was used for the bottom controller. In their study, the controllers proposed were
developed from discrete-time transfer function. From the simulation, the advantage
Basualdo et al. (1994) implemented the neural network in the IMC design for
was used as the network architecture and the system identification of this process
was developed from the simulation data. The reflux stream and its consequent
product composition at different operation points were selected as the input to the
was capable to generalize the dynamic process. In their study, a feedforward neural
network model was developed to estimate the distillate composition of the column to
represent the internal model of the IMC. The results of the top composition control
of the distillation column showed that the IMC based neural network demonstrated a
Fieg et al. (1996) studied the performance of the IMC in comparison with a
column. The research was proposed to deal with measurement delay by using online
gas chromatography (GC) and problems in using the tray temperature control
because the oleo chemical products temperature was less sensitive. The IMC was
developed based on the linear transfer function model. Based on the simulation
38
study, the IMC control scheme gave better results in less settling time and a small
Shaw and Doyle (1997) used the IMC based on the Recurrent Dynamic
Neural Network (RDNN) model for a two input and two output for a high purity
distillation column. The RDNN structure was similar to the Recurrent Neural
feedback. The difference was the RDNN weight which was based on the nonlinear
functions of the outputs. This gave the network additional ability to cover a wider
technique was used for the decoupling of the MIMO RDNN system to linearize the
internal model into a linearized transform system in order to apply the linear IMC
controller design. The RDNN control system was then tested on a high purity
distillation column model. From the study, the integration of a priori information
such as the two-time constant of the distillation behavior helped to improve the
model for open loop simulation. From the closed loop simulation and Singular Value
Decomposition (SVD) analysis, a model was identified in which both open loop and
closed loop data gave better control performance than the model identified solely
space form. The control system was implemented in the experimental methanol-
water distillation column. The new control strategy showed good performance in the
39
servo test with a fast settling time. Besides that, based on the model mismatch test,
control (CIMIC) scheme in controlling a pilot scale distillation column. In the study,
the CIMIC performed better than the standard IMC and 2DOF IMC control scheme.
However, the performance of the CIMIC was not as distinct as expected due to the
primary output (tray temperature) response which was almost as fast as the
inferential technique in the CIMIC at that time was not very profound.
Nonlinear IMC (NIMC) with an on-line estimator with a globally linearizing control
(GLC) and generic model control (GMC) strategies for a distillation column start-up
and operation. The NIMC approach used the nonlinear filter that can be adjusted for
model mismatch. The new input for the NIMC was derived to obtain the control
algorithm and the design was based on the continuous linear transfer function. The
rigorous dynamic column distillation model from Wood and Berry (1973) was used
in the research. The study concluded that the NIMC strategy was recommended due
to easy tuning and it produced good performance during the operation transition
Wang et al. (2002) proposed a new approach to the IMC analysis and design
for the decoupling and stabilizing of multivariable stable processes with multiple
time delay. In the study, the characteristics of a stable IMC were formulated in terms
of time delays and non-minimum phase zeros for the open-loop system. Based on
this characterization, the control design procedure was developed for best
performance and was further simplified using the model reduction technique. The
40
resulting controller was in the form of proper second order transfer functions in a
state space form. The performance evaluation of the proposed decoupling IMC
design was done in the binary distillation column. The results showed that the
performance of the proposed IMC was better than the Smith predictor controller.
Jones and Tham (2004) presented the IMC-PID design and compared it with
the Gain Phase Margin (GPM) based PID. Both control strategies were applied to a
in a Woods and Berry distillation column. The IMC-PID design method was easier
to use as it had only one design parameter related to the desired time constant of the
closed loop response. In the study, the Simplified IMC (SIMC) was used based on
the First Order plus Time Delay (FOPTD) model for the bottom loop and the Integral
plus Time Delay (IPTD) model in the distillate loop control. Based on the
while the GPM-PID approach performed well in handling the bottom response.
cost of variability in the controlled variables (control cost) with capital and operating
process. Conventionally, the capital and operating costs are first determined and
minimized to obtain the optimum process design based on a steady state model.
Then, in a subsequent step, the process control design is considered. However, in the
evaluated based on the closed loop control performance. Here, a robust IMC control
integrated and conventional method, the proposed method has significantly reduced
41
A combination of feedback and feedfoward control with the IMC concept
was also proposed for a two-point temperature control in a binary distillation column
the structure, construction, and tuning aspects of the control design problem of the
linear two-point temperature was explained. The decentralized one-way and two-way
decoupling control structure was also studied. The control model consisted of a
linear integrator based on the control input, an effective load disturbance and steady
state parameters that gathered the relevant interaction data. The controller consisted
control loops, with setpoint adjusters for feed temperature, built from the static
output temperature correlation on the feed temperature. It was found that the
proposed control system was able to capture the behavior that was similar to a model
Razzaghi and Shahraki (2007) studied the application of a control for a high
purity distillation column model. The column model was developed with the
uncertainty and dynamic behavior of the high purity column for the entire operating
condition. A Structured Singular Value (SSV) which was defined in terms of the H-
norm of the weighted sensitivity function was used to synthesize the controller and
to evaluate the control performance. The decentralized linear IMC based PID model
was developed and good set-point tracking and disturbance rejection of the controller
Shamsuzzoha and Lee (2008) proposed a simple analytical design method for
the PID controller based on the IMC concept for integrating and first order unstable
processes with time delay. In their work, a new tuning for the IMC-PID was tested
on several cases including the distillation column model. In the distillation column
42
test, the results demonstrated that the proposed technique was better than the
conventional PID tuning methods available in the literature for solving servo,
summarized in Table 2.3. It should be noted that the remarks column in Table 2.3
refer to issues in the particular reference that are related to the current research
Nonlinear
3 Basualdo et al. 1994 Binary (TF) Simulation SISO -
(FFNN)
Multi-
Linear Simulation Delayed CV
4 Fieg et al. 1996 component MIMO
(TF) (Industrial case) response
(TF)
Linear Inferential
5 Hggblom 1996 Binary (SS) Experiment MIMO
(SS) disturbance
Linear
6 Murad et al. 1996 Binary (TF) Simulation MIMO -
(TF)
Nonlinear Binary
7 Shaw & Doyle 1997 Simulation MIMO -
(RNN) (FPM)
Venkateswarlu Nonlinear
8 1997 Binary (TF) Simulation MIMO -
& Gangiah (IOL)
Linear
9 Murad et. al 1997 Binary (TF) Simulation MIMO -
(TF)
Linear Delayed CV
10 Wang et al. 2002 Binary (TF) Simulation MIMO
(TF) response
43
Linear
11 Jones & Tham 2004 Binary (TF) Simulation MIMO -
(TF)
Multi-
Chawankul et Linear Simulation
12 2005 component SISO -
al. (TF) (Aspen)
(TF)
Castellanos- Linear Binary
13 2005 Simulation MIMO -
Sahagu et al. (TF) (FPM)
Shamsuzzoha Linear
15 2008 Binary (TF) Simulation SISO -
& Lee (TF)
Based on Table 2.3, many of the applications of the IMC in the continuous
distillation control were based on a linear model which was either the transfer
function or the state space model. Furthermore, the majority of the work focussed on
the binary component distillation process. The most preferred binary distillation
process was the Woods & Berry distillation column (Wood and Berry, 1973). In
addition, only a few of the work used the industrial distillation process as a case
study.
primary variable which can deteriorate the controller performance. This slow
response can be caused by the measurement delay and slow process dynamic. Apart
from the slow response, disturbance can also give a significant effect to the
threat to the distillation product quality. However, these problems are not highlighted
measures in the control scheme to reject disturbance more effectively in the process.
Among the two control schemes that were mentioned, the CIMIC was
considered as a better option. The main problem with the feedforward controller is
44
the disturbance in the process needs to be modelled in order to be eliminated. Thus,
this will become a disadvantage for the feedforward control scheme when dealing
with unmeasured or unexpected disturbance. On the other hand, the CIMIC uses an
secondary variable process. Since all disturbances in the primary variable can be
perceived properly by the secondary variable fast behaviour, the disturbance can be
distributed processor made up of simple processing units called neuron that has a
natural propensity for storing experiential knowledge and making it available for use.
Neural network resembles the human brain in two aspects which are knowledge is
acquired by the network from its environment through a learning process and
interneuron connection strengths, known as synaptic weights are used to store the
45
Figure 2.4 Schematic diagram of a single neuron (Seborg et al., 2004)
figure, the neuron computes the weighted sum of the input signals and compares the
result with the threshold value from the activation function. The activation function
will determine the permissible amplitude range of the output signal to some finite
the weights inside the network in accordance through a predefined learning process
and algorithm.
Baughman and Liu (1995). The advantages are; nonlinearity input-output mapping,
fault tolerance and learning capability. In contrast, some of the limitations of the
ANN are also summarized by Baughman and Liu (1995). They are; long training
time, require a large amount of training data and no guarantee of optimal results.
Neural network has been applied in many fields including the control system.
The capability of the neural network as an effective system identification and control
system for the dynamic nonlinear process has been discussed by Narendra and
46
Parthasarathy (1990). Bhat and McAvoy (1990) successfully proved the neural
network ability to generalize and model chemical process system using the back
propagation technique. Basically, in the IMC control scheme, the neural network is
mostly used as the inverse model based controller. Hussain (1999) summarized the
chemical process control. He reported that many researchers used the IMC scheme
when applying the neural network as the inverse model based controller. Moreover,
Meireles et al. (2003) reviewed the neural network application in the industrial
problems. Based on the review, the neural network has been successfully
control.
Based on the previous works, the neural network modelling had been used in
many areas including distillation control. However, the approach of using neural
network for CIMIC control scheme has yet to be established. Thus, in this work, the
(NLCIMIC) design. Due to the inherent design of the IMC, the neural network
modelling technique can be used explicitly to represent the control scheme blocks.
Thus, with the generalization of the nonlinearity ability from the neural network, the
47
CHAPTER 3
METHODOLOGY
3.1 Overview
Figure 3.1. In this chapter, the methodology of the Nonlinear CIMIC development is
Project Initialization
Performance Test
No
Accepted?
Yes
Completion of the project
48
3.2 Research Software
In this research, two main software is used to develop the Nonlinear CIMIC
(NLCIMIC) control scheme. The first software is the process simulator, Aspen Plus
and its dynamic extension, Aspen Dynamic. This software is used to develop the
distillation column simulation model. The second software is the Matlab, which is
used in the development of linear and nonlinear modelling and control scheme for
order to implement the proposed NLCIMIC control scheme for the n-butane/i-butane
distillation process.
3.2.1 Matlab
and output results are expressed in a familiar mathematical notation. In this study,
the Neural Network Toolbox is used to aid the neural network model and the control
development processes especially in terms of the training algorithm. Besides that, the
dynamic systems or models by using block sets. This enables a representation of the
real time online system where everything is interconnected and computation happens
at a specific period of time. The Matlab System Identification Toolbox is also used
49
3.2.2 Aspen Plus and Aspen Dynamic
Aspen Plus is widely used in academic and industrial fields as a steady state
process simulator for the chemical process. Aspen Plus provides a user-friendly
package, variety process capabilities and unit operations available. Moreover, the
dynamics of the process can also be simulated by exporting models from Aspen Plus
into Aspen Dynamic. In Aspen Dynamic, the user is given the freedom to manipulate
the process based on a dynamic and practical simulation. Based on this, the dynamic
capability and nonlinearity behaviour of the plant can be further studied. Aspen has
also provided the AMSystem, a link between Aspen Dynamic and Matlab Simulink
to enhance a wider application for both software. In this work, Aspen Plus and
Aspen Dynamic are used to develop the steady state and dynamic model of the
distillation process.
Klemola and Ilme (1996) and Ilme et al. (2001) is considered which originated from
the Neste Oil refinery plant in Finland. The aim of the column is to separate i-butane
and n-butane from the upstream hydrocarbon feed containing approximately 29.4
wt% of i-butane, 67.7 wt% of n-butane, 1.5 wt% of propane and 1.0 wt% of
pentanes. The rest of the feed contains 0.5 wt% C4 olefins which are then lumped
into i-butene and 1-butene due to their small concentrations. Thus, there are a total of
eight components in the feed stream. In this case, i-butane and n-butane are
considered to be the light and heavy key components for the distillation system,
respectively. The column operating pressure is 650 kPa. It is 2.9 meters in diameter
50
and has 74 Glitsch Ballast two-pass type V-1 valve trays. A simplified version of the
45.1 `C
650 kPa
Tray 9 18.5 `C
47.5 `C
Feed Distillate
26234 kg/h 8011 kg/h
Propane 1.5 wt% Tray 65
i-butane 29.4 wt% 62.2 `C Propane 5.3 wt%
n-butane 67.7 wt% i-butane 93.5 wt%
Tray 74
C4 olefins 0.5wt% n-butane 0.2 wt%
63.2 `C
Pentanes 1.0 wt% C4 olefins 1.0 wt%
Pentanes 0.0 wt%
10.24 MW
The distillation column model is first developed based on its steady state
condition using Aspen Plus. In steady state modelling, all the values and conditions
used are based on the distillation column at its steady state operating condition and
the results achieved are compared with the actual distillation column data for
validation purposes (Fruehauf and Mahoney, 1993). The design parameters and
condition of the distillation column are based on the work from Ilme et al. (2001)
and Klemola and Ilme (1996) which are shown in Table 3.1, Table 3.2 and Table 3.3,
51
method because the process under consideration deals with high temperature and
pressure (Aspen Tech, 2009). For validation purposes, the overall column efficiency
is set at 110% which is based on the value considered by Ilme et al. (2001). The
efficiency of a distillation column can reach over 100% due to non perfect mixing in
the liquid phase as the liquid flows across the tray. This usually happens for a low
52
Table 3.2: Distillation column specification (Klemola and Ilme, 1996)
Downcomer Area
Column height (m) 51.8 0.86
(centre) (m2)
Column diameter (m) 2.9 Tray Spacing (m) 0.6
Weir length side (mm) 1.859 Total Hole Area (m2) 0.922
Table 3.3: Distillation column operation data (Klemola and Ilme, 1996)
Reflux rate (kg/h) 92838 Pressure drop per tray (kPa) 0.47
After the steady state distillation column model is validated, the model is
simulation and limitation on the steady state model are discussed by Mahoney and
Fruehauf (1997). Here, the nonlinearity and multiplicity study is carried out using the
model that has been developed previously. Some recommendations on designing the
reflux drum and sump are taken from Luyben (2006a) and Kister (1992). The sump
and reflux drum dimension are selected based on the heuristic assumption to set up a
5 minutes liquid holdup while the vessel is 50% full when entering and leaving the
vessel. The liquid hydraulics and temperature within the stages are calculated by
rigorous tray correlations provided in Aspen Plus using the calculation procedure
53
proposed by Glitsch (1974). By using the AMSystem provided in Aspen Dynamic,
the distillation column is then connected with the Matlab Simulink for further study.
depending on the vapour liquid hydraulic inside the column. There are also other
factors that can influence the temperature inside the column such as the temperature
of the feed, the reaction for reactive distillation and the pressure of the column. In
the absence of the concentration analyzer, the temperature of the stages can provide a
fairly good estimation of the composition. Thus, selecting the best stage temperature
is crucial in order to have an accurate inferential of the composition and for control
purposes. The general rule of thumb for choosing the tray temperature to be
controlled is based on the location of the manipulated input thus, to control the
column using the reboiler duty or the steam stream valve, the appropriate trays to
choose should be located near at the bottom. Similarly, trays at the top should be
used for the reflux flowrate control. However, since every column is different in
many aspects from one another, the rule of thumb only serves as a guideline and a
best stage temperature for a multicomponent distillation system. The SVD equation
(3.1)
54
where G is the process steady state gain matrix, U is the left singular vector matrix,
is the diagonal matrix of singular values and V is the right singular vector matrix. U
indicates the most sensitive output direction for simultaneous change in the MV in
the same direction in G. The ratio of highest to lowest number in the matrix is used
this work, U represents the tray sensitivity measurement matrix where the highest
Usvd value refers to the most sensitive tray temperature towards the respective MV.
The steady state gain matrix is developed from the effect of the small MV change
For a continuous distillation column, there are many inputs and outputs that
operate throughout the system. Thus, the sensitivity analysis is done to determinate
the significance of the input towards the output of the process. In this study, the
selected inputs of the process are the reboiler duty, the feed flowrate and feed
composition. The considered outputs of the process are the product composition and
the tray temperature. The input-output relationship can provide a useful process
of step inputs to observe the behaviour of the output response. The step-test with
and the feed composition are introduced to the system. As for the feed flowrate, a
55
used. The selection of the reboiler duty range is made based on the predicted
based on the suggested practical industrial disturbance range for the feed
One of the difficult parts in model based control strategies is the selection of
the appropriate model structure for the nonlinear system. Normally, the selection of
the model is made based on linear intuition and heuristic approach, which is still
vague and not practical. Pearson (2003) proposed several methods to properly
their complexity. Thus, with this method, the degree of nonlinearity of a system can
be properly determined and a suitable model can be developed based on it. The
the odd symmetry of a linear system. By using symmetrical inputs, a system that
addition, the asymmetric behaviour test is similar with the sensitivity analysis study.
Thus, in this work, the result of the asymmetric behaviour is referred to the
existence of the input multiplicity can be observed when several steady state process
inputs generate the same output. In this work, the approach adopted by Zheng et al.
(1998) is used to identify the input multiplicity in the distillation column. In the test,
the reboiler duty is manipulated between 10 and 15MW and the reflux ratio are
56
changed from 5 to 15 in order to study the product composition response. The output
multiplicity is the existence of multiple steady states output from a fixed input and is
categorized under the strong nonlinearity behaviour. The study of the output
multiplicity adopted here is based on the work done by Guttinger et al. (1997) and
Koggersbl et al. (1996). The existence of output multiplicity is verified when a pair
of steady state values overlaps during the distillation operation shift. The operation
shift is done by first increasing the reboiler duty from 0.9 to 1.2 MW for the
operation increasing profile. Then, the reboiler duty is decreased from 1.2 to 0.9 MW
for the decreasing profile. Chokshi and Malik (2006) showed that the multiplicity
study in the distillation process can be done using a process simulator. Thus, in this
work all of these tests are carried out using the Aspen software.
The purpose of data generation is to collect a set of data to describe how the
system behaves for a certain period of time. The idea is to vary the input(s), , of the
system and to observe the response of the output(s), . In order to develop a good
neural network model, the amount and range of data generated must be adequate to
represent the whole operation of the system (Baughman and Liu, 1995). There are
several types of input available such as step, staircase, ramp, generalized multilevel
noise (GMN) and white noise which can be introduced to the system.
In this work, the uniform random noise and multilevel step are used as the
input signals in the data generation process. The uniform random noise is used to
generate uniformly distributed random numbers over a period of time. The data from
the uniform random noise is used for the development of the neural network
modelling and inverse model. Meanwhile, the multilevel step is used to produce data
57
based on several specified levels of input. This type of input is used for setpoint
controller data generation. Since the behaviour of the setpoint controller in the
NLCIMIC is similar to an open loop model based controller, this type of input is
more suitable. In addition, these two types of input signals are generally used for
nonlinear system identification using a neural network (Norgaard et al., 2000; Shaw
et al., 1997). The uniform random noise is generated randomly between 15% from
the nominal reboiler duty value for 2400 hours of simulation time with 24 hours of
change time. The sampling time used in the simulation is 1 minute. On the other
hand, the multilevel step noise is generated for five specified levels of step input for
1800 hours with 60 hours during each step change and 10 minutes of sampling time.
The data from both inputs are generated in two sets, each set for neural network
developed from a state space model based on the closed loop identification. In order
to revisit the concept, the linear model of the CIMIC is initially developed in order to
further understand the capability of the control scheme and later, to be used as a
The identification of the linear model involves introducing the step test into the
open loop distillation system in Aspen Dynamic in order to obtain the input-output
profile. In the step test, a gain of +5% from its nominal MV value is introduced to
the system. The value chosen is enough to cover the linear operation region of the
58
distillation column. Based on the acquired data, a linear transfer function model is
developed using the Matlab System Identification Toolbox which is in the Laplace
domain.
The controller for the linear based CIMIC is developed based on the IMC and
2DOF IMC control scheme. Here, the details for the controller development for the
The development of the CIMIC structure begins with the development of the
IMC. Initially, the linear model developed in Section 3.7.1 is used as a process
model in the IMC control scheme. In order to imitate the practical implementation, a
15% in the process gain and time constant (Shamsuzzoha and Lee, 2008). The IMC
summarized in Figure 3.3. The full IMC scheme is shown in Figure 3.4.
59
Develop the process model
based on transfer function
No
Accepted?
Yes
Finish
From Figure 3.4, the closed loop relationship between the output y(s) and the
setpoint r(s) and disturbance d(s) is given by Brosilow and Joseph (2002):
60
(3.2)
(3.3)
transfer functions that can be adjusted independently (Araki and Taguchi, 2003).
Hence, the 2DOF IMC is introduced to optimize the setpoint tracking and
disturbance rejection problem separately. Therefore, the trade off between the
more efficiently compared to the IMC. The full scheme of the 2DOF IMC is shown
in Figure 3.5. Based on the figure, the feedback controller, , is designed to reject
the disturbance while the setpoint filter, , is designed to shape the response to the
desired MV. The development of the feedback controller and setpoint filter is based
on Brosilow and Joseph (2002) and Morari and Zafiriou (1989) work. Technically,
both of the blocks are inverse model based controllers which are similar to the IMC
controller. However, Brosilow and Joseph (2002) suggested the introduction of the
parameter is used to make the controller more robust towards the model mismatch
error. Based on the 2DOF IMC structure, the perfect model closed loop output and
(3.4)
(3.5)
61
Figure 3.5: 2DOF IMC Structure (Brosilow and Joseph, 2002)
(1996) and the Decoupling and Disturbance Rejection (DRD) scheme from Sandelin
et al. (1991). Based on Figure 3.6, the Linear CIMIC structure is basically the
modification of the 2DOF IMC with the additional inferential loop. The dashed line
represents the process of the distillation column which also incorporates disturbance.
In this scheme, the primary CV loop, y is typically associated with a large time delay
associated with a variable which has a fast sampling time and is easy to measure
such as temperature. Since the primary CV loop response is delayed, any disturbance
that occurs in the system will have a significant impact on the process due to the
delayed controller response. Thus, in order to resolve this matter, the disturbance in
the primary variable is inferred from the performance of the secondary variable.
Besides that, the second variable is normally more sensitive to disturbance and its
62
Figure 3.6: Linear CIMIC control scheme
The purpose of the inverse controller for both the primary and the secondary
loop is to invert the mismatch between the actual process and the model into
appropriate controller effort changes. In addition, for the case of the perfect model,
the mismatch between the process and the model would be explicitly due to the
counter the disturbance effect in the process. The block is particularly used as
primary control variable and the secondary variable is the tray temperature. The time
sampling for the primary variable loop is estimated at 10 minutes while the time
sampling for the secondary variable loop is 1 minute. The reason for the selection of
1 minute as the sampling time for the tray temperature loop is due to the
the simulation calculation would become more extensive and would consume more
63
time to complete. Thus, in order to simplify the simulation process, the 1 minute
3.7.3 Tuning
All the tuning parameters used in the IMC, the 2DOF IMC and the linear
CIMIC are obtained using the IMCTUNE software (Brosilow and Joseph, 2002)
based on the Mp tuning method (Stryczek et al., 2000). However, the tuning method
for the linear CIMIC with the Aspen distillation column model (LCIMIC-AS) is
done by using the heuristic approach with the filter properties proposed by Bequette
(2003). The tuning for the LCIMIC-AS needs to be done separately since the linear
CIMIC and the Aspen distillation process has different properties. The steps
(3.6)
2) The Qsp is first tuned for setpoint tracking performance with the Qy and Qv
with the scheme and tuned for the same test while the Qv loop is still
disconnected.
4) After tuning Qy is completed, the Qv loop is connected and tuned using the
same test. In the end, there are three tuning parameters which are for Qsp,
64
3.8 Development of Nonlinear CIMIC
in Figure 3.7. Based on the figure, the thick line of the block implies that it is
nonlinear. Here, the block represents the Aspen Dynamic distillation column
model, block and are the distillation column nonlinear process models, block
and are the nonlinear inverse model controllers which act as disturbance
compensators and block is the setpoint controller which handles the setpoint
change task.
the system which makes it hard to measure its effect on the process outputs
immediately. Thus, this gives an advantage to the NLCIMIC since it can compensate
the disturbance earlier based on the secondary variable behaviour before it can
65
the primary variable loop can act as the corrective effort to the secondary variable
CIMIC is applying the IMCs inverse model control law, the perfect inverse model is
also necessary. However, the perfect inverse of the nonlinear model is nearly
to develop such good nonlinear models and inverse models, the neural network
modelling is used. In the later part, the neural network technique modelling is further
3.8. Generally, the neural network uses an input-output mapping through a series of
supervised training in order to represent the nonlinear system. Thus, in order for the
neural network to be trained and validated, several sets of data are needed. These
datasets are generated from the Aspen Dynamic simulation which is explained
modelling is when the validation of the neural network fails to meet expectations, the
trouble shooting procedure can go as far as the initial stage of the development.
66
Data Generation
Input/output Selection
Model Structure
Selection
Training
Validation
No
Accepted?
Yes
Neural Network
Model
select the most significant input towards the desired output for a successful neural
network model (Principe, 2006). Furthermore, the selection of the appropriate input-
output can help to preserve the parsimonious behaviour of the network. Thus, the
partly related to the plant model order. The plant model order can be obtained based
on the linearization of the plant. Besides that, it could also avoid using such an
excessive input-output lag structure which would increase the network complexity,
training time and inconsistency of the output. The risk of over parameterization or
over fitting can also be avoided with a proper network size. Hence, the selection of
67
the network input-output can be made from the sensitivity analysis, past experience
or process insight. However, the final decision on selecting the best input-output is
still based on a case-to-case method (Hussain and Kershenbaum, 2000). In this work,
a heuristic approach is taken into consideration with the number of lags in the
control due to its simplicity and ease of development (Irwin et al., 1995). The MLP
network is a fully connected network in layers that consist of input layer, hidden
layer and output layer. Each layer has neurons taking only inputs from neurons in the
previous layer and producing outputs for the next neuron in the next layer. Due to
this forward structure and with updated features (i.e. activation function and
training algorithm), this type of neural network is often referred as the Feedforward
Neural Network (FFNN). According to Hussain (1999), the FFNN is the most used
in the chemical and process control applications which demonstrates its capability to
perform system identification and control for a wide range of dynamic and nonlinear
systems.
closed loop version of the FFNN. In the RNN scheme, the output of the neural
network is supplied back to the input of the network in a feedback loop. Despite the
long training time, in theory the RNN would be able to approximate the arbitrary
dynamical systems with arbitrary precision (Jaeger, 2002). In addition, the RNN is
(Shaw et al., 1997) and time series forecasting (Brezak et al., 2012).
68
In this work, the RNN is used to develop the distillation column model since
theoretically the RNN can perform better as a nonlinear model than the FFNN. On
the other hand, the time delayed FFNN is used for the development of the inverse
Network topology refers to how the neural network is connected between the
hidden units inside the network. In the FFNN, the hidden nodes connect parallel with
each other in one direction only. Meanwhile, for the RNN, there are cycles of
connections between the hidden nodes or a feedback connection from the output
layer into the input layer. Furthermore, the number of input-output to the network
can vary depending on the applications. In this work, the neural network structure
based on the Autoregressive Exogenous with Eternal Input (ARX) is used for the
RNN while the Finite Impulse Response (FIR) structure is used for the FFNN
(Janczak, 2005).
One of the characteristics of the neural network is the multi layer parallel
connection of neurons. This structure serves as the foundation of the neural network
processing capabilities. Hornik et al. (1989) and Cybenko (1989) have shown that all
uniform norms with a one hidden layer neural network. Thus, in this work, the neural
Sarle (2002) and Hagan et al. (1996) had provided several guidelines in order
to choose the best hidden neurons. Sarle (2002) concluded from his work that the
69
selection of hidden neurons is best determined via case by case basis. Furthermore,
using a large dataset for training and early stopping method can help to improve the
neural network generalization. Hagan et al. (1996) suggested that a simple and
smaller size of network for a same case is preferable to avoid over fitting. Besides
that, in this work, the amount of hidden neurons is also determined based on the
holdout method in the cross validation technique. In the holdout method, the data set
is separated into two sets which are called the training and the validation dataset. The
training dataset is used during the network training and the validation dataset is used
afterwards to measure the network performance for unseen data. The Mean Square
nonlinearity response from the network output. This feature would give the neural
network the capability to predict and model the nonlinearity relationship. During
training, the neuron computes the weighted sum of the input signal and compares it
with a threshold value which is determined by the activation function. Whether the
net input is greater, less or equal to the threshold, the network output will be based
In this work, the tangent sigmoid is selected as the activation function. The
tangent sigmoid function differs from the sigmoid function only through a linear
transformation. Thus, the neural network response whose hidden neuron uses the
tangent sigmoid is equivalent to the ones using the sigmoid function despite having
different values for the weights and biases. Empirically, the tangent sigmoid function
70
provides a faster convergence during training than the sigmoid function (Bishop,
1995). The tangent sigmoid function is governed by the equation below and its
(3.7)
Furthermore, the linear activation function is generally used in the output neural
network layer to map the input signal in the range space corresponding to the output
order for the neural network to generalize better. The neural network model is
plays an essential role for the neural network training to be more efficient. Here, the
input data is standardized to 1 and -1 range to suit the tangent sigmoid activation
function used and to avoid the data dissimilarity influence. The data dissimilarity
influence can happen when larger input variables are trained together with smaller
input variables. This inequality of magnitude in the neural network input would
71
affect the weight training. In addition, the neural network can be trained more
effectively based on a standard value rather than using the real data.
structure shown in Figure 3.10. Based on the figure, the neural network model is
trained based on the error of the plant (or process) output with the model output
itself. This error signal will be used as the training signal in the neural network
learning algorithm with the objective to minimize this error. This learning structure
is a classical supervised learning problem where the teacher (i.e. the process) teaches
the learner (i.e. the network) how to achieve the target (i.e. process output) (Jordan
Yp
U
Plant
+
-
Neural Network
Model
Ym
Error
Learning
Algorithm
training with the target and then uses this information to systematically modify the
weights throughout the neural network in a backward manner. Generally, there are
72
network training. In the simplest implementation of the backpropagation algorithm,
the network weights and biases are updated in the direction in which the
performance function decreases most rapidly, which is the negative of the gradient.
(3.8)
where is a vector of current weights and biases, is the current derivative error
gradient and is the learning rate. However, this simple backpropagation algorithm
is not suitable for practical problems due to its slow training process. Hence, a faster
backpropagation algorithm has been developed to improve the training process and
algorithm for the neural network. The LM algorithm is developed to approach the
second-order training speed without having to compute the Hessian matrix. The LM
(3.9)
with
(3.10)
(3.11)
where is the Jacobian matrix that contains the first derivatives of the network errors
with respect to the weights and biases and is a vector of network errors. When the
73
scalar is zero, Equation 3.9 becomes the Newton's method with the Hessian matrix
approximation. When is large, this equation becomes the gradient descent method
with a small step size. Since the Newton's method is faster and more accurate near an
error minimum, the aim is to shift the equation towards the Newton's method as
performance function) and is increased only when a tentative step would increase the
In addition, the early stopping method is used in the neural network training
to avoid network over fitting. In order to validate the neural network model, another
set of unseen data is used to test the network performance. The neural network
model with the best MSE and R2 from the validation test is selected.
Here, the method for training the neural network as an inverse model
controller is presented. Generally, there are two ways to train the neural network as
an inverse model. The first method is known as the direct method as shown in Figure
3.11. Here, a training signal is introduced to the system. Then, the system output will
be used as input to the network. Afterwards, the output of the network is then
compared with the training signal (system input) and the network will be trained
based on this error. This structure is intended to force the network to generalize the
plant inverse. The second method is called specialized inverse learning. In this
approach, the network inverse model preceded the system and receives the input of
training signals which spans the desired operational output range of the controlled
system. The system can also be replaced with a neural network model if the real
74
system is not available. The inverse model in this scheme is trained based on the
u y umodel +
Inverse Neural -
System
Network Model
Errormodel
Figure 3.11: Neural network inverse using direct method (Norgaard et al., 2000)
In this work, the direct method architecture is selected to develop the neural
network inverse. Although the specialized training is more sensible and goal
directed, the scheme can only be realized using on-line training. In contrast, the
direct method technique is more straight-forward and the training can be done
offline.
shown in Figure 3.12. Based on the figure, the dark line represents the slow sampling
loop (10 minutes time) for the primary control variable while the light grey is the fast
sampling loop (1 minute) for the secondary inferential variable. The DR Y and DR V
whereas the SP controller shapes the reference signal into the desired controller
in the NLCIMIC. The application of the tuning parameter for the neural network
control scheme has been done previously by many researchers (Biyanto et al., 2010;
Hunt and Sbarbaro, 1991; Hussain and Kershenbaum, 1999; Shaw et al., 1997). In
75
this study, the tuning parameter is used as the low-pass filter to handle the inverse
neural network controllers sensitivity and robustness and the selection of the tuning
The performance test is done as the evaluation for the control scheme
performance and reliability. It is important to assess and evaluate the control scheme
in order to know its strengths and limitations. Thus, it serves as the benchmark in
In the neural network modelling, the criteria that are used to evaluate the
network performance are the Mean Square Error (MSE) and Coefficient of
Determination (R2). The MSE is used to quantify the performance error between the
neural network new output and its target. In addition, R2 is a measurement on how
the disparity and likeness of the new network outcome from the target values.
76
Furthermore, in order to evaluate the performance of the control schemes used in this
when a new specification of product is needed. In this test, a 98% purity setpoint
change is introduced from the nominal value at t=10h and followed by a 90% purity
step down at t=100h. Both of the setpoint changes occur for 50 hours and are reset to
nominal value after each step test is completed. The performance of the control
respond and reject appropriately the disturbance effect. The disturbance effect can be
defined as the behaviour of the control variable under the influence of the
disturbance variable. In the distillation column, there are two significant disturbances
that occur at the feed stream, which is the feed composition and feed flowrate. The
feed composition upset would shift the composition profile through the column
which will result in a significant upset in the product purity. Since many of the
distillation columns do not have any composition analyzer at the feed stream, the
flowrate can become a disturbance if the feed from the upstream process is not
77
In this test, the disturbances are introduced after the process has reached
steady state at upper setpoint (98% purity). During the process, the disturbances are
employed at t=25h for a 5% increase in feed flowrate and at t=47h for a 10%
decrease of n-butane overall purity in the feed stream. Both disturbances lasted for 2
78
CHAPTER 4
based on the specifications and conditions given in Ilme et al. (2001) and Klemola
and Ilme (1996). The distillate product of the column is i-butane at 93.5% purity and
the bottom product is n-butane at 98.1% purity. The distillation column simulation is
In order to validate the steady state distillation column model, the results
from Aspen Plus is compared with the actual data taken from the original reference
(Ilme et al., 2001; Klemola and Ilme, 1996). The stream validation results are shown
in Table 4.1 and the tray temperature validation is shown in Figure 4.1. Based on
Table 4.1, all the streams results obtained from the Aspen Model almost match the
data from the actual column. The error for trace components such as Isobutene and
Neopentane can be neglected due to its small fraction in the process. As for the tray
temperature in Figure 4.1, the actual and simulated results differ at an acceptable
margin of below 4% of error. Based on these validations, the steady state distillation
79
Table 4.1: Distillation column model validation results
Distillate Bottom
Components Error Error
Actual* Simulation Actual* Simulation
% %
Propane wt% 0.0494 0.0495 -0.202 0.000 0.000 0.000
i-butane wt% 0.942 0.942 0.000 0.003 0.003 0.000
n-butane wt% 0.00200 0.00168 16.0 0.981 0.981 0.000
Isobutene wt% 0.00230 0.00311 -35.2 0.0008 0.00048 40.0
1-butene wt% 0.00410 0.00412 -0.488 0.00100 0.00104 -4.00
Neopentane wt% 0.000 0.000 0.000 0.00170 0.00160 5.88
Isopentane wt% 0.000 0.000 0.000 0.0112 0.0112 0.000
n-Pentane wt% 0.000 0.000 0.000 0.00110 0.00116 -5.455
Total Flowrate
8123 8123 0 17999 17999 0
(kg/h)
*As reported in Ilme et al. (2001)
The dynamic model is developed based on the validated steady state model of
the distillation column. The development of the dynamic modelling using Aspen
Dynamic is shown in Figure 4.2. Based on the figure, it can be observed that there
are three control scheme lines implemented in the dynamic distillation column. The
control lines represent the reflux tank level control, top pressure control and
80
distillation bottom level control. These control schemes are essential to operate a
Based on Neste Oil (2013) sales report, the n-butane is sold in the market at
98% and 90% purity. Hence, in order to comply with the market demand and to
evaluate the controller performance, the current steady state condition of the
distillation column is changed to a new value that lies at the middle of the desired
output. The current and proposed value is tabulated in Table 4.2. By using the new
proposed value of the reboiler duty, the product purity is now at 94%, which is in the
Table 4.2: Comparison of current and proposed value for the distillation column
steady state
Current Value Proposed Value
Reboiler Duty (KW) 10240.0 9225.0
n-butane (kg/kg) 0.9804 0.9400
81
4.1.3 Distillation Control Configuration
Besides basic control loops, the distillation column model is also required to
be equipped with the composition control configuration in order to obtain the desired
product purity. The control configuration used in this distillation column model is
the (L/D,V) configuration where the top composition is maintained by the reflux
ratio and the bottom composition is controlled by the stream vapour flowrate. Since
the steam vapour variable is not available in the Aspen Dynamic, the reboiler duty is
selected as the replacement since the energy from the reboiler is correlated with the
amount of the steam input. This control configuration is suitable for a high reflux
ratio column and when the bottom product is more important than the distillate
(Riggs, 2001). In this work, the NLCIMIC scheme is used to regulate the reboiler
In addition, manipulating the reboiler duty can be a practical option since the
short time due to the fast change in the vapour boil up rates (Riggs, 2001). Thus, any
control action taken by the reboiler duty will have an immediate effect on the
distillation system. In contrast, the reflux ratio control is maintained at its steady
In this work, the tray temperature selection is made based on the Singular
Value Decomposition (SVD) analysis and the result is shown in Figure 4.3. Based on
Figure 4.3, the highest Usvd magnitudes for the reboiler duty and the reflux ratio are
located at tray 29 and tray 61 respectively. These imply that both the trays are the
82
which has the second highest Usvd value for the reboiler duty is used instead of tray
29 because in general practice, the tray temperature used to infer or control the
dynamically correlated with each other during constant pressure (Hoffman et al.,
several degrees of step tests to excite the distillation process. Based on the step test
can be made. In this study, the effect on the bottom product composition (n-butane)
and temperature tray 68 is evaluated based on changes in the reboiler duty, feed
83
4.3.1 Effect of Reboiler Duty
The reboiler step test result for n-butane purity (XB) and tray 68 temperature
(T68) is shown in Figure 4.4. Based on the results, the negative step tests in XB
produce larger deviations than the positive step tests. Since the maximum bottom
composition of the distillation column model design is only around 98.4%, any
positive gain approaching this value will become saturated. On the contrary, a
steps tests rather than negative step tests. This is due to the effect of increasing
vapour boiling up rate in the vapour-liquid transfer in the column which is much
more significant in the higher purity region (i.e. the positive step). The distillation
column system is known to be more sensitive and nonlinear at a higher purity region
(Fuentes and Luyben, 1983). From an overall view, both XB and T68 share a similar
trend in the positive and negative step test. This signifies a correlation between XB
and T68 output which can justify the application of the inferential technique between
Figure 4.4: Step test results for bottom product purity (left) and temperature Tray 68
(right) by manipulating the reboiler duty at 5%, 10% and 15% change from the
nominal condition
84
4.3.2 Effect of Feed Flowrate
The effect of feed flowrate variation on the XB and T68 profile can be seen
in Figure 4.5. The effect of feed flowrate is most significant at 10% change for
both cases. Based on the figure, the effects of the feed flowrate change produce an
opposite result on XB and T68. This is due to an increase in the feed flowrate which
will lead to the increment of the columns material balance. Thus, as the distillate
and bottom flowrate rise due to this additional material source, the product
composition will drop. The column temperature will also decrease as the vapour-
liquid volume transfer in the column is increased. Based on the figure, it can be
observed that feed flowrate parameter has a significant influence on both XB and
Figure 4.5: Step test results for bottom product purity (left) and Tray 68 temperature
(right) by manipulating feed flowrate at 3%, 5% and 10% change from the
nominal condition
The effects of disturbance in the feed composition for XB and T68 are shown
in Figure 4.6. The figure shows that the increase of the n-butane amount in the feed
stream would favour the increase in the bottom product composition and vice versa.
85
Since the n-butane is a heavy key component, the additional availability of this
component in the rectifying section would shift the VLE profile and thus influence
the columns temperature as well. In this case, the T68 profile trend is similar with
the XB profile trend which indicates the strong relationship between the T68 and XB
Figure 4.6: Step test results for bottom composition (left) and Tray 68 temperature
(right) by manipulating the n-butane feed composition at 5%, 10% and 15%
change from the nominal condition
Based on the results in the sensitivity study for the reboiler duty, feed
flowrate and feed composition, the XB and T68 parameters have shown asymmetric
response profiles in all the figures. This observation shows that the distillation
86
certain benchmark, the nonlinearity behaviour of the system can be further explored
generalize an overview of the distillation column behaviour, the reboiler duty (QB)
and reflux ratio (RR) is manipulated by a certain range to measure the response in
the bottom product purity (n-butane). The result achieved is shown in Figure 4.7.
Based on the figure, it is obvious that the nonlinearity behaviour occurs when all the
parameters are simulated together as the plane surface is not flat at certain regions.
Similar observation can be found in Fox and Stine (2001) and Ma et al. (2010) work.
In order to further clarify the nonlinear behaviour, the parameters are studied in
Figure 4.7: Effect of reflux ratio and reboiler duty variations towards bottom product
purity
87
4.4.1 Asymmetric Response
In the asymmetric response test, the system is tested with a specified range of
step inputs to observe the systems output. This test is actually similar to the
sensitivity analysis test. Thus, based on the remarks from the sensitivity analysis test,
The input multiplicity is the existence of a set of steady state input that
generates a same output response. In this process, the existence of this behaviour is
shown in the small circles in Figure 4.8 and Figure 4.9, respectively. In the Figure
4.8, the reflux ratio of 6.5 and 10.7 produced the same response i.e. 98% of n-butane
purity for the bottom stream. In Figure 4.9, 98.3% of n-butane purity at the bottom
stream is produced when the reboiler duty applied are 10950KW and 15000KW. The
reason for the decrease of n-butane composition in the figures is because when the
reboiler duty and reflux flowrate is increased, the lighter components in the bottom
section at the column decreases. As the inventory of the lighter components becomes
would even remove the n-butane component from the bottom stream. As a result, the
88
Figure 4.8: Effect of reflux ratio variations towards the bottom product (n-butane)
purity. The input multiplicity occurrences are located inside the drawn circle.
Figure 4.9: Effect of reboiler duty variations towards the bottom product (n-butane)
purity. The input multiplicity occurrences are located inside the drawn circle.
from a single process input. In this work, the effect of output multiplicity occurs at
the point shown in an oval as in Figure 4.10. Based on the observation, the reboiler
duty values from increasing and decreasing operation lines are matched with each
89
other except at 9900 KW, 10100 KW and 10300 KW. At these three points, there are
two values of MV that can produce the same product composition. Apparently, in
order to achieve the specified output concentration, two values of inputs can be used.
However, in this case, the discrepancy between the two points is very close which
behaviour.
Figure 4.10: Effect of reboiler duty using two different operating lines towards the
bottom product (n-butane) purity. The output multiplicity occurrences are located
inside the drawn oval shape.
Based on the study mentioned earlier, the distillation column has been
According to Pearson (2003), the availability of such attributes signify mild degree
90
4.5 Data Generation Results
The data needed for the neural network modelling is generated using the
butane composition and temperature tray 68 variation profiles are shown in Figure
4.11, Figure 4.12, and Figure 4.13, respectively. The data for the neural network
91
Figure 4.13: Tray 68 temperature profile
In this section, the results for the linear based CIMIC development and its
original schemes, the IMC and the 2DOF IMC are shown. These linear controllers
In order to develop a linear model, the data from the small gain step test is
used to generalize the system dynamic. The input and output results from the +5%
step test are used and the distillation column is assumed to behave similar to a first or
second order process. It can be observed that the positive and negative 5% step test
results from the section 4.3.1 have shown a symmetric linear response throughout the
process. Thus, it indicates that the process has behaved linearly at the specified
region. The transfer function model is developed using the Matlab System
Identification Toolbox. The results for the process identification are tabulated in
Table 4.3. The letter P represents the pole, Z is the zero and D is the delay in each
92
Based on the table, it can be observed that model P2Z produces the best
response for both the distillation column composition model (XB) and the
temperature model (T68) based on highest best fits values. The best fit equation is
similar to the coefficient of determination (R2) equation where the highest value
indicates that the simulated response is more similar to the original response. The
term P2Z shows that the transfer function model has two poles and single zero.
Based on the system identification results, the transfer function model for the bottom
composition XB model is P2Z (referred after this as Model Y) and P2Z for tray
temperature T68 model is P2Z (referred after this as Model V). Both of these transfer
function model parameters are shown in Table 4.3 and used in the linear controller
design.
Table 4.3 System identification results for Model XB and Model T68
Model XB Model T68
Models Best Fits % Models Best Fits %
P2Z 97.88 P2Z 97.7
P1D 93.05 P1Z 97.61
P1 93.05 P1D 87.40
P2 91.47 P2 83.93
(4.1)
(4.2)
93
4.6.2 Controller Scheme Design
function model with a modified 2DOF IMC design. Here, the results for all the
IMC Controller
IMC Controller:
(4.3)
2DOF IMC
(4.4)
Linear CIMIC
(4.5)
94
4.6.3 Performance Evaluation
tracking and disturbance rejection test. The IAE analysis is used as a tool to quantify
The setpoint tracking test overall result is shown in Figure 4.14. Based on
figure, the Linear CIMIC (LCIMIC) performed better than the 2DOF IMC and the
IMC in terms of a faster settling time. Moreover, during the step up test at t=5 hours,
the LCIMIC showed an overdamped response compared with the 2DOF and the IMC
slight overshoot behaviour as shown in Figure 4.15. This observation is in line with
the IAE value obtained where the LCIMIC gives the smallest value (IAE =0.0689)
followed by the 2DOF IMC (IAE = 0.1018) and the IMC (IAE = 0.1037). The
controller output for the setpoint tracking test can be seen in Figure 4.16. From the
figure, it can be observed that all the controllers produced a similar response.
However, from a closer observation in Figure 4.17, it can be seen that the LCIMIC
had exerted a faster controller action to track the appropriate MV value if compared
to the 2DOF IMC and the IMC. Since the reboiler duty energy consumption is large,
95
Figure 4.14: Setpoint tracking response results for LCIMIC, 2DOF IMC and IMC
Figure 4.15: Setpoint tracking result for a step up change at t=5 hours
96
Figure 4.16: Setpoint tracking MV profile results for LCIMIC, 2DOF IMC and IMC
The overall performance of the LCIMIC, the 2DOF IMC and the IMC for
setpoint tracking test is predicted to be almost the same since the controller has the
same setpoint controller parameters. However, the additional inferential loop in the
LCIMIC control scheme has led to a faster response. This is obvious because the
Hence, the LCIMIC controller can act faster based on the inferential response.
97
4.6.3 (b) Input Disturbance
perturbation in the MV due to certain situations. The value selected for the
disturbance is equivalent to 10% of the process input. In this work, the disturbance is
introduced at t=1 hours for 5 hours as shown in Figure 4.18. Based on the figure, the
Linear CIMIC scheme displays its profound ability to effectively reject the
disturbance compared to the other control schemes. The reason for the poor
performance for the IMC and the 2DOF IMC is because the effect of input
disturbance is not considered by the controllers. Since the IMC and the 2DOF IMC
only perform based on model mismatch, the effect of input disturbance is not
explicitly included in the mismatch. Thus, the controller is unable to produce the
correct gain to compensate the input disturbance (Wassick and Tummala, 1989).
Since the linear CIMIC has another loop to infer the disturbance from the tray
temperature profile, the existence of the disturbance at the beginning of the process
can be rejected. In order to reject the input disturbance promptly, the Linear CIMIC
controller has to exert a sudden extent of effort if compared to the other controllers
98
Figure 4.18: Input disturbance rejection results for CIMIC, 2DOF IMC and IMC
Figure 4.19: Input disturbance rejection MV profile for CIMIC, 2DOF IMC and IMC
the perturbation that occurs inside the process such as tray condition, sudden drop in
column pressure and equipment failure which can affect the process response
directly or indirectly. Thus, a step disturbance of 10% from the nominal value of the
99
process output stream is introduced at t=1 hours for 14 hours as shown in Figure
4.20. Based on the figure, all the control schemes are able to equally reject the output
disturbance using almost the same magnitude of the reboiler duty as shown in Figure
4.21. However, the controller needs at least 14 hours to fully reject the disturbance.
Since n-butane/i-butane distillation process has a slow process gain, thus, such
anticipated since such disturbances can lead to an additional model mismatch in the
control scheme. If a perfect process model is available, then such mismatch is solely
due to the disturbance in the process. Based on the quantitative error test, the
LCIMIC produced a better performance with the IAE of 0.0346 compared with the
2DOF IMC (IAE = 0.0364) and IMC (IAE = 0.0370). Thus, the addition of the
inferential loop in the Linear CIMIC has again helped the controller to properly
reject the disturbance. The 2DOF IMC is found to perform slightly better than the
IMC since the disturbance is compensated using a different controller. Unlike the
IMC scheme, both the setpoint tracking and the disturbance rejection case are
100
Figure 4.20: Output Disturbance rejection results for Linear CIMIC, 2DOF IMC and
IMC
Figure 4.21: MV profile for Linear CIMIC, 2DOF IMC and IMC output disturbance
rejection test
A summary of the results from the setpoint tracking, input disturbance and
output disturbance test is tabulated in Table 4.4. It can be observed that the LCIMIC
101
produced a better performance than the 2DOF IMC and the IMC in all the tests
conducted.
Table 4.4: Summary results for Linear CIMIC, 2DOF IMC and IMC performance
IAE
Control
Scheme Step Input Output
Test Disturbance Disturbance
Linear CIMIC 0.0689 0.0080 0.0497
column model would serve as the nonlinear system compared to the linear model
used earlier in the test. In order to integrate it with the Aspen distillation model, the
LCIMIC-AS has to be retuned. This is because the tuning parameter used in the
LCIMIC earlier is based on a linear model. Thus, when using a nonlinear process
the nonlinear process model. The tuning parameter selected for the LCMIC-AS is
LCIMIC. In the setpoint tracking test as shown in Figure 4.22, the LCIMIC shows a
better tracking capability (IAE = 0.0697) with shorter rise time and less overshoot
when compared with LCIMIC-AS (IAE = 0.2009). However, for the step down
profile, the LCIMIC-AS produced a larger deviation than the LCIMIC. This can be
102
explained by the sudden controller action taken by the LCIMIC-AS as shown in
Figure 4.23. This controller action performance is typical for any linear controller.
feed flowrate in the LCIMIC-AS scheme. For the LCIMIC, the disturbance is
4.24, all the disturbances occur at t=2 and last for 4 hours. Based on the figure, the
LCIMIC showed a faster disturbance rejection capability (IAE = 0.0076) than the
LCIMIC-AS (IAE = 0.0098). This can be linked to the MV profile of the disturbance
rejection test as shown in Figure 4.25. Figure 4.25 shows that the LCIMIC produced
a fast and sudden controller action to reject the disturbance, compared to the
LCIMIC-AS that shows a slow controller response. However, both of them have
103
Figure 4.22: Setpoint tracking CV profile results for LCIMIC-AS and LCIMIC
Figure 4.23: Setpoint tracking MV profile results for LCIMIC-AS and LCIMIC
104
Figure: 4.24: Disturbance rejection results for LCIMIC-AS and LCIMIC
Figure 4.25: Disturbance rejection MV profile results for LCIMIC-AS and LCIMIC
A summary of the comparison test results are tabulated in Table 4.5. Based
on the overall performance of the comparison test, the LCIMIC outperformed the
LCIMIC-AS. Since the LCIMIC is based entirely on transfer function model and
controller, the good result is expected. However, for the LCIMIC-AS, the linear
controller is used to control the distillation that is represented by the nonlinear model
105
simulated in Aspen which caused the deterioration of the control performance. This
In this section, the results for the Nonlinear CIMIC development are
presented. The Nonlinear CIMIC design scheme consists of neural network models
of the distillation column and the neural network inverse models for setpoint tracking
and disturbance rejection. The development results for each of the neural network
model used here is presented. In the final stage, all the neural network models are
combined to construct the Nonlinear CIMIC scheme and simulated with Aspen
(Model Y) and tray temperature (Model V). Model Y is developed to simulate the
bottom composition response with a slow sampling time. Model V is the tray 68
temperature response model with a fast sampling time which is used in the
inferential scheme.
106
4.7.1 (a) Input/output Scheme
The neural network input and output scheme is shown in Table 4.6 for Model
Y and Table 4.7 for Model V. The number of lags for Model Y and Model V is
The full validation results for the neural network model are shown in Table
4.8 for Model Y and Table 4.9 for Model V. In the both tables mentioned above, the
number of hidden neurons used in the each model is in bold. In general, the best
hidden neuron for any model is selected based on the lowest validation MSE.
However, some guidelines as mentioned in subchapter 3.8.1(d) are also taken into
107
consideration. Therefore, although in Table 4.8, hidden neuron 10 produce the
lowest MSE value, hidden neurons 4 until 9 also show similar comparable results.
Thus, since the results are so close, hidden neuron 4 is selected since it has the
smallest number of hidden neurons. In a similar manner, in Table 4.9, hidden neuron
6 is selected. In order to visualize the models performance, the validation result for
each model is plotted in Figure 4.26 and Figure 4.27. The accuracy of the target and
value. Based on Figure 4.26 and Figure 4.27, both models clearly give very good
Model Y shows a slightly more accurate response with R2=0.9998 when compared
2 6.54E-07 0.99957
3 3.96E-07 0.99974
4 2.78E-07 0.99982
5 2.66E-07 0.99983
6 2.63E-07 0.99983
7 2.56E-07 0.99983
8 2.70E-07 0.99982
9 2.53E-07 0.99983
10 2.48E-07 0.99984
108
R2=0.9998
2 0.004644 0.997638
3 0.001565 0.999205
4 0.003923 0.998008
5 0.004107 0.997915
6 0.001097 0.999444
7 0.000813 0.999587
109
R2=0.9994
inverse model to Model Y. Based on the CIMIC scheme, the setpoint controller role
is similar to the open-loop model based controller which drives the setpoint reference
The input/output scheme for the setpoint controller is shown in Table 4.10.
Input Output
110
4.7.2 (b) Training and Validation
The results for the neural network validation for the setpoint controller are
shown in Table 4.11. Based on Table 4.11, the lowest validation error is at 6 hidden
neurons. In Figure 4.28, the selected model is simulated to visualize its response and
the R2 is 0.9974. Based on the figure, the spikes that occurred during each step
changes are the result of the neural network over fitting behavior and poor
generalization. This is due to the lack of input and variation of data that have been
used to developed the inverse model. However, this performance is still acceptable
2 5260.566 0.996067
3 5609.572 0.99607
4 5399.61 0.996086
5 4362.946 0.996538
6 3450.514 0.997461
7 5415.757 0.996261
8 7751.012 0.995207
9 5144.077 0.996626
10 8991.557 0.994126
111
R2=0.9974
The neural network is trained into the inverse model of Model Y and Model
compensator is to compensate the model mismatch of the process, given that the
112
Table 4.13: Input/output for neural network disturbance compensator V
Input Output
The results for the neural network disturbance compensator Y and V are
shown in Table 4.14 and Table 4.15, respectively. The selected models, which are
the ones with the lowest MSE, are highlighted in the tables. Furthermore, the model
response profiles are also plotted in Figure 4.29 and Figure 4.30. The R2 values for
both models are 0.9916 and 0.9918, respectively. Based on Figure 4.29 and Figure
4.30, it has been observed that the spikes that occurred are similar to ones in Figure
4.28. Thus, it shows that both neural network inverse models also suffered from
same problem. In this case, the inverse models are still acceptable since the R2 values
2 13078.21 14112.69
3 10653.93 12135.83
4 10594.81 11615.12
5 10591.67 11675.11
6 9897.869 10340.1
7 9850.946 10042.04
113
8 10025.02 10206.29
9 9879.597 10208.66
10 10079.68 10687.36
11 9862.656 10239.01
12 9967.157 10328.19
13 9708.707 9998.66
14 9722.862 10029.01
15 9868.571 10354.17
R2=0.9916
2 14012.86 15848.49
3 11487.14 12352.88
4 9557.791 10522.62
114
5 10144.55 10916.83
6 698918.5 720102.2
7 9318.82 10208.53
8 8768.691 9890.006
9 10075.51 11020.8
10 14026.6 15519.35
11 9888.813 10571.83
12 9854.003 10724.1
13 9445.498 9730.953
14 9812.606 10504.02
15 12747.13 14309.73
R2=0.9918
In this last section, the performance of the Linear (LCIMIC-AS) and the
Nonlinear CIMIC (NLCIMIC) are compared and discussed. Here, the Aspen
115
distillation column model is used to represent the real process for both schemes. In
addition, both schemes are tested in advance in order to find the best tuning
parameter. The tuning parameters used in both schemes are shown in Table 4.16.
Table 4.16: The tuning parameter used in the NLCIMIC and LCIMIC-AS
Tuning parameter
Control Setpoint Disturbance Disturbance
Scheme Controller Compensator Y Compensator V
The results for the overall setpoint tracking test is shown in Figure 4.31 and
Figure 4.32, respectively. From a general observation, it can be seen that the
the overall setpoint tracking test is divided into two parts which are the step up and
116
Figure 4.31: Overall setpoint tracking test results for NLCIMIC and LCIMIC-AS
Based on the step up test results in Figure 4.33, the NLCIMIC performed
better than the LCIMIC-AS with a faster trajectory response and a shorter setting
time with IAE = 0.0349. In contrast, the LCIMIC-AS performance (with IAE =
117
0.0636) is slightly oscillated towards reaching the new setpoint. This can be further
justified based on the MV profile for the step up test which is shown in Figure 4.34.
Based on the figure, the fast and stable controller action helped the NLCIMIC to
reach the new setpoint within a shorter period of time. In contrast, the oscillatory
controller effort from the LCIMIC signified the controllers slow compensate action.
Figure 4.33: The step up test response for NLCIMIC and LCIMIC-AS
Figure 4.34: The MV profile for NLCIMIC and LCIMIC-AS in the step up test
118
Figure 4.35 shows the step down test response for both controllers. From the
figure, it is observed that the LCIMIC-AS produced a faster and less overshoot
response if compared to the NLCIMIC. This observation can be verified via the IAE
analysis which indicates that the LCIMIC-AS performance (IAE = 0.0399) is better
than the NLCIMIC performance (IAE = 0.0402). The reason for such NLCIMIC
performance can be further evaluated from the MV profile of the step down test as
shown in Figure 4.36. Based on the figure, the controller effort from the LCIMIC-AS
is much more stable without much oscillation if compared to the NLCIMIC even
though both controllers have showed a nearly similar MV profile. This is might be
due to the imperfectness of the neural network inverse controller in the NLCIMIC in
tracking the lower setpoint. However, it should be noted that the error difference is
relatively small and NLCIMIC response has settled earlier than LCIMIC-AS.
119
Figure 4.35: The step down test response for NLCIMIC and LCIMIC-AS
Figure 4.36: The MV profile for NLCIMIC and LCIMIC-AS in the step up test
Based on Figure 4.37, it can be observed that the NLCIMIC was able to
reject both sources of disturbances more effectively than the LCIMIC-AS. Moreover,
the error analysis also showed that the NLCIMIC (IAE = 0.0107) outperformed the
120
LCIMIC (IAE = 0.0154). In the feed flowrate disturbance, the NLCIMIC
compensated the disturbance faster than LCIMIC-AS. For the disturbance in the feed
composition, both controllers responded nearly the same during the occurrence of the
responded with a long tail response and the LCIMIC-AS behaved with oscillatory
decay. Based on such responses, the LCIMIC reached the setpoint slightly earlier
than the NLCIMIC. Nonetheless, based on the MV profile in Figure 4.38, the
NLCIMIC response was preferred where much less energy was required to reject the
121
Figure 4.37: Disturbance rejection test response profile for the NLCIMIC and the
LCIMIC-AS
Figure 4.38: MV profile for the disturbance rejection test between the NLCIMIC and
the LCIMIC-AS
All the performance results in the comparison study are tabulated in Table
4.17. Based on the table, most of the performance results are showing the advantage
of the NLCIMIC over LCIMIC-AS. This outcome is expected since the benefits of
122
the nonlinear model and controller application in handling nonlinear process
behaviour is commonly known. However, there were some issues regarding the
performance of the nonlinear controller using neural network model that need to be
addressed.
setpoint controller to handle the setpoint tracking problem. Thus, the imperfectness
of neural network inverse in the setpoint controller will influence directly the
NLCIMIC setpoint tracking ability. In this matter, the significant error in the
NLCIMIC overall setpoint tracking result is due to the occurrence of the bump at t
= 60h as seen in Figure 4.31. The bump is caused by the sudden slump followed by
Figure 4.32. This irregular behaviour is caused by the inverse neural network model
response which can be verified via Figure 4.28. Based on Figure 4.28, the noises on
the setpoint controller profile signify the inverse neural network sturdy effort to
represent the target profile. Such behaviour can happen when the neural network
model has poor generalization due to many reasons such as lack of sufficient training
data, over fitting, improper training method and not enough input (Vemuri, 1993).
large training data to generalize properly. In addition, the data also need to cover the
whole operating region and enough to represent the problem. In the NLCIMIC
setpoint controller case, there is a possibility that the inverse neural network model
failed to generalize properly due to insufficient size of training data. Moreover, the
selection of training method can also affect the inverse neural network
generalization. In this work, the general training method is used to train the inverse
123
implemented, there are some issues that can influence the neural network output
performance such as sensitive to noise, poor robustness, large output signal and
possibility of unstable inverse (Norgaard et al., 2000). This is due to the fact that the
objective of the general training is to minimize the difference between the actual and
estimated output signal rather than goal directed. Moreover, the selection of neural
network input is also important to avoid the over fitting problem. Through a
selection of significant inputs, the neural network will be able to recognize and learn
the data pattern more properly. Generally, relaying on a single input maybe more
convenient and practical. However, there are some possibilities that the neural
developed similarly to the NLCIMIC setpoint controller. However, the training data
compensator has shown superior performance than the LCIMIC-AS. In the Figure
4.37 and Figure 4.38, the NLCIMIC was able to reject disturbance more quickly and
compensators (refer Figure 4.29 and Figure 4.30) shows similar observation of spike
noises (as in Figure 4.28), the final controller performance was good. Thus, it can be
assumed that the poor generalization of the inverse neural network model only
affecting certain region only. This assumption can be further proven with the
NLCIMIC different performance in tracking the high and lower purity setpoint.
Theoretically, the effectiveness of the IMC (even the CIMIC) control scheme
is mostly depends on how accurate the process model and the inverse model
controllers are developed. If a perfect process model can be developed and the
inverse of such a process model is available, then total control of the process can be
124
achieved. The linear modelling technique such as the transfer function model is
commonly used in the IMC scheme due its simplicity and straightforward design.
One of the advantages of using the transfer function model is the dynamic of the
response can be explicitly manipulated using a filter. In addition, the perfect inverse
model can also be realized by just inverting the transfer function model. Thus, with
the perfect inverse model controller and robust filter tuning, the LCIMIC-AS is
However, the linearization of the nonlinear process has its limitations since
there are various nonlinear processes that have certain dynamic behaviours that
cannot be modelled linearly. Thus, this would certainly affect the performance of the
control scheme since the controller is developed directly from the process model. In
addition, Pearson (2003) has demonstrated that neural network model has the ability
issue.
IAE
Control
Scheme Step Test Step Up Step Down Disturbance
(Overall) test test Rejection test
125
CHAPTER 5
5.1 Conclusions
A model based control (MBC) strategy, the CIMIC, has been developed to
control a distillation process. The CIMIC has the advantage of handling a slow
capability. In order to further enhance the ability of the CIMIC, the CIMIC control
scheme is embedded with a nonlinear model and controller. Thus, in this work the
In this study, the steady state and dynamic model of the n-butane/i-butane
distillation column was developed using Aspen software. In addition, the steady state
model was successfully validated based on data from the literature (Ilme et al.,
multicomponent feed stream, low relative volatility and slow dynamics. Thus, in
order to explore and verify the nonlinearity of the distillation column process, a set
have mild degree of nonlinearity with the occurrence of the input multiplicity
behaviour.
The linear MBC control strategy based on the transfer function i.e. IMC,
2DOF IMC and linear based CIMIC (LCIMIC) was also developed to further
126
the comparison tests made between the IMC, 2DOF IMC and LCIMIC, the LCIMIC
outperformed the rest in all areas especially in rejecting input disturbance. However,
when the LCIMIC was tested with an Aspen based nonlinear distillation model
develop the nonlinear process models and controllers. Generally, the neural network
is known for its ability in modelling and generalization of a nonlinear system. Since
the MBC performance mainly depends on the availability of a good process model,
the selection of neural network modelling is justified. In this work, the distillation
column model was developed based on the Recurrent Neural Network (RNN) while
the controller part was developed using the Feedforward Neural Network (FFNN)
inverse modelling technique. Since two control loops were used in the NLCIMIC
scheme, two process models and disturbance compensator controllers with different
In the final part, the NLCIMIC was evaluated and compared with the
was made based on the controller schemes performance in setpoint tracking and
stream with IAE = 0.0107 compared to LCIMIC-AS that produced IAE = 0.0154.
Moreover, the controller output produced by NLCIMIC was far more efficient in
LCIMIC-AS. On the other hand, the setpoint tracking for the NLCIMIC and the
127
98%), the NLCIMIC performed better with IAE = 0.0349. For the step down
setpoint (n-butane purity = 90%), the LCIMIC-AS had a better tracking ability with
IAE = 0.0399.
good control scheme with a straightforward controller design and fast disturbance
nonlinear process behaviour, the NLCIMIC is introduced. From the final comparison
test, it is proven that NLCIMIC has the advantage of handling nonlinear process than
5.2 Recommendations
The following are some recommendations that can be implemented for future work:
1. The distillation column used in this work has a slow dynamic and low separation
factor. Thus, using a distillation system which has a big relative volatility and
fast dynamic can further illustrate the significance of the NLCIMIC inferential
scheme.
the inverse nonlinear model. Thus, a deeper study on the nonlinear model
inversion technique and using a better method can definitely improve the
NLCIMIC performance.
3. In this work, a SISO NLCIMIC is developed to control the bottom product purity
128
further expanded in a multivariable control scheme which can be used to control
both the top and bottom product of the distillation column simultaneously.
129
REFERENCES
Abdullah, Z., Aziz, N., and Ahmad, Z. (2007). Nonlinear modelling application in
distillation column. Chemical Product and Process Modeling 2.
Agachi, P. S., Nagy, Z. K., Cristea, M. V., and Lucaci, A. I. (2006). "Model Based
Control: Case Studies in Process Engineering," Wiley-VCH Verlag GmbH
&Co. KGaA, Weinheim.
Alina-Simona, B., Nicolae, P., and Daniel, M. (2011). Using an internal model
control method for a distillation column. In "Mechatronics and Automation
(ICMA), 2011 International Conference on", pp. 1588-1593.
Arkun, Y., Canney, W. M., Hollett, J., and Morari, M. (1986). Experimental study of
internal model control. Industrial & Engineering Chemistry Process Design
and Development 25, 102-108.
Aske, E. M. B., Strand, S., and Skogestad, S. (2008). Coordinator MPC for
maximizing plant throughput. Computers & Chemical Engineering 32, 195-
204.
Aspen Tech 2009, Aspen Physical Properties System V7.1. Available form: Aspen
Technology Inc.[2 August 2013].
130
Basualdo, M. S., Calvo, R. A., and Ceccatto, H. A. (1994). Neural control strategies
of a binary distillation column. In "IEEE International Symposium on
Industrial Electronics", pp. 77-81.
Berber, R., and Brosilow, C. (1999). Algorithmic Internal Model Control of Unstable
Systems. In "Proceedings of the 7th Mediterranean Conference on Control
and Automation (MED99)", Haifa, Israel.
Bettoni, A., Bravi, M., and Chianese, A. (2000). Inferential control of a sidestream
distillation column. Computers & Chemical Engineering 23, 1737-1744.
Bhat, N., and McAvoy, T. J. (1990). Use of neural nets for dynamic modeling and
control of chemical process systems. Computers & Chemical Engineering 14,
573-582.
Biyanto, T. R., Widjiantoro, B. L., Jabal, A. A., and Budiati, T. (2010). Artificial
neural network based modeling and controlling of distillation column system.
International Journal of Engineering, Science and Technology 2(6).
Brezak, D., Bacek, T., Majetic, D., Kasac, J., and Novakovic, B. (2012). A
comparison of feed-forward and recurrent neural networks in time series
forecasting. In "Computational Intelligence for Financial Engineering &
Economics (CIFEr), 2012 IEEE Conference on", pp. 1-6.
131
Brosilow, C., and Joseph, B. (2002). "Techniques of Model Based Control," Prentice
Hall, New Jersy.
Chawankul, N., Budman, H., and Douglas, P. L. (2005). The integration of design
and control: IMC control and robustness. Computers and Chemical
Engineering 29, 261-271.
Cheng, C., and Chiu, M. S. (2007). Adaptive IMC Controller Design for Nonlinear
Process Control. Chemical Engineering Research and Design 85, 234-244.
Chia, T.-L., and Lefkowitz, I. (2010). Internal model-based control for integrating
processes. ISA Transactions 49, 519-527.
Chidambaram, M., and Reddy, G. P. (1996). Nonlinear control of systems with input
and output multiplicities. Computers & Chemical Engineering 20, 295-299.
Dez, E., Langston, P., Ovejero, G., and Romero, M. D. (2009). Economic feasibility
of heat pumps in distillation to reduce energy use. Applied Thermal
Engineering 29, 1216-1223.
132
Economou, C. G., and Morari, M. (1986). Internal model control. 6. Multiloop
design. Industrial & Engineering Chemistry Process Design and
Development 25, 411-419.
Economou, C. G., Morari, M., and Palsson, B. O. (1986). Internal Model Control. 5.
Extension to Nonlinear Systems, Industrial & Engineering Chemistry,
Process Design and Development 25, 403-411.
Fieg, G., Landwehr, B., and Wozny, G. (1996). About the possibility of a direct
concentration control for distillation columns. Chemical Engineering and
Technology 19, 299-307.
Fruehauf, P. S., and Mahoney, D. P. (1993). Distillation column control design using
steady state models: Usefulness and limitations. ISA Transactions 32, 157-
175.
Garcia, C. E., and Morari, M. (1982). Internal model control. A unifying review and
some new results. Industrial & Engineering Chemistry Process Design and
Development 21, 308-323.
Garcia, C. E., and Morari, M. (1985a). Internal model control. 2. Design procedure
for multivariable systems. Industrial & Engineering Chemistry Process
Design and Development 24, 472-484.
133
Garcia, C. E., and Morari, M. (1985b). Internal model control. 3. Multivariable
control law computation and tuning guidelines. Industrial & Engineering
Chemistry Process Design and Development 24, 484-494.
Glitsch, F. W. (1974). "Ballast Tray Design Manual Bulletin 4900 " 6th/Ed. Glitsch
Inc., Dallas, Texas.
Guttinger, T. E., Dorn, C., and Morari, M. (1997). Experimental Study of Multiple
Steady States in Homogeneous Azeotropic Distillation. Industrial &
Engineering Chemistry Research 36, 794-802.
Hagan, M. T., Demuth, H. B., and Beale, M. H. (1996). "Neural network design,"
PWS Boston, MA.
Hagan, M. T., Demuth, H. B., and Jess, O. D. (2002). An introduction to the use of
neural networks in control systems. International Journal of Robust and
Nonlinear Control 12, 959-985.
Haykin, S. S. (2009). "Neural networks and learning machines," Prentice Hall, New
York.
134
Henson, M. A., and Seborg, D. E. (1991). An internal model control strategy for
nonlinear systems. AIChE Journal 37, 1065-1081.
Hernjak, N., Doyle III, F. J., Ogunnaike, B. A., and Pearson, R. K. (2004). Chapter
A2 Chemical process characterization for control design. In "Computer
Aided Chemical Engineering" (S. Panos and C. G. Michael, eds.), Vol.
Volume 17, pp. 42-75. Elsevier.
Hoffman, H. L., Lupfer, D. E., Kane, L. A., Jensen, B. A., and Liptk, B. G. (2006).
Distillation: Basic Controls. In "Process Control And Optimization" (B. G.
Liptk, ed.), Vol. 2. CRC Press, Boca Raton, FL.
Hu, Q., and Rangaiah, G. P. (1999). Adaptive internal model control of nonlinear
processes. Chemical Engineering Science 54, 1205-1220.
Hunt, K. J., and Sbarbaro, D. (1991). Neural networks for nonlinear internal model
control. Control Theory and Applications, IEE Proceedings D 138, 431-438.
Hurowitz, S., Anderson, J., Duvall, M., and Riggs, J. B. (2003). Distillation control
configuration selection. Journal of Process Control 13, 357-362.
135
Hussain, M. A., and Kershenbaum, L. S. (2000). Implementation of an inverse-
model-based control strategy using neural networks on a partially simulated
exothermic reactor. Chemical Engineering Research and Design 78, 299-
311.
Ilme, J., Klemola, K., Aittamaa, J., and Nystrom, L. (2001). Calculating distillation
efficiencies of multicomponent i-butane/n-butane column. Chemical
Engineering Communications 184, 1-21.
Irwin, G. W., Warwick, K., and Hunt, K. J. (1995). "Neural network applications in
control," The Institution of Electrical Engineers, London.
Jensen, B. A., and Abonyi, J. (2006). Neural Networks for Process Modeling. In
"Process Control And Optimization" (B. G. Liptk, ed.), Vol. 2. CRC Press,
Boca Raton, FL.
Juan, C., Lu, W., and Bin, D. (2008). Modified internal model control for chemical
unstable processes with time-delay. In "Intelligent Control and Automation,
2008. WCICA 2008. 7th World Congress on", pp. 6353-6358.
136
Juwari, Chin, S. Y., Samad, N. A. F. A., and Aziz, B. B. A. (2008). A structure of
two-degree-of-freedom internal model control from feedback/feedforward
scheme. In "10th International Conference on Control, Automation, Robotics
and Vision, ICARCV", pp. 2044-2048, Hanoi, Vietnam.
Kano, M., Showchaiya, N., Hasebe, S., and Hashimoto, I. (2003). Inferential control
of distillation compositions: selection of model and control configuration.
Control Engineering Practice 11, 927-933.
Kister, H. Z., Mathias, P. M., Steinmeyer, D. E., Penney, W. R., Crocker, B. B., and
Fair, J. R. (2007). Equipment for Distillation, Gas Absorption, Phase
Dispersion and Phase Separation. In "Perrys Chemical Engineers
Handbook" (R. H. Perry and D. W. Green, eds.), pp. 14-51. McGraw-Hill,
New York.
Koggersbl, A., Andersen, T. R., Bagterp, J., and Jrgensen, S. B. (1996). An output
multiplicity in binary distillation: Experimental verification. Computers &
Chemical Engineering 20, S835-S840.
Lee, Y., Lee, J., and Park, S. (2000). PID controller tuning for integrating and
unstable processes with time delay. Chemical Engineering Science 55, 3481-
3493.
137
Liu, T., and Gao, F. (2011). Enhanced IMC design of load disturbance rejection for
integrating and unstable processes with slow dynamics. ISA Transactions 50,
239-248.
Ma, K., Valds-Gonzlez, H., and Bogle, I. D. L. (2010). Process design in SISO
systems with input multiplicity using bifurcation analysis and optimisation.
Journal of Process Control 20, 241-247.
Morari, M., and Zafiriou, E. (1989). "Robust Process Control," Prentice Hall,
Englewood Cliffs, New Jersey.
Mujtaba, I. M., Aziz, N., and Hussain, M. A. (2006). Neural network based
modelling and control in batch reactor. Chemical Engineering Research and
Design 84, 635-644.
Murad, G., Postlethwaite, I., and Gu, D. W. (1997). A discrete-time internal model-
based H infinity controller and its application to a binary distillation column.
Journal of Process Control 7, 451-465.
138
Nahas, E. P., Henson, M. A., and Seborg, D. E. (1992). Nonlinear internal model
control strategy for neural network models. Computers & Chemical
Engineering 16, 1039-1057.
Neste Oil, Neste Oil Product and Safety data sheets (2013) [Online], [Accessed on
7th July 2013]. Available from World Wide Web:
http://www.neste.fi/tuotteet_haku.aspx?path=2589;2655;2698;2699
Norgaard, M., Ravn, O., Poulsen, N. K., and Hansen, L. K. (2000). "Neural
Networks for Modelling and Control of Dynamic Systems," Springer-Verlag,
London.
Psaltis, D., Sideris, A., and Yamamura, A. A. (1988). A multilayered neural network
controller. Control Systems Magazine, IEEE 8, 17-21.
Psichogios, D. C., and Ungar, L. H. (1991). Direct and indirect model based control
using artificial neural networks. Industrial & Engineering Chemistry
Research 30, 2564-2573.
139
Rhinehart, R. R. (2006). Model-Based Control. In "Process Control And
Optimization" (B. G. Liptk, ed.), Vol. 2. CRC Press, Boca Raton, FL.
Riggs, J. B., and Ford, P. E. (2010). Chapter 34 - Advanced Control for the Plant
Floor. In "Instrumentation Reference Book (Fourth Edition)", pp. 619-628.
Butterworth-Heinemann, Boston.
Rivera, D. E., Morari, M., and Skogestad, S. (1986). Internal model control. 4. PID
controller design. Industrial & Engineering Chemistry Process Design and
Development 25, 252-265.
Sarle, W. S. (2002) Neural Network FAQ [Online]. [6th October 2012]. Available
from World Wide Web: ftp://ftp.sas.com/pub/neural/FAQ.html
Saxena, S., and Hote, Y. (2012). Advances in Internal Model Control Technique: A
Review and Future Prospects. IETE Tech Rev 29, 461-472.
Seborg, D. E., Edgar, T. F., and Mellichamp, D. A. (2004). "Process Dynamics and
Control: Second Edition," John Wiley & Sons, Hoboken, NJ.
140
Shamsuzzoha, M., and Lee, M. (2008). Analytical design of enhanced PID filter
controller for integrating and first order unstable processes with time delay.
Chemical Engineering Science 63, 2717-2731.
Shaw, A. M., Doyle III, F. J., and Schwaber, J. S. (1997). A dynamic neural network
approach to nonlinear process modeling. Computers and Chemical
Engineering 21, 371-385.
Skogestad, S. (2003). Simple analytic rules for model reduction and PID controller
tuning. Journal of Process Control 13, 291-309.
Skogestad, S., Lundstrm, P., and Jacobsen, E. W. (1990). Selecting the best
distillation control configuration. AIChE Journal 36, 753-764.
Skogestad, S., and Morari, M. (1987). The dominant time constant for distillation
columns. Computers & Chemical Engineering 11, 607-617.
Stryczek, K., Laiseca, M., Brosilow, C., and Leitman, M. (2000). Tuning and design
of single-input, single-output control systems for parametric uncertainty.
AIChE Journal 46, 1616-1631.
Subawalla, H., Paruchuri, V. P., Gupta, A., Pandit, H. G., and Rhinehart, R. R.
(1996). Comparison of Model-Based and Conventional Control: A Summary
of Experimental Results. Industrial & Engineering Chemistry Research 35,
3547-3559.
141
Tan, W., Marquez, H. J., and Chen, T. (2003). IMC design for unstable processes
with time delays. Journal of Process Control 13, 203-213.
Varshney, T., Varshney, R., and Sheel, S. (2009). ANN based IMC scheme for
CSTR. In "Proceedings of the International Conference on Advances in
Computing, Communication and Control", pp. 543-546. ACM, Mumbai,
India.
Wang, Q.-G., Zhang, Y., and Zhang, Y. (2001). A New Internal Model Control
Scheme with Simplified Design and Implementation. Chemical Engineering
Communications 184, 35 - 47.
Wang, Q. G., Zhang, Y., and Chiu, M. S. (2002). Decoupling internal model control
for multivariable systems with multiple time delays. Chemical Engineering
Science 57, 115-124.
142
Yamada, K. (1999). Modified Internal Model Control for unstable systems. In
"Proceedings of the 7th Mediterranean Conference on Control and
Automation (MED99)", pp. 293-302, Haifa, Israel.
Yamada, T. (2011). Discussion of neural network controllers from the point of view
of inverse dynamics and folding behavior. In "SICE Annual Conference
(SICE), 2011 Proceedings of", pp. 2210-2215.
Yang, X.-P., Wang, Q.-G., Hang, C. C., and Lin, C. (2002). IMC-Based Control
System Design for Unstable Processes. Industrial & Engineering Chemistry
Research 41, 4288-4294.
Zheng, A., Grassi, V., and Meski, G. (1998). On Control of Distillation Columns
with Input Multiplicity. Industrial & Engineering Chemistry Research 37,
1836-1840.
143
APPENDICES
APPENDIX A
The overall steady-state material balance of a distillation column based on Figure 2.1
is express as:-
A.3
or
A.4
Equation A.3 and A.4 define the distillation cut i.e. the percentage of the total feed
exits the column as distillate or bottom product for a specified inlet and outlet purity.
where is the vapour boil-up rate and is reflux flowrate. The subscripts denote
144
The material balance around the reboiler:
and assuming that the molar flow of liquid and vapour are constant throughout the
A.10
A.11
Therefore, the relationship between the distillate product flowrate ( ) and bottom
product flowrate ( ) with reflux flowrate ( ) and vapour boil-up rate ( ) can be
described by:-
A.12
A.13
145
APPENDIX B
146
LIST OF PUBLICATIONS
Processes", Singapore.
147