Model-Based Controller Design for General Nonlinear Processes

A Thesis

Submitted to the Faculty

of

Drexel University

by

Chanin Panjapornpon

in partial fulfillment of the

requirements for the degree

of

Doctor of Philosophy

September 2005
 @ Copyright 2005
Chanin Panjapornpon. All Rights Reserved.
 ii
Dedications

To my parents
 iii
Acknowledgements

First, I would like to express my deep gratitude to my research advisor, Prof. Masoud

Soroush, for his help and supervision on this project. This work would not be possible

without his invaluable suggestions and guidance. He motivated and helped me to expand

my knowledge through self-learning. He showed me the importance of organization and

clarity in presentations and publications. I am also grateful to my thesis committee

members, Professors Richard A. Cairncross, Bor-Chin Chang, Warren D. Seider, and

Charles B. Weinberger for their valuable comments and suggestions.

I would like to thank the rest of the chemical engineering faculty members for

their help with the graduate classes. I would also like to credit Dorothy Porter and Orit

Darwish for all the help with administrative work, and Dan Luu for all the assistance on

technical matters.

My research and study would not have been possible without financial support

from the Office of Civil Service Commission, the Royal Thai Government and the

National Science Foundation.

There are a number of people, family and friends that I am grateful and deserve to

be a part of my success. I would first like to thank my sister and my brother who help me

take care of my father. Special thanks to Nasir Mehranbod, Congling Quan, Felix S.

Rantow, and Swa S. Metta who are my best colleagues in Prof. Soroush’s research group.

They are always open for discussions on my research struggles. I would also like to

thank Kaveh Komaee, Kevin B. Towles, Ehsan Jabbarzadeh and Veena Pata who have

made the lab an enjoyable place to work in. I also would like to give special thanks to the
 iv
Thai community in Philadelphia and Tanawadee Dechakupt, a special person in my life,

who consistently supports me and makes my life more enjoyable.

Most importantly, I would like to thank my parents for giving me unconditional

love and support from the other part of the world.

 v
Table of Contents

List of Tables ................................................................................................................... viii

List of Figures .................................................................................................................... ix

Abstract .............................................................................................................................. xi

1. Introduction................................................................................................................... 1

1.1 Motivation.............................................................................................................. 1

1.2 Previous Work ....................................................................................................... 2

1.2.1 Nonlinear Model-Based Control.................................................................. 2

1.2.1.1 Differential-Geometric Control ...................................................... 3

1.2.1.2 Model-Predictive Control ............................................................... 5

1.2.1.3 Lyapunov-Based Control ................................................................ 8

1.2.2 Software for Analytical Model-Based Controller Design ........................... 9

1.3 Scope and Objectives........................................................................................... 12

1.4 Significance of this Research............................................................................... 13

1.5 Organization of the Thesis ................................................................................... 14

2. Approximate Input-State Linearization ...................................................................... 16

2.1 Introduction.......................................................................................................... 16

2.2 Scope and Mathematical Preliminaries................................................................ 17

2.2.1 Approximate Input-Output Linearization .................................................. 19

2.2.2 Reduced-Order State Observer Design...................................................... 20

2.3 Nonlinear Control Method................................................................................... 22

2.3.1 State Feedback Design............................................................................... 22

 vi
2.3.2 Reduced-Order State Observer .................................................................. 25

2.3.3 Integral Action ........................................................................................... 26

2.3.4 Controller System ...................................................................................... 26

2.4 Illustrative Examples ........................................................................................... 28

2.4.1 Single-Input Single-Output Chemical Reactor .......................................... 28

2.4.2 Single-Input Single-Output Bioreactor ...................................................... 35

2.4.3 Multi-Input, Multi-Output Chemical Reactor............................................ 40

2.5 Conclusions.......................................................................................................... 43

2.6 Notation................................................................................................................ 47

3. Input-Output Linearization with Stability Constraint ................................................. 49

3.1 Introduction.......................................................................................................... 49

3.2 Scope and Mathematical Preliminaries................................................................ 51

3.2.1 MPC Formulation of Input-Output Linearization...................................... 52

3.2.2 Reduced-Order State Observer Design...................................................... 54

3.3 Nonlinear Control Method................................................................................... 56

3.3.1 Lyapunov Stability Constraint ................................................................... 56

3.3.2 Nonlinear State Feedback Design.............................................................. 58

3.3.2.1 Implementation of Nonlinear State Feedback............................... 59

3.3.3 Reduced-Order State Observer .................................................................. 63

3.3.4 Integral Action ........................................................................................... 63

3.3.5 Controller System ...................................................................................... 66

3.4 Illustrative Examples ........................................................................................... 68

3.4.1 Single-Input Single-Output Chemical Reactor .......................................... 68

 vii
3.4.2 Two-Input Two-Output Chemical Reactor................................................ 72

3.5 Conclusions.......................................................................................................... 76

3.6 Notation................................................................................................................ 77

4. Software for Analytical Model-Based Controller Design .......................................... 79

4.1 Introduction.......................................................................................................... 79

4.2 Mathematical Preliminaries ................................................................................. 80

4.3 Software Environment ......................................................................................... 84

4.3.1 User Interface............................................................................................. 85

4.3.2 MATHEMATICA Packages...................................................................... 92

4.3.3 Simulation Algorithm ................................................................................ 94

4.4 Examples.............................................................................................................. 94

4.4.1 Continuous Bioreactor ............................................................................... 95

4.4.2 Single-Input Single-Output Chemical Reactor .......................................... 97

4.4.3 Multi-Input Multi-Output Jacketed Chemical Reactor ............................ 101

4.4.4 Continuous Chemical Reactors in Series ................................................. 106

4.5 Conclusions........................................................................................................ 108

4.6 Notation.............................................................................................................. 110

5. Conclusions and Future Research Directions ........................................................... 113

5.1 Conclusions........................................................................................................ 113

5.2 Future Research Directions................................................................................ 117

List of References ........................................................................................................... 120

Vita.................................................................................................................................. 129
 viii
List of Tables

2.1 Closed-loop eigenvalues of the linear example for several p1 and p2 values..........25

2.2 Values of the parameters of the non-isothermal reactor model ................................29

2.3 Closed-loop eigenvalues of the CSTR example for several p1 and p2 values.........32

2.4 Closed-loop eigenvalues of the bioreactor example for several p1 and p2 values ...37

2.5 Parameter parameters of the non-isothermal reactor with cooling jacket.................41

3.1 Process values of the non-isothermal reactor............................................................69

4.1 Controller design packages .......................................................................................93

4.2 Design analysis packages..........................................................................................94

4.3 Process parameters of the non-isothermal reactor ....................................................98

4.4 Process parameters of the non-isothermal jacketed reactor....................................102

 ix
List of Figures

2.1 Parameterized controller system ...............................................................................27

2.2 Schematic of the non-isothermal continuous stirred tank reactor.............................28

2.3 Closed-loop response of the state variables of the chemical reactor ........................33

2.4 Manipulated input and unmeasured disturbance profiles corresponding to

Figure 2.3 ..................................................................................................................34

2.5 Closed-loop response of the bioreactor state variables.............................................38

2.6 Manipulated input and unmeasured disturbance profiles corresponding to

Figure 2.5 ..................................................................................................................39

2.7 Schematic of the multivariable non-isothermal chemical reactor.............................40

2.8 Closed-loop response of the first two state variables of the chemical reactor..........44

2.9 Closed-loop response of the last two state variables of the chemical reactor...........45

2.10 Manipulated input profiles corresponding to Figures 2.8 and 2.9 ............................46

3.1 Illustration of a controlled output response achieved under

the control system .....................................................................................................62

3.2a State responses under the controller system and the controller flag .........................64

3.2b Manipulated input response and the Lyapunov function under

the controller system .................................................................................................65

3.3 Parameterization of the controller system.................................................................67

3.4 Schematic of non-isothermal continuous stirred tank reactor (CSTR).....................67

3.5a State responses of the non-isothermal CSTR under the controller

and the controller flag ...............................................................................................70

3.5b Manipulated input response and the Lyapunov function,

corresponding to Figure 3.5a ....................................................................................71

3.6a State response of the non-isothermal MIMO CSTR under

the control system and the controller flag.................................................................74
 x
3.6b Manipulated input responses and the Lyapunov function under the control
system, corresponding to Figure 3.6b .......................................................................75

4.1 Flow diagram of the major components of the model-based controller

design software .........................................................................................................84

4.2 Flow diagram of the tasks in the model-based controller design software
(numbers correspond to the steps in the user interface)............................................86

4.3 The system information window (step 1) of model-based controller design

software.....................................................................................................................87

4.4 The process model window (step 2) of model-based controller

design software .........................................................................................................88

4.5 Diagram showing how the software directs the user to the controller methods
that are applicable to the continuous process under consideration ..........................89

4.6 The code viewer window showing controller equations in the FORTRAN
language format ........................................................................................................90

4.7 The simulation setup (step 4) and the plot selection (step 5) windows
of the software...........................................................................................................91

4.8 A simulation plot generated by plot selection step of the software ..........................92

4.9 Simulated closed-loop response of the state variables and manipulated

input of the bioreactor...............................................................................................97

4.10 Simulated closed-loop response of the controlled output and manipulated

input of the non-isothermal CSTR..........................................................................100

4.11 Simulated closed-loop response of the controlled outputs and manipulated

input of the MIMO chemical reactor ......................................................................105

4.12 Simulated closed-loop response of the controlled outputs and manipulated

input of the isothermal chemical reactors in series .................................................109

5.1 Illustration of the output prediction and control action in model-predictive

control with prediction horizon P and control horizon M........................................116
 xi
Abstract
Model-Based Controller Design for General Nonlinear Processes
Chanin Panjapornpon
Masoud Soroush, Ph.D.

Non-minimum-phase behavior of a process limits the degree of achievable control quality

and complicates the controller design for the process. This behavior can be due to the

presence of an unstable mode in the zero dynamics, a time delay, or both. To achieve

greater profitability, process designers have been creating designs in regions involving

complex nonlinearity where the controllers continue to face stiff challenges. Operation is

often more profitable at an unstable steady state or a stable steady state close to an

unstable one, sometimes involving non-minimum-phase behavior. This project aims to

advance the existing non-minimum-phase control theory and to improve the operation of

processes that are unstable and exhibit non-minimum-phase behavior.

The research project has two main objectives: (i) the derivation of a general

model-based control system and (ii) the development of a software package for model-

based controller design. Two novel nonlinear control laws that are applicable to general

multivariable processes, whether non-minimum- or minimum-phase, were developed.

The first control law ensures closed-loop stability by forcing all state variables to follow

their corresponding linear reference trajectories. The second control law guarantees the

closed-loop stability by using a Lyapunov hard constraint that requires all state variables

to evolve within a shrinking state ‘tube’.

A prototype controller design software package was developed to simplify the

implementation of differential geometric, model-based controllers. The software package

carries out symbolic manipulations to automatically generate the analytical model-based

controllers and subsequently test the designed controllers. It has a user-friendly interface
 xii
that allows the user to enter process model equations and process parameters easily.

Using the controller design software, one can design the model-based controllers with

much less effort, which is expected to result in more industrial applications of the

controllers.
 1
Chapter 1: Introduction

1.1 Motivation

In an effort to achieve maximum profitability, chemical and biochemical processes have

to find a compromise between product quality and operating cost. Process optimization

has become a preferred choice for a process engineer to achieve this goal. Often

operation is more profitable at either an unstable steady state or at a stable steady state in

the proximity of an unstable steady state, sometimes involving non-minimum-phase

behavior (inverse response). In processes exhibiting such behavior, the controlled output

initially responds in the opposite of the desired direction. Examples of these processes

include chemical reactors [1-5], fermentation reactors [6-9], fluidized catalytic crackers,

reactor-separator-recycle plants [10], azeotropic distillation columns [11, 12], and

reboilers [13]. Non-minimum-phase behavior limits the degree of achievable control

quality and complicates the controller design for the process. This behavior can be due to

the presence of an unstable mode in the zero dynamics (finite right-half-plane zero in the

linear case), a time delay (infinite right-half-plane zero in the linear case), or both. In

model-based control, significant advances have been made in an effort to address the

control problem of non-minimum-phase processes. Differential-geometric control is a

model-based control method which involves direct controller synthesis. However, the

majority of the existing differential-geometric methods that are applicable to non-

minimum-phase processes are not easy to implement because these methods require

solving a set of partial differential equations. In addition, there are a few control methods

in differential-geometric approach that are able to control the unstable non-minimum-

 2
phase processes. These methods are limited to a few classes of nonlinear processes. The

existing nonlinear differential-geometric model-based control methods to handle the

general multivariable processes whether minimum-phase or non-minimum-phase are still

inadequate. In addition, unavailable is a software package that fully automates the design

of differential-geometric controllers, given a process model in the form of differential and

algebraic equations. The advances in the control of unstable, non-minimum-phase

systems will allow for tighter control of severely nonlinear processes with complex

dynamics and is expected to lead to greater interest in model-based control. Furthermore,

the advances in controller design software are expected to increase the industrial use of

differential-geometric control.

1.2 Previous Work

The relevant research work either in the nonlinear model-based control for non-

minimum-phase processes or in the software for analytical model-based controller design

are briefly discussed in the following subsections.

1.2.1 Nonlinear Model-Based Control

In the area of nonlinear model-based control, the main interest has been focusing on the

frameworks of model-predictive control, differential-geometric control, and Lyapunov-

based control. In model-predictive control, controller action is the solution to a

constrained optimization problem that is solved on-line, involving numerical model

inversion [14]. Local closed-loop stability is guaranteed by imposing designed constraints

or penalty terms on optimization problems. In the differential-geometric control, the

controller is derived by requesting a desired closed-loop response in the absence of input

 3
constraints [15], involving analytical model inversion. In Lyapunov-based control,

closed-loop stability plays a fundamental role in the controller design [16-18]. The

asymptotic decay of a norm of the state variables is ensured by applying a proper

Lyapunov function in the controller design. More details in each framework are discussed

in the following subsections.

1.2.1.1 Differential-Geometric Control

The differential-geometric control is the dominant approach in nonlinear model-based

control due to direct synthesis of the controllers and the proof on closed-loop stability.

Input-Output (I-O) linearization has been the most widely used differential-geometric

control, because it is easy to implement (i.e. it does not require solving a set of partial

differential equations). Initially, I-O linearization was developed for unconstrained,

minimum-phase (MP) processes [19, 20]. The I-O linearization cannot be applied directly

to non-minimum-phase (NMP) processes because it includes an inverse of the process

model, causing closed-loop instability in the case of NMP processes.

During the past fifteen years, several successful attempts have been made to

develop differential-geometric controllers applicable to NMP nonlinear processes [1, 2,

15, 21-33]. For example, total linearization [15, 27, 28], equivalent outputs [21, 29],

approximate I-O linearization [1, 25, 32], trajectory linearization [31], H ∞ optimal

controller [33], and stable inversion [22-24, 26]. Isidori [15], Isidori and Byrnes [28],

and Isidori and Astolfi [27] studied total linearization, which is applicable to a limited

class of minimum- and non-minimum-phase processes and requires solving a nonlinear

partial differential equation. Kravaris and Daoutidis [2] present a nonlinear state-

feedback controller for second-order, non-minimum-phase, nonlinear systems. Niemiec

 4
and Kravaris [21] propose a systematic procedure for the construction of statically

equivalent outputs with prescribed transmission zeros. They then design a nonlinear

state-feedback controller on the basis of the synthetic outputs. Kanter et al. [1] developed

nonlinear control laws for input-constrained, multiple-input, multiple-output stable

processes, whether their delay-free part was minimum- or non-minimum-phase. They

addressed the nonlinear control of the processes by inducing an approximately linear,

input-output response that exploited the connection between model-predictive control and

I-O linearization. Kravaris et al. [29] present a systematic method of arbitrarily assigning

the zero dynamics of a nonlinear system by constructing the requisite synthetic output

maps. The minimum-phase, synthetic output maps constructed can be made statically

equivalent to the original output maps, and therefore, they can be directly used for non-

minimum-phase compensation purposes. The method requires solving a system of first-

order, nonlinear, singular PDEs. Mickle et al. [31] develop a tracking controller for

unstable, non-minimum-phase, nonlinear processes by using trajectory linearization.

Tomlin and Sastry [32] derived tracking control laws for non-minimum-phase, nonlinear

systems with both fast and slow, possibly unstable, zero dynamics. van der Schaft [33]

develops a nonlinear state feedback H ∞ optimal controller. Devasia et al. [24] and

Devasia [23] introduce an inversion procedure for nonlinear systems that constructs a

bounded input trajectory in the pre-image of a desired output trajectory. The pre-image

trajectory is noncausal (rather than unstable) in the non-minimum-phase case. Hunt and

Meyer [26] show that under appropriate assumptions the bounded solution of the partial

differential equation of Isidori and Byrnes [28] for each trajectory of an exosystem must

be given by an integral representation formula of Devasia et al. [24]. Chen and Paden
 5
[22] studied the stable inversion of non-minimum-phase nonlinear systems. They derive a

stable, but non-causal, inverse that can be incorporated into a stabilizing controller for

output tracking. Doyle et al. [25] present a control synthesis scheme for nonlinear, single-

input, single-output systems which have completely unstable zero dynamics. The

approach is similar to linear approaches for non-minimum-phase systems and involves

the derivation of an input-output linearizing controller for a suitably-defined, nonlinear,

minimum-phase approximation to the original system. The linearizing controller achieves

an approximately linear input-output response and internal stability. McLain et al. [30]

propose a controller design strategy for nonlinear systems with more manipulated inputs

than controlled outputs. The controller can be used for non-minimum-phase processes.

While controllers in [1, 22-26, 30] are applicable to multi-input, multi-output (MIMO),

NMP processes, either sets of partial differential equations must be solved [21, 27, 29], or

the controllers are applicable to a very limited classes of processes [2, 15, 22-28].

1.2.1.2 Model-Predictive Control

Model-predictive control is a discrete-time control in which control action is the solution

to an open-loop constrained optimization problem at each time step. The current values

of the process states are used for calculating the control action, and the optimization

process is repeated at the next time horizon. A small time interval and a long prediction

horizon are required to maintain closed-loop robustness. However, the smaller the time

interval is, the heavier the computational load. The model-predictive control generally

formulates the optimization problem over a finite prediction horizon to decrease the

computational load. In general, the nonlinear model-predictive control leads to a non-

convex optimal control problem. The solution of the optimization problem is a local
 6
optimum. However, the model-predictive control can achieve a global optimum by using

suboptimal model-predictive control [34].

The model-predictive control has been used widely in the process industries

because of its many appealing features such as handling multivariable systems with time-

delays. In addition, the constraints on manipulated inputs, states and output variables are

explicitly handled in the formulation of the optimization problem. However, the closed-

loop stability and feasibility are major concerns in the model-predictive control. The local

optimization in a finite horizon does not guarantee closed-loop stability [34, 35]. Thus,

the model-predictive control formulates the optimization problem with special constraints

or penalty terms based on a Lyapunov function to ensure the closed-loop stability. The

stability is guaranteed by imposing an equality or inequality constraint on the final state

in the prediction horizon ( the terminal state constraint [36-38] ), adding a weight on the

final state in the objective function ( the terminal state penalty [35] ), or using an infinite

prediction horizon with a finite control horizon [39]. However, the measurable process

states, perfect process-model, and high computational load to determine the attraction

domain boundaries of the linear controller are required. Furthermore, a large number of

tunable parameters are needed when the optimization problem of the terminal state

constraint is complex. To avoid the high computational load, the contractive constraint

method [40] introduces a stabilizing state constraint and requires the process states at the

end of the prediction horizon to be norm-contracted with respect to the process states at

the beginning of the prediction horizon. The stability of the closed-loop system in the

model-predictive control can be tested by employing a Lyapunov function, or a sequence

 7
of monotonic objective functions only when the prediction horizon is infinite or when a

terminal state constraint is applied.

The control problems of unstable processes and non-minimum-phase processes

have drawn attention for many years. The control problem of the unstable process

requires availability of sufficient control action to keep the process variables within a

limited region around the desired operating point. Gobin et al. [41] employed a dynamic

matrix control technique that requires linearized process model around an unstable

operating point. Nagrath et al. [42] proposed state-estimation-based model-predictive

control with no terminal state penalty term to control the unstable processes. The control

method employed a concept of cascade control that uses the secondary measurement to

aid disturbance rejection. However, the stability and stabilizing constraint are dependent

on proper tuning parameters. The control problem of non-minimum-phase processes is

handled simply by requesting a sufficiently large prediction horizon and a sufficiently

short control horizon [43, 44]. Hernandez and Arkun [45] proposed a pth inverse

controller with long prediction horizon for processes. The Jacobian of the closed loop

system comes close to the Jacobian of the open-loop system as prediction horizon

approaches infinity.

Although the differential-geometric and model-predictive control are respectively

formulated in the continuous and discrete time, the relationship between the two

approaches has been discussed [46, 47]. Soroush et al. [46, 47] proposed a continuous-

time model-predictive control method that exploited the connections between model-

predictive and input-output linearizing control methods. The derived model-predictive

control laws have the shortest possible prediction horizon and explicit analytical form.
 8
The controller is derived by minimizing a function norm of the deviations of the

controlled outputs from desired linear reference trajectories. However, this method is not

applicable to unstable, non-minimum-phase processes.

1.2.1.3 Lyapunov-Based Control

In Lyapunov-based control, a concept of Lyapunov stability is integrated into control

design. Two dominant methods of Lyapunov stability are used to investigate the stability

of nonlinear systems, called linearization method and direct method. The linearization

method or indirect method determines the stability of the nonlinear system though the

eigenvalues of the Jacobian matrix of the linearized system around equilibrium. The

direct method determines the stability of nonlinear system by constructing a Lyapunov

function formulated as real positive definite. If the Lyapunov function is continuously

dissipated, then the system must eventually settle down to an equilibrium point, whether

minimum- or non-minimum-phase systems. Lyapunov-based control control methods are

usually of the following types [48]:

• Assume a suitable control law and then find a proper Lyapunov function to

prove closed-loop stability, and

• Assume a Lyapunov function candidate and then find a control law to make

this candidate a real Lyapunov function, usually called converse Lyapunov

theorem.

Most Lyapunov-based control methods are of the second type. Examples are the

controllers in [49-51] derived by solving inverse optimal control problem exploited with

the concept of control Lyapunov functions. Complementary to the Lyapunov direct

 9
method are feedback control with backstepping [52, 53], energy-predictive control [54,

55] and other approaches [56-60].

For the control problem of non-minimum-phase processes, Kazantzis and

Kravaris [61] developed the control method for unstable non-minimum-phase nonlinear

processes by integrating both feedback linearization and pole replacement design in a

single step via the adapted Luenberger approach. The Lyapunov auxiliary theorem is

employed to guarantee the existence and uniqueness of solution. However, this method is

required to truncate the control problem formulated in the form of a singular partial

differential equation system. In addition, it is restricted to the first order ODE system. Wu

[62] proposed the Lyapunov-based linearization method to ensure the asymptotic output

regulation. This method uses differential-geometric approach with the Lyapunov

function, and the partial state feedback is obtained by observer-based. However, the

method presented by Wu is applicable for non-minimum-phase nonlinear processes with

the actuator constraint and one controlled output. In addition, the application of this

technique requires the definition of the error variables and chooses the appropriate

Lyapunov function to stabilize the closed-loop system.

A problem that has received attention in recent years is how to expand the domain

of attraction for the closed-loop system in Lyapunov control. To solve this problem, the

use of combined control methods [18, 63, 64] and multiple Lyapunov functions in the

controller design [17] have been proposed.

1.2.2 Software for Analytical Model-Based Controller Design

Conventional linear controllers, PI and PID, were widely used in the industrial processes

for many decades. However, not all processes can be controlled with such controllers.
 10
More advanced control technique such as the model-based control becomes a preferred

controller design approach for processes exhibiting high nonlinearity. Model-predictive

control and differential-geometric control have been the most widely used model-based

control methods. The model-predictive control involves numerical model inversion,

while differential-geometric control involves analytical model inversion. The analytical

inversion required in differential-geometric control requires analytical partial derivatives

and symbolic manipulations, which become cumbersome as the relative order and/or the

level of complexity of the model increases. Indeed, the burden of taking analytical partial

derivatives and performing symbolic manipulations have prevented this theoretically-

sound, efficient controller from being implemented widely in the process industries. In

addition, many steps are involved in the design and implementation of differential-

geometric controllers including process modeling, model verification, process analysis,

controller design, controller verification, and code implementation. To complete the

design tasks, there are very elaborate and error prone steps. Furthermore, the differential-

geometric controller design still has a gap between the controller verification (by

simulation) and the code implementation. The problem with the target platform is not

found until the implementation begins. Thus, there are increasing needs for a systematic

approach to support controller design algorithms that would lead to reduction in the

amount of hand written code, reduction of redundant work in redesigning a new process,

and decrease in testing time. The design of the model-based controllers can be used

effectively and economically to reduce overall effort.

The increased role of computers in solving complex scientific and engineering

problems has led to software packages that facilitate the design of model-based
 11
controllers [65-74]. Symbolic manipulation software for controller design has developed

using packages such as MATHEMATICA [75], MAPLE [76], and the Symbolic Toolbox

in MATLAB [77]. In one approach, the software is designed as a toolbox for an existing

symbolic manipulation package [68, 70, 72, 73]. For example, Blankenship et al. [68]

developed a package for the design of nonlinear controllers using MATHEMATICA to

provide the controller equations in the format suitable for simulation in SIMULINK.

However, the software lacks a visual interface and cannot be used directly for controller

design and implementation. In general, existing software has numerical or symbolic

computation limitations and requires users to implement lengthy sequences of input

commands that follow specific patterns.

To avoid such limitations, an alternate approach to the development of software

for the controller design uses multiple calculating engines [69, 72]. For example,

Kitamoto et al. [72] created H ∞ controller-design software in which the design algorithm

is programmed in MATHEMATICA and the front-end interface is written in MATLAB.

Users input the control system in the form of block diagram. Unfortunately, however, the

software is unable to analyze the performance of the designed controller, and the

capabilities of the symbolic computation are not fully utilized. The software developed

by Campbell et al. [69] uses MAPLE to provide symbolic computation of the feedback

linearization design for mechanical multi-body systems and Scilab [78] to carry out the

numerical simulation.

Although symbolic computation software packages for model-based controller

design have been available for many years, their applications have been limited due to

deficiencies in available computer-aided design tools and user-interface platforms. For

 12
example, they do not design the controller fully when the process model is a set of

ordinary differential and/or algebraic equations. Furthermore, they are not user-friendly

for design integration and closed-loop simulation. Thus, the development of a software

package devoid of these limitations is of interest in the practice of process control.

1.3 Scope and Objectives

The I-O linearization methods have been widely used because these methods do not

require solving a set of partial differential equations and require fewer number of tuning

parameters. However, the I-O linearization could not be applied directly to unstable non-

minimum-phase processes because of an inverse of the process model causing closed-

loop instability. In addition, there are a few control methods in differential-geometric

approach that are able to control the unstable non-minimum-phase processes. These

methods are limited to a few classes of nonlinear processes.

Thus, one of the objectives of this research effort is to develop new control laws

that do not possess the limitations of I-O linearization. The specific research objectives

are as follows.

(1) To develop new control laws that are applicable to general nonlinear processes

within the framework of the differential-geometric control. The control laws should:

• Guarantee offset-free response in the presence of constant disturbances and

process-model mismatch,

• Guarantee the asymptotic stability of the closed-loop system, and

• Be applicable to processes with input constraints.

 13
(2) To develop a software package for designing model-based controllers for general

nonlinear processes. The newly developed software will help control engineers to

design conveniently model-based controllers for general nonlinear processes. The

software should:

• Automate the design of differential-geometric controllers,

• Have a user-friendly visual interface, and

• Be capable of generating controller equations in C, FORTRAN, or MATLAB

formats and also carry out closed-loop simulations.

Two novel control laws are presented to achieve the first objective, approximate input-

state linearization and input-output linearization with a stability constraint. The

controllers derived based on the developed control laws are applicable to stable and

unstable processes, whether non-minimum- or minimum-phase.

1.4 Significance of this Research

The impact of this research is two-fold. First, the problem of controller design for general

nonlinear processes is studied. The unstable non-minimum-phase behavior limits the

degree of achievable control quality of the process. The new control law allows complex

processing plants to accomplish greater profitability. Second, the new controller design

software enables control engineers to implement differential-geometric nonlinear

controllers with little effort. This is expected to lead to an improvement in the industrial

popularity of differential-geometric nonlinear controllers and better control of severely

nonlinear processes. This work can positively impact the profitability of processing plants

through the production of higher quality products at lower costs and improved control

quality.
 14
1.5 Organization of the Thesis

The dissertation is organized into two parts. The first part comprises Chapters 2-3 that

address the theoretical issues concerning two new nonlinear model-based control laws for

general multivariable nonlinear processes with input constraints. The second part

comprises Chapter 4 that describes the development of software for analytical model-

based controller design

In the part one, the approximate input-state linearization is introduced in Chapter

2. The proposed method ensures the closed-loop stability by forcing all process state

variables to follow their corresponding reference trajectories that have orders higher than

the state-variable relative orders. The control system includes a nonlinear state feedback,

integrator and a reduced order nonlinear state observer. The scope of the study,

mathematical preliminaries of the approximate input-output linearization and reduced-

order state observer, and nonlinear feedback control method are discussed in the

subsections of Chapter 2. Furthermore, Chapter 2 also demonstrates the performance of

the controller by numerical simulation of a bioreactor and chemical reactors.

In Chapter 3, the input-output linearization with stability constraint is discussed.

The second control method formulates the optimal output-tracking controller based on I-

O linearization and guarantees closed-loop asymptotic stability by including a Lyapunov

function as a hard constraint. The scope of the study, mathematical preliminaries of I-O

linearization and the Lyapunov’s direct method, and nonlinear control method are

discussed in the subsections of Chapter 3. In addition, Chapter 3 also demonstrates the

performance of the control method by numerical simulation of several process examples.

 15
In part two, the focus is shifted to the analytical model-based controller design

software. The software automates the design of differential-geometric controller for

general nonlinear processes, thus avoiding laborious mathematical derivations. The scope

of the study, mathematical preliminaries of geometric model-based control, the details of

the software, and application of the software are discussed in the subsections of Chapter

4. Finally, Chapter 5 presents the overall conclusions of this work and the suggestions for

future research directions.

 16
Chapter 2: Approximate Input-State Linearization

2.1 Introduction

Differential-geometric control is a direct synthesis approach in which the controller is

derived by requesting a desired closed-loop response in the absence of input constraints.

Input-Output (I-O) linearization has been the most widely used differential-geometric

control method because it does not require solving a set of partial differential equations.

However, I-O linearization cannot be applied directly to non-minimum-phase processes

(NMP) because it includes an inverse of the process model, thereby causing closed-loop

instability in the case of NMP processes. The differential-geometric control requires a

special treatment to handle the non-minimum-phase behavior.

During the past fifteen years, differential-geometric controllers have been

extended to NMP nonlinear processes [1, 2, 15, 21-30, 32, 33, 79]. While controllers in

[22-26, 30] are applicable to multi-input multi-output (MIMO) NMP processes, either

sets of partial differential equations must be solved [21, 27, 29], or the controllers are

applicable to a very limited class of processes [2, 15, 22-28, 30].

This chapter presents a new control method that uses the same continuous-time

model-predictive control framework employed in Kanter et al. [1] and Soroush and

Soroush [80]. However, it is conceptually different from those presented previously in [1,

80] as follows:

a) The method is now applicable to processes operated at an unstable non-minimum-

phase steady state in the presence of input constraints.

 17
b) The new controller is derived by requesting desired responses for the state

variables (instead of outputs).

c) The nonlinear state feedback is derived by minimizing a function norm of the

deviations of state variables from linear reference trajectories in each time instant,

that have orders higher than the state-variable relative orders.

In the proposed method, the closed-loop stability is ensured by forcing all process state

variables to follow their corresponding reference trajectories. The proposed control

system includes a nonlinear state feedback, integrator and a reduced order nonlinear state

observer. The resulting state feedback approximately induces linear responses to the state

variables in the presence of constraints. The integral action is added to ensure offset-free

output responses in the presence of constant disturbances and process-model mismatch.

The state observer is used to estimate state variables that are not measured.

This chapter is organized as follows. The scope of the study and some

mathematical preliminaries are given in Section 2.2 Section 2.3 presents the nonlinear

feedback control method. Finally, in Section 2.4, the application and performance of the

control method are illustrated by numerical simulation of a bioreactor and two chemical

reactors with multiple steady states.

2.2 Scope and Mathematical Preliminaries

Consider general, square, multi-input, multi-output processes having a mathematical

model in the form:

dx
= f ( x, u ) x(0) = x0
dt (2.1)
y = h( x )
 18
where x = [ x1 xn ] ∈ R
T n
is the vector of state variables, u = [u1 um ] ∈ R
T m
is the

vector of manipulated inputs, y = [ y1 ym ]T ∈ R m is the vector of controlled outputs,

f ( x, u ) = [ f1 ( x, u ) f n ( x, u )]T and h( x) = [h1 ( x) hm ( x)]T are smooth vector functions.

The relative order (degree) of a state variable, xi , is denoted by ri , where ri is the

smallest integer for which

∂ ⎛ d ri xi ⎞
⎜ ⎟ ≠ 0.
∂u ⎝ dt ri ⎠

It is assumed that the relative orders, r1 ,..., rn , are finite, and the process is controllable

and observable locally around a desired steady state.

Let H 0 ( x) = x , and define the following notation:

dxi
H i1 ( x) =
dt

d ri −1 xi
H iri −1 ( x) =
dt ri −1
d ri x
H iri ( x, u ) = ri i (2.2)
dt
ri +1
d x
H iri +1 ( x, u (0) , u (1) ) = ri +1i
dt

( ji − ri ) d ji xi
H i ( x, u , u , … , u
ji (0) (1)
)=
dt ji

d lu
where ji ≥ ri , i = 1,… , n , and u (l )
= l
dt

For a given output set-point, ysp , the corresponding desired steady-state pair ( xss , uss )

satisfy:
 19
0 = f ( xss , uss )
ysp = h( xss )

These relations are used to describe the dependence of a nominal (desired) steady

state, xssN , on the set-point, say xssN = F ( ysp ) .

2.2.1 Approximate Input-Output Linearization

To place the new control method in perspective, it is instructive to review briefly the

control method presented in [1]. For a process in the form of (2.1), in [1], closed-loop

process output responses of the linear form:

(τ 1 D + 1) P1 y1 = ysp1
(2.3)
(τ m D + 1) ym = yspm
Pm

are requested. Here D = d / dt , P1 ≥ R1 , , Pm ≥ Rm ( R1 , , Rm are the relative orders of

the process outputs, y1 ,… , ym , respectively), and τ 1 , ,τ m are positive constants that set

the speed of the responses of the process outputs, y1 ,… , ym , respectively. Taking the

output time-derivatives in (2.3) leads to the derivation of a dynamic state feedback in the

compact form:

Φ P ( x, u (0) , u (1) , , u ( P − R ) ) = ysp (2.4)

where P=max(P1,…,Pm) and R=min(R1,…,Rm). Setting all the time derivatives of u in

(2.4) to zero results in to the static state feedback:

φP ( x, u ) = Φ P ( x, u (0) , 0, , 0) = ysp (2.5)

If (2.5) has a real root for u at each time instant and ∂ φ p ( x , u ) / ∂ u ≠ 0 at the root, a state

feedback of the form exists:

 20
u = ψ P ( x, ysp ) (2.6)

Theorem 1 [1]. The closed-loop system under the state feedback of (2.6) is asymptotical-

ly stable, if the following conditions hold:

a) The nominal equilibrium pair of the process corresponding to ysp is hyperbolically

stable;

b) The tunable parameters P1 , , Pm are chosen to be sufficiently large; and

c) The tunable parameters, τ l = τ 1 , ,τ m , are chosen such that for every l = 1,… , m , all

eigenvalues of [ I + τ l J ol ( xss , uss )] lie inside the unit circle.

Furthermore, as P1 , , Pm → ∞, the eigenvalues of the Jacobian of the closed-loop system

[ J cl ( xss , uss )] approach the eigenvalues of the Jacobian of the open-loop process

[ J ol ( xss , uss )] .

2.2.2 Reduced-Order State Observer Design

This subsection presents a brief review of the nonlinear, reduced-order, state observer

described in [81]. The state observer will be used in this paper to reconstruct unmeasured

state variables of the process under consideration. Consider a nonlinear process in the

form of (2.1) with additional output measurements Y1 , , Yq ; that is,

dx
= f ( x, u ) x(0) = x0
dt
y = h( x )
Y = K ( x)

The non-redundancy of the measured outputs ensures the existence of a locally-invertible

state transformation of the form

 21
⎡η ⎤ ⎡ Qx ⎤
⎢ y ⎥ = Τ( x) = ⎢ h( x) ⎥
⎢ ⎥ ⎢ ⎥
⎢⎣Y ⎥⎦ ⎢⎣ K ( x) ⎥⎦

T
where η = ⎡⎣η1 ,… ,η n − m − q ⎤⎦ , and Q is a constant ( n - m − q ) × n matrix which for the sake

of simplicity, is chosen such that:

a) Each row of Q has only one nonzero term equal to one; and

b) Locally

⎧ ⎡ Qx ⎤ ⎫
⎪∂ ⎪
rank ⎨ ⎢ h( x) ⎥ ⎬ = n
⎢ ⎥
⎪ ∂u ⎢ K ( x) ⎥ ⎪
⎩ ⎣ ⎦⎭

Thus, the state transformation [η y Y ] = Τ( x) is locally invertible. The system of

T

(2.1), in terms of the new state variables, η1 ,… ,ηn − m − q , y, Y, takes the form

η = Fη (η , y, Y , u )
y = Fy (η , y, Y , u ) (2.7)
Y = FY (η , y, Y , u )

where

Fη (η , y, Y , u ) = Qf ⎡⎣Τ−1 (η , y, Y ), u ⎤⎦

∂h( x)
Fy (η , y, Y , u ) = f ⎡⎣Τ−1 (η , y, Y ), u ⎤⎦
∂x x =Τ−1 (η , y ,Y )

∂K ( x)
FY (η , y, Y , u ) = f ⎡⎣Τ −1 (η , y, Y ), u ⎤⎦
∂x x =Τ−1 (η , y ,Y )
 22
A closed-loop, reduced-order observer is then designed in the form:

z = Fη ( z + L1 y + L2Y , y, Y , u ) − L1Fy ( z + L1 y + L2Y , y, Y , u) − L2 FY ( z + L1 y + L2Y , y, Y , u )

xˆ = Τ−1 ( z + L1 y + L2Y , y, Y )

(2.8)

where the constant [ (n − m − q) × m ] and [ (n − m − q) × q ] matrices, L1 and L2 , are the

observer gains. The observer gains should be set such that the observer error dynamics

are asymptotically stable, which requires all eigenvalues of the [(n − m − q ) × (n − m − q )]

matrix

∂Fη (η , y, Y , u ) ∂Fy (η , y, Y , u ) ∂FY (η , y, Y , u )

− L1 − L2
∂η ∂η ∂η

evaluated at the nominal steady-state pair, to be in the left half of the complex plane.

2.3 Nonlinear Control Method

A state feedback that induces approximately linear responses to the state variables is first

derived. A reduced-order state observer is then designed to reconstruct unmeasured state

variables from the output measurements. To add integral action to the state feedback-state

observer system, a dynamic subsystem is finally added.

2.3.1 State Feedback Design

For a process in the form of (2.1) with input constraints:

uli ≤ u ≤ uhi , i = 1, ,m

a linear response is requested of the following form for each of the state variables:

(ε1 D + 1) p1 x1 = xssN
1

(2.9)
(ε n D + 1) xn = xssN
pn
n
 23
where p1 ≥ r1 , , pn ≥ rn , and ε1 , , ε n are positive constants that set the speed of the

state responses. The state responses in (2.9) can be achieved only when m ≥ n . However,

since in many processes m < n (there are more state variables than manipulated inputs),

the state responses in (2.9) can rarely be achieved. To relax the requirement of linear

responses for all state variables, state responses are requested that are as close as possible

to the linear ones described by (2.9). To derive a state feedback that achieves relaxed

state responses, the following constrained optimization problem is solved at each time

instant:

2
n⎡ (ε i D + 1) pi xi − xssi ⎤
min ∑ θi ⎢ ⎥ (2.10)
u
i =1 ⎣⎢ ε i pi ⎥⎦

subject to:

ui (l ) = 0, l ≥ 1, i = 1, ,m
uli ≤ u ≤ uhi , i = 1, ,m

where θ1 ,..., θ n are adjustable, positive, scalar weights whose values are set according to

the relative importance of the state variables: the higher the value of θi , the smaller the

deviation of the xi response from the desired linear response for xi .

For a process in the form of (2.1), the optimization problem in (2.10) takes the form:

2
⎡ ri −1
l ⎛ pi ⎞
pi
⎛p ⎞ ⎤
⎢ i ∑ i⎜ ⎟ i
+ ε + ∑ ε il ⎜ i ⎟ H il ( x, u, 0,… , 0) − xssNi
l
x H ( x ) ⎥
⎝l ⎠ ⎝l ⎠
n
min ∑ θ i ⎢ ⎥
l =1 l = ri
(2.11)
u
i =1
⎢ εi ip
⎥
⎢ ⎥
⎣ ⎦

subject to

uli ≤ u ≤ uhi , i = 1, ,m
 24
where

⎛ pi ⎞ pi !
⎜ ⎟=
⎝ l ⎠ ( pi − l )!l !

In the case that m ≥ n , the performance index in (2.10) may take the value of zero,

implying that the linear, closed-loop, state responses of (2.9) are achieved. The preceding

state feedback is denoted in the compact form:

u = Ψ ( x, xssN ) (2.12)

Example: Linear Unstable, Non-Minimum-Phase Process

Consider the linear process:

dx1
= x2
dt
dx2
= 10 x1 + 9 x2 + u
dt
y = 2 x1 − x2

This process is unstable (has poles at -1 and 10), and non-minimum-phase (has a

transmission zero at 2). The relative orders of the states, x1 and x2 , are: r1 = 2 and r2 = 1.

For ε 1 = 0.8 and ε 2 = 0.01, the closed-loop eigenvalues of this process under the state

feedback in (2.11) are given in Table 2.1. The state feedback of (2.11) is capable of

operating the process at any steady state when p1 and p2 are chosen such that p1 ≥ 2 and

p2 ≥ 1 .
 25
Table 2.1 Closed-loop eigenvalues of the linear example for several p1 and p2 values.

p1 p2 λ1 λ2
1 1 -100.00 0.000
1 2 -47.89 0.000
1 3 -30.59 0.000
1 4 -21.99 0.000
1 5 -16.86 0.000
2 1 -1.44 -1.090
2 2 -1.52 -1.020
2 3 -1.60 -0.970
2 4 -1.67 -0.930
2 5 -1.73 -0.900
3 1 -1.00 -0.150
3 2 -1.00 -0.153
3 3 -1.00 -0.153
3 4 -1.00 -0.153
3 5 -1.00 -0.153
4 1 -1.00 -0.017
4 2 -1.00 -0.017
4 3 -1.00 -0.017
4 4 -1.00 -0.017
4 5 -1.00 -0.017

2.3.2 Reduced-Order State Observer

In general, measurements of all state variables are not available. In such cases, estimates

of the unmeasured state variables can be obtained from the output measurements. Here,

we use the ‘closed-loop’, reduced-order, nonlinear, state observer of (2.8) to reconstruct

 26
the unmeasured state variables. The state observer is applicable to both stable and

unstable processes.

2.3.3 Integral Action

To ensure offset-free response of the closed-loop system in the presence of constant

disturbances and model errors, the control system should have integral action. An

estimate of disturbance-free process outputs is first calculated by using the closed-loop

process model:

w = f ( w, Ψ ( w, xssN ))
(2.13)
ξ = h( w)

where ξ is the estimate of the disturbance-free controlled output. The difference between

this estimate and the measurement of the controlled outputs, y, is then added to the output

set-point, leading to:

xssN = F ( ysp + (h( w) − y ) ) (2.14)

An interesting feature of this approach to adding integral action is that the addition of the

dynamic system of (2.13) to the state feedback of (2.12) [calculation of xssN according to

(2.14)] adds no additional conditions for closed-loop asymptotic-stability. In other words,

if the closed-loop system is asymptotically stable under the state feedback of (2.12)

alone, it is asymptotically stable under (2.12), (2.13) and (2.14).

2.3.4 Controller System

Combining the state feedback of (2.12), the reduced-order observer of (2.8), and the

dynamic sub-system of (2.13) and (2.14), leads to the following controller system that has

integral action:
 27
z = Fη (ηˆ, y, Y , u ) − L1 Fy (ηˆ , y, Y , u ) − L2 FY (ηˆ , y, Y , u )
ηˆ = z + L1 y + L2Y
xˆ = Τ−1 (ηˆ, y, Y )
w = f ( w, Ψ ( w, v)) (2.15)
v = F ( ysp − y + h( w) )
u = Ψ ( xˆ, v)

The control system parameters, ε1 , , ε n , set the speed of the closed-loop state responses;

the smaller the value of ε i , the faster the xi response. When the process is to operate at a

minimum-phase steady state, choosing p1 = r1 ,… , pn = rn is adequate to ensure the

asymptotic stability of the closed-loop system. When the process is to operate at a non-

minimum-phase steady state, one should choose p1 > r1 ,… , pn > rn . A block diagram of

the controller system is shown in Figure 2.1.

Y =K(x)
Y
ysp Θ v u = Ψ(xˆ, v) u
v = F(Θ) x = f (x,u)
y
y = h( x)
x̂

z = Fη (ηˆ, y,Y, u) − LF
1 y (η, y,Y, u) − L2FY (η, y,Y, u)
ˆ ˆ
ηˆ = z + L1y + LY
2

xˆ = T−1(ηˆ, y,Y)

ξ
w = f (w, Ψ(w, v)) ξ = h( w)

Figure 2.1 Parameterized controller system.

 28
2.4 Illustrative Examples

2.4.1 Single-Input Single-Output Chemical Reactor

Consider the constant-volume, non-isothermal, continuous, stirred-tank reactor shown in

Figure 2.2, in which the reaction A → B takes place in the liquid phase. The reactor

model has the form:

dC A ⎛ E ⎞ F
= − s exp ⎜ − a⎟ C A + (C Ai − C A )
dt ⎝ RT ⎠ V
dT ⎛ E ⎞ F
= γ s exp ⎜ − a ⎟ C A + (Ti − T ) + Q
dt ⎝ RT ⎠ V

where F is the volumetric flow rate of the reactor feed and product streams, V is the

reactor volume, and C Ai is the concentration of A in the feed stream. The reactor

parameter values are given in Table 2.2. This reactor has multiple steady states.

F, CAi

CA T
F, CA
Figure 2.2 Schematic of the non-isothermal continuous stirred tank reactor.
 29
Table 2.2 Values of the parameters of the non-isothermal reactor model.

Parameter Value Unit

s 5.0 × 108 s −1
Ea / R 8100 K
γ 3.9 m3 K kmol −1
Q −2.519 × 10 −2 K s −1
C Ai 12 kmol m −3
Ti 300 K
V 0.1 m3

The reactor temperature, T , is the controlled output; it is desired to operate the reactor at

its middle (unstable) steady state by manipulating the feed flow rate, F . The operating

range of the feed flow rate is 0.2 ≤ F ≤ 1.5 m3 / h . Here, x = [C A T ] , u = [ F ] , and

T

y = [T ] .

The steady-state pair corresponding to ysp = 302 is ( x1ss = 6.319 , x2 ss = 302 ,

uss = 0.45 ) which is unstable (eigenvalues of the process Jacobian evaluated at the

steady-state pair are -4.5 and 0.309). The relative orders of the state variables are both

unity ( r1 = r2 = 1 ). The zero dynamics of this process are governed by:

d ζ ⎪⎧ C A − ζ ⎫⎪ ⎛ − Ea ⎞ CA − ζ
= ⎨ −1 − γ i ⎬ s exp ⎜⎜ ⎟⎟ ζ − i Q
dt ⎪⎩ Ti − ysp ⎭⎪ ⎝ Rysp ⎠ Ti − ysp

whose Jacobian evaluated at the desired steady state has an eigenvalue in the right half of

the complex plane at 36.27. Thus, the middle steady state is unstable and non-minimum-

phase.
 30

Controller System

For this process, the controller system of (2.15) with θ1 = ε12 , θ 2 = ε 22 , p1 = 2 , and p2 = 2 ,

takes the form:

{
u = Ψ ( xˆ, v) = arg min ⎡⎣(−v1 + A10 + A11u + A12u 2 ) 2 + (−v2 + A20 + A21u + A22u 2 ) 2 ⎤⎦
u
}
subject to:

0.2 ≤ u ≤ 1.5

where

ε1 xˆ1s ⎧⎪ ⎛ E ⎞ ⎛ 2E ⎞ ⎫⎪
A10 = xˆ1 + 2 ⎨
−(ε1QEa + 2 Rxˆ22 ) exp ⎜ − a ⎟ + ε1s (− Eaγ xˆ1 + Rxˆ22 ) exp ⎜ − a ⎟⎬
Rxˆ2 ⎪⎩ ⎝ Rxˆ2 ⎠ ⎝ Rxˆ2 ⎠ ⎪⎭

ε1 (2C A − 2 xˆ1 ) ε12 s ⎛ E ⎞

A11 = i
+ {
exp ⎜ − a ⎟ − R(C Ai − 2 xˆ1 ) xˆ22 + Ea xˆ1 ( xˆ2 − Ti )
2 }
V VRxˆ 2 ⎝ Rxˆ2 ⎠

ε12 (C A − xˆ1 )
A12 = − i

V2

⎪⎧ ⎛ −E ⎞ ⎪⎫ ε 22γ sxˆ1 ⎛ 2E ⎞ ⎪⎧ ⎛ Ea ⎞ ⎫⎪
A20 = x2 + 2ε 2 ⎨Q + γ xˆ1s exp ⎜ a ⎟⎬ + 2
exp ⎜ − a ⎟ ⎨QEa exp ⎜ ⎟ + xˆ1sγ Ea − xˆ2 Rs ⎬
2

⎩⎪ ⎝ Rxˆ2 ⎠ ⎭⎪ Rxˆ2 ⎝ Rxˆ2 ⎠ ⎩⎪ ⎝ Rxˆ2 ⎠ ⎭⎪

ε 2 (2Ti − 2 xˆ2 − ε 2Q) ε 22 γ s ⎛ E ⎞

A21 = + {
exp ⎜ − a ⎟ (Ti − xˆ2 ) xˆ1 Ea + (C Ai − 2 xˆ1 ) Rxˆ22
2 }
V VRxˆ 2 ⎝ Rxˆ2 ⎠

ε 22 (Ti − xˆ2 )
A22 = −
V2

xˆ1 = z + Ly
xˆ2 = y

⎧ ⎛ E ⎞ u⎫ ⎧ ⎛ E ⎞ u ⎫
z = ⎨− s exp ⎜ − a ⎟ ( z + Ly ) + (C Ai − z − Ly ) ⎬ − L ⎨γ s exp ⎜ − a ⎟ ( z + Ly ) + (Ti − y ) + Q ⎬
⎩ ⎝ Ry ⎠ V⎭ ⎩ ⎝ Ry ⎠ V ⎭
 31
⎛ E ⎞ Ψ ( w, v)
w1 = − s exp ⎜ − a ⎟ x1 + (C Ai − w1 )
⎝ Rw2 ⎠ V
⎛ E ⎞ Ψ ( w, v)
w2 = γ s exp ⎜ − a ⎟ x1 + (Ti − w2 ) +Q
⎝ Rw2 ⎠ V

⎛ E ⎞ ⎛ ⎪⎧ ⎛ E ⎞ ⎪⎫ ⎛ E ⎞ ⎞
exp ⎜ a ⎟ QV − sV (CAi γ + Ti − v2 ) + V 2 ⎜ ⎨exp ⎜ a ⎟ Q + s(CAi γ + Ti − v2 )2 ⎬ + 4exp ⎜ a ⎟ sQ(v2 − Ti ) ⎟
⎜ ⎪ ⎝ Rv2 ⎠ ⎟
⎝ Rv2 ⎠ ⎝⎩ ⎭⎪ ⎝ Rv2 ⎠ ⎠
v1 = −
2 γ sV
v2 = ysp − xˆ2 + w2

The following controller parameter values are used: ε 1 = 468 s , ε 2 = 468 s , and L = 20 .

The closed-loop eigenvalues of this process under the state feedback of (2.11), for several

p1 and p2 values, are given in Table 2.3. MATLAB optimization toolbox is used to solve

the constrained optimization problem of the controller system. In this optimization, the

number of optimizing variables is equal to the number of manipulated inputs. Therefore,

the computational cost of this controller is much less than that of a typical model-

predictive controller.

Controller Performance

Servo and regulatory responses of the controller are shown in Figures 2.3 and 2.4. Figure

2.3 depicts the evolution of the process state variables for two initial conditions,

[ x1 (0) , x2 (0) ] = [9.162, 290] in the minimum-phase region, and [ x1 (0) , x2 (0) ] =

[2.923, 320] in non-minimum-phase region, with C A = 12 kmol / m3 . The set-point,

i

ysp = 302 , corresponds to ( x1ss = 6.319, x2 ss = 302) . In addition, when the process is at

the desired set-point, the regulatory performance of the controller is studied. A step

change from 12 to 10 kmol / m3 in C Ai is made at t= 2 hr (unmeasured disturbance). The

simulation results show that the controller successfully operates the reactor at the desired
 32
steady state, which is non-minimum-phase and open-loop unstable. The controller is

capable of operating the process at the desired steady state, regardless of the initial

conditions of the process. The integral action of the controller ensures offset-free

response in the presence of the unmeasured disturbance.

Table 2.3 Closed-loop eigenvalues of the CSTR example for several p1 and p2 values.

p1 p2 λ1 λ2
1 1 7.1628 -3.3332
1 2 -3.1375 - 0.5928 i -3.1375 + 0.5928 i
1 3 -4.264 -1.4298
2 1 16.118 -3.024
2 2 -3.4071 + 0.7204 i -3.4071 - 0.7204 i
2 3 -4.3881 -1.5140
3 1 0.0003 -4.9778
3 2 -3.4752 + 0.3415 i -3.4752 - 0.3415 i
3 3 -4.3489 -1.5317
 33

10

initial condition1
8
initial condition2
x1 (kmol m )
-3

320 initial condition1

ysp
initial condition2

310
x2 (K)

300

290
0 2 4 6
time (hr)

Figure 2.3 Closed-loop response of the state variables of the chemical reactor
 34

1.6

initial condition1
initial condition2
1.2
u (m h )
-1
3

0.8

0.4

0.0
13
CAi (kmol m )
-3

11

9
0 2 4 6
time (hr)

Figure 2.4 Manipulated input and unmeasured disturbance profiles corresponding to

Figure 2.3
 35
2.4.2 Single-Input Single-Output Bioreactor

Consider the constant-volume, continuous bioreactor described in [7]. The process

dynamics are described by:

x2
x1 = − x1u + x1 (1 − x2 ) exp( )
γ1
x2 1 + β1
x2 = − x2u + x1 (1 − x2 ) exp( )
γ 1 1 + β1 − x2
y = x1

where x1 is the dimensionless cell-mass concentration in the bioreactor, x2 is the

dimensionless substrate concentration, and u is the dilution rate. The model parameters

are β1 = 0.02 and γ 1 = 0.48 . It is desired to maintain the cell-mass concentration, x1 , at a

set-point by manipulating the dilution rate, u , within the operating range, 0.4 ≤ u ≤ 1.0 .

The process steady-state pair corresponding to ysp = 0.1448 ( x1ss = 0.1448, x2 ss = 0.8455,

uss = 0.9) is unstable (eigenvalues of the open-loop Jacobian evaluated at the steady state

are 0.061 + 1.731 i and 0.061 − 1.731 i ). The relative orders of the state variables are:

r1 = 1 and r2 = 1 . The zero dynamics of the process is governed by:

dζ ζ ζ 1 + β1
= − yspζ (1 − ζ ) exp( ) + ysp (1 − ζ ) exp( )
dt γ1 γ 1 1 + β1 − ζ

whose Jacobian evaluated at the steady state has an eigenvalue in the right half of the

complex plane at 1.374. Thus, the desired steady state is unstable and non-minimum-

phase.

Controller System

For this process, the controller system of (2.14), with θ1 = ε1 , θ 2 = ε 2 , p1 = 1 and p2 = 1 ,

takes the form:

 36

{ }
u = Ψ ( xˆ, v) = arg min ⎡⎣(−v1 + A10 + A11u )2 + (−v2 + A20 + A21u )2 ⎤⎦
u

subject to:

0.4 ≤ u ≤ 1.0

where

xˆ2
A10 ( xˆ ) = xˆ1 + ε1 xˆ1 (1 − xˆ2 ) exp( )
γ1
A11 ( xˆ ) = − xˆ1ε1
xˆ2 1 + β1
A20 ( xˆ ) = xˆ2 + ε 2 xˆ1 (1 − xˆ2 ) exp( )
γ 1 1 + β1 − xˆ2
A21 ( xˆ ) = − xˆ2ε 2

xˆ1 = y
xˆ2 = z + Ly

⎧ z + Ly 1 + β1 ⎫
z = ⎨−( z + Ly )u + x1 {1 − z − Ly} exp( ) ⎬
⎩ γ 1 1 + β1 − z − Ly ⎭
⎧ z + Ly ⎫
− L ⎨− x1u + x1 {1 − z − Ly} exp( )⎬
⎩ γ1 ⎭
w
w1 = − w1q + w1 (1 − w2 ) exp( 2 )
γ1
w2 1 + β1
w2 = − w2 q + w1 (1 − w2 ) exp( )
γ1 1 + β1 − w2

ν 1 = ysp − xˆ1 + w1
ν 2 = F (v1 )

q = Ψ ( w, v)

The following controller parameter values are used: ε 1 = 0.001, ε 2 = 0.15 , and L =0.01.

The closed-loop eigenvalues of this process under the state feedback of (2.11), for several

p1 and p2 values, are given in Table 2.4.

 37
Table 2.4 Closed-loop eigenvalues of the bioreactor example for several p1 and p2
values.

p1 p2 λ1 λ2
1 0 -10 3.4557
1 1 -2.91 -0.85
1 2 -2.91 -0.83
2 1 -2.91 -0.83
2 2 -2.91 -0.83

Controller Performance

Figures 2.5 and 2.6 show the servo and regulatory responses of the controller. For two

initial conditions, [ x1 (0), x2 (0)] = [0.1, 0.75] and [0.25, 0.2] , the process state profiles are

depicted in Figure 2.5. The set-point, ysp = 0.1448, corresponds to the steady state ( x1ss =

0.1448, x2 ss = 0.8453). Figure 2.5 shows that the controller successfully operates the

bioreactor at the desired steady state, irrespective of the initial conditions of the process.

To evaluate the regulatory performance of the controller, a step change from 0.02 to

0.022 in β1 of the process (unmeasured disturbance) was made at t=8 hr when the

process was at the desired steady state. As Figure 2.5 shows, the controller rejects the

unmeasured disturbance asymptotically.

 38

0.5

initial condition1
0.4 ysp
initial condition2

0.3
x1

0.2

0.1

0.0
1.0

0.8

0.6 initial condition1

x2

initial condition2

0.4

0.2
0 10 20 30
time (hr)

Figure 2.5 Closed-loop response of the bioreactor state variables.

 39

1.1

u 0.9

0.7

initial condition1
0.5 initial condition2

0.3
0.023

0.022
β1

0.021

0.020

0.019
0 10 20 30
time (hr)

Figure 2.6 Manipulated input and unmeasured disturbance profiles corresponding to

Figure 2.5.
 40
2.4.3 Multi-Input, Multi-Output Chemical Reactor

Consider the constant-volume, non-isothermal, continuous, chemical reactor shown in

Figure 2.7, in which the series reactions, A → B → C , take place in the liquid phase. The

process dynamics are represented by the following model:

dC A F E
= (C Ai − C A ) − s1 exp(− 1 )C A2
dt V RT
dCB F E E
= − CB + s1 exp(− 1 )C A2 − s2 exp(− 2 )CB
dt V RT RT
dT F ( −Δ H ) E (−ΔH 2 ) E US
= (Ti − T ) + 1
s1 exp(− 1 )C A2 + s2 exp(− 2 )CB + (T j − T )
dt V ρ cp RT ρ cp RT ρ c pV
dT j Fj US
= (T ji − T j ) − (T − T )
dt V ρ j cp j V j j
y1 = CB
y2 = T

The reactor parameter values are given in Table 2.5. It is desired to maintain CB and T at

set-point values by manipulating F and Fj within the operating ranges, 10 ≤ F ≤ 150 l / h

and 10 ≤ Fj ≤ 150 l / h . In this process, only the state variable, C A , is not measured.

F, CAi , Ti
Fj , Tji

Tj
C A , CB , T

Figure 2.7 Schematic of the multivariable non-isothermal chemical reactor.

 41
Table 2.5 Process parameters of the non-isothermal reactor with cooling jacket.

Parameter Value Unit

F 80.95 l hr −1
Fj 100 l hr −1
C Ai 12 mol l −1
s1 2.5 × 1010 lmol −1hr −1
s2 1.5 × 1010 hr −1
− E1 / R 8000 K
− E2 / R 9100 K
Ti 320 K
T ji 298.15 K

−ΔH1 −20 kJ mol −1

−ΔH 2 −80 kJ mol −1
ρ 1 kg l −1
ρj 1.1 kg l −1

cp 2.25 kJ kg −1 K −1

cp j 3 kJ kg −1 K −1

U 3825 kJ m −2 K −1s −1
S 0.225 m2
V 5 l
Vj 5 l
 42
Here, x = [C A CB T T j ] , u = [ F Fj ] ,
T T
y = [CB T ]
T
and Y = [T j ] . The steady state

corresponding to the set-points, ysp1 = 5.233 and ysp2 = 443.92 , is ( x1ss = 0.701 ,

x2 ss = 5.233 , x3 ss = 443.92 , x4 ss = 403.24 ), which is unstable (eigenvalues of the

Jacobian of the process evaluated at the steady state are −491.65 , −90.71 , 86.29 , and

−15.36 ). The relative orders of the state variables are: r1 = 1, r2 = 1, r3 = 1 , and r4 = 1 .

The zero dynamics of this process is governed by:

dζ 1 1 ⎡ E1 E ⎤ E
= ⎢ s1 exp(− )ζ 12 − s2 exp(− 2 ) y1sp ⎥ (C A0 − ζ 1 ) − s1 exp(− 1 )ζ 12
dt y1sp ⎣⎢ Ry2 sp Ry2 sp ⎦⎥ Ry2 sp

whose Jacobian evaluated at the steady state has eigenvalues at 589.59 . Thus, the desired

steady state is unstable and non-minimum-phase. For this process, the controller system

(2.15), with θ1 = ε12 , θ 2 = ε 22 , θ3 = ε 32 , θ 4 = ε 4 , p1 = 2 , p2 = 2 , p3 = 2 and p4 = 1 , is

implemented, with 10 ≤ u1 ≤ 150 and 10 ≤ u2 ≤ 150 . The following controller parameter

values are used: ε1 = 0.17 , ε 2 = 0.17 , ε 3 = 0.17 , ε 4 = 0.17 , L11 = −0.03 , L12 = −0.03 ,

and L21 = 0 .

Controller Performance

It is desired to maintain the controlled outputs at ysp1 = 5.233 and ysp2 = 443.92 , which

correspond to the steady state ( x1ss = 0.701 , x2 ss = 5.233 , x3 ss = 443.92 , x4 ss = 403.24 ) .

Figures 2.8, 2.9 and 2.10 show the servo and regulatory responses of the controller. The

process state variable profiles for two initial conditions, [ x1 (0), x2 (0), x3 (0), x4 (0)]

= [1.814, 9.446, 387.69, 380.67 ] and [ 0.519, 3.5, 455, 410] , are shown in Figures 2.8
 43
and 2.9; the controller is capable of operating the process at the unstable non-minimum-

phase steady state. After the process reached the desired steady state, a step change from

12 kmol / m3 to 11.8 kmol / m3 in C Ai (unmeasured disturbance) was made at t=8 hr. From

Figures 2.8 and 2.9, it can be seen that the controller provides offset-free responses in the

presence of the unmeasured disturbance. The corresponding manipulated input profiles

are shown in Figure 2.10.

2.5 Conclusions

A control law is presented that is applicable to general multivariable processes, whether

minimum- or non-minimum-phase. The controller forces process state variables to follow

their corresponding reference trajectories. The state feedback is obtained by requesting

state responses closest to a set of desired linear state responses. The integral action

ensures offset-free, closed-loop, output response in the presence of constant disturbances

and process-model mismatch. A closed-loop, reduced-order observer is used to estimate

the unmeasured state variables. Compared to a general, multivariable, model-predictive

controller, this control system has less tuning parameters to achieve closed-loop,

asymptotically stability. The control system does not have the limitation of the control

method presented by Kanter et al. [1]. However, because of the optimization (numerical)

and shortest-prediction-horizon forms of the controller system, analytical proof of

asymptotic, closed-loop stability of the closed-loop system remains an open problem.

 44

2.0
initial condition1
initial condition2
1.6
x1 (mol l )
-1

1.2

0.8

0.4
11
initial condition1
ysp1
9 initial condition2
x2 (mol l )
-1

3
0 4 8 12
time (hr)

Figure 2.8 Closed-loop response of the first two state variables of the chemical reactor.
 45

460

440
x3 (K)

420

400
initial condition1
ysp2
initial condition2
380
420

410
x4 (K)

400

390

initial condition1
initial condition2
380
0 4 8 12
time (hr)

Figure 2.9 Closed-loop response of the last two state variables of the chemical reactor.
 46

100

u1 (l h )
-1

60

initial condition1
initial condition2
20
120

80
u2 (l h )
-1

40

initial condition1
initial condition2
0
0 4 8 12
time (hr)

Figure 2.10 Manipulated input profiles corresponding to Figures 2.8 and 2.9.
 47
2.6 Notation

A, B, C Chemical species
cp Heat capacity of feed and product, kJ kg −1 K −1

cpj Heat capacity of jacket fluid, kJ kg −1 K −1

C Ai −3
Inlet concentration of the reactant, kmol m
CA −3
Outlet concentration of the reactant, kmol m
CB Outlet concentration of B , mol l −1
D Differential operator, D = d/dt.
Ea , E1 , E2 Activation energy, kJ mol −1
F Feed flow rate to reactor
Fj Jacket coolant flow rate, l h −1
J ol Open-loop Jacobian
J cl Closed-loop Jacobian
L1 , L2 Observer gains
m Number of manipulated inputs and controlled outputs
n Number of state variables
ri Relative order of state variable xi

R Universal gas constant, kJ kmol −1 K −1

s, s1 , s2 Pre-exponential factor, s-1, l mol-1hr-1, hr-1
S Heat transfer surface area, m 2
t Time, s
T Reactor outlet temperature, K
Ti Reactor inlet temperature, K
Tj Jacket temperature, K

T ji Jacket inlet temperature, K

u Vector of manipulated inputs
 48
U −2 −1 −1
Overall heat-transfer coefficient, kJ m K s
V Reactor volume, m3
Vj Jacket volume, m3
x Vector of state variables
y Vector of controlled outputs
ysp Vector of set-points

−ΔH1 , − ΔH 2 Heat of reaction, kJ mol −1

Greek

ε 1 ,… , ε n Adjustable parameters of controller

w Vector of the controlled state variables

γ Reactor model parameter, K m3 kmol -1
ζ Vector of the state variable of zero dynamics
ρ Density of mixture in reactor, kg l −1
ρj Density of jacket fluid, kg l −1

Superscripts

^ Estimate

Subscripts

A, B Chemical species
ss Steady State
sp Set-point
 49
Chapter 3: Input-Output Linearization with Stability Constraint

3.1 Introduction

Output tracking has been studied very extensively. The control problem is to design a

controller that forces the controlled output(s) to asymptotically track reference

trajectory(ies). Since the early 1990’s, many research efforts have been made to solve the

problem of output tracking of non-minimum-phase (NMP) nonlinear systems. These

efforts have been mainly within the frameworks of model-predictive control and

differential geometric control.

In model-predictive control (MPC), controller action is the solution to a

constrained optimization problem that is solved repeatedly on-line. In MPC, global

optimality may not imply closed-loop stability. To ensure closed-loop stability in MPC, it

has been proposed to add Lyapunov stability constraints or penalty terms to the

optimization problem. For example, terminal state constraints in [34, 36, 38], terminal

state penalty in [35], terminal inequality constraint in [35], and contractive constraint in

[40] have been used. In non-minimum-phase systems, output tracking through MPC

requires longer prediction horizon(s) than in minimum-phase systems.

Differential geometric control is a direct synthesis method in which the controller

is derived by requesting a desired closed-loop output response in the absence of input

constraints. A widely used differential geometric control method is input-output

linearization, which cannot be used to operate a process at a NMP steady state. Efforts to

make input-output linearization applicable to processes with a NMP steady state include

the use of equivalent output(s) for the controller design [82-84], coordinated control [30],
 50
controller design by inverting the minimum-phase part [2, 25], and approximate input-

output linearization [1, 85, 86]. Other methods that are applicable to nonlinear systems

with a non-minimum-phase steady state include stable inversion [23, 24, 26], H ∞ control

[27, 33], and total linearization [48].

Another popular control framework is Lyapunov-based control, in which the

central focus in the controller design is on stability through Lyapunov’s direct method.

The asymptotic decay of a norm of the state variables is ensured by the use of a proper

Lyapunov function in the controller design.

This chapter presents a new method that addresses a major limitation of input-

output linearization. The proposed method is applicable to stable and unstable processes,

whether non-minimum- or minimum-phase. A novel Lyapunov stability constraint that

does not suffer from the singularity (at the desired steady state) problem of the standard

Lyapunov constraints is presented. The control method is obtained by further exploiting

the connections between input-output linearization and model-predictive control [46, 80,

87]. It performs optimal output tracking via input-output linearization and guarantees

closed-loop asymptotic stability within an assessable domain of attraction by satisfying a

hard Lyapunov stability constraint. This hard constraint requires all state variables to

evolve within a shrinking state tube. Whenever output tracking alone is unable to ensure

closed-loop asymptotic stability, the closed-loop system evolves while being at the hard

constraint. Upon the arrival of the closed-loop system in a state-space region in which

output tracking alone can ensure asymptotic stability, the hard constraint becomes

inactive.
 51
This organization of this chapter is as follows. The scope of the study and some

mathematical preliminaries are given in Section 3.2. Section 3.3 presents the nonlinear

control method. The application and performance of the control method are illustrated by

numerical simulation of two examples in Section 3.4.

3.2 Scope and Mathematical Preliminaries

Consider the general class of multivariable processes with a mathematical model in the

form:

x = f ( x, u ), x (0) = x0
(3.1)
y = h( x )

with the input constraints

uli ≤ ui ≤ uhi , i = 1, ,m

where x ∈ X ⊂ R n is the vector of state variables, u ∈ U ⊂ R m is the vector of

manipulated inputs, y ∈ R m is the vector of controlled outputs, and f ( x, u ) and h( x) are

smooth functions on X × U and X , respectively. X is an open connected set that

contains the nominal steady state value of x . U is a closed connected set that contains

the nominal steady state value of u : U={u| uli ≤ ui ≤ uhi , i = 1, , m }. The relative

orders (degrees) of the controlled outputs y1 ,… , ym are denoted by R1 ,… , Rm ,

respectively, where Ri is the smallest integer for which ∂ ⎣⎡ d Ri yi / dt Ri ⎦⎤ / ∂u ≠ 0 . The

following notation will be used:

 52
h ( x) = hi ( x) = yi
i
0

dyi
hi1 ( x) =
dt
(3.2)
d Ri −1 yi
hiRi −1 ( x) =
dt Ri −1
d Ri yi
hiRi ( x, u ) =
dt Ri

Assumptions:

A1. The relative orders (degrees), R1 … Rm , are finite.

A2. The characteristic (decoupling) matrix of the process is nonsingular on X × U .

A3. The process is controllable and observable on X × U .

The steady state pair(s) ( xss , uss ) corresponding to a given output setpoint, ysp , satisfy:

0 = f ( xss , uss )
ysp = h( xss )

These relations are used to describe the dependence of a nominal steady state, xssN , on the

set point: xssN = F ( ysp ) .

3.2.1 MPC Formulation of Input-Output Linearization

This section presents a very short review of the MPC formulation of input-output

linearization that can be found in detail in [46, 80, 87]. This formulation will be used in

the main body of this paper.

A linear response of the following form:

(ε1 D + 1) R1 y1 = ysp1
(3.3)
(ε m D + 1) ym = yspm
Rm
 53
where D = d / dt , and ε 1 ,… , ε m are positive constants that set the speed of the closed-loop

output responses, can be induced to an unconstrained process of the form (3.1) by

implementing the solution to the following unconstrained optimization problem at each

time instant [46, 80, 87]:

m
min ∑ yˆi (τ ) − yd (τ )
2
(3.4)
u (t ) i
i =1

where q(τ ) denotes the 2-norm of a scalar function q (τ ) , given by:

t +Th
q(τ ) ∫t
q (τ ) 2 dτ

yˆi (τ ) is model-predicted future value of yi at time τ , given by a Taylor series expansion

of yi around the present time, t:

Ri −1
[τ − t ]l [τ − t ]Ri
yˆ i (τ ) = hi ( x(t )) + ∑ hil ( x (t )) + hiRi ( x (t ), u (t )) + h.o.t.
l =1 l! Ri !

and ydi (τ ) is the predicted future value of the reference trajectory of yi at time τ , given

by:

(ε1 D + 1) R1 yd (τ ) = ysp11

(3.5)
(ε m D + 1) yd (τ ) = yspm
Rm
m

initialized at:

d l ydi (t )
l
= hil ( x(t )), l = 0, , Ri − 1, i = 1, ,m
dt

Taylor series expansion of ydi around the present time, t, yields:

 54
Ri −1
⎛ Ri ⎞ l
yspi (t ) − hi ( x(t )) − ∑ ε il ⎜ ⎟ hi ( x(t ))
Ri −1
[τ − t ] l
⎝l ⎠ [τ − t ]Ri
yd (τ ) = hi ( x(t )) + ∑ hil ( x(t )) +
l =1
i
l =1 l! ε i
Ri
Ri !
+h.o.t.

After substituting for yˆi (τ ) and ydi , for a very small prediction horizon Th , the

minimization problem of (3.4) takes the simplified form:

2
⎡ Ri −1
l ⎛ i ⎞ l
R ⎤
⎢ y sp − hi ( x ) − ∑ ε i ⎜ ⎟ hi ( x ) − ε i
Ri Ri
hi ( x , u ) ⎥
⎝ l ⎠
m i

min ∑ ⎢ l =1 ⎥
⎢ ⎥ (3.6)
u
i =1 εiRi

⎢ ⎥
⎣⎢ ⎦⎥

Equation (3.6) represents a state feedback, which is denoted by:

u = Ψ ( x, ysp )

If this state feedback makes the performance index in (3.4) zero at each time instant, the

linear closed-loop response of (3.3) is induced; that is, the state feedback of (3.6) is input-

output linearizing. Needless to mention that the closed-loop system under the input-

output linearizing state feedback of (3.6) is stable when the desired steady state is

minimum-phase.

3.2.2 Reduced-Order State Observer Design

This subsection presents a brief review of the nonlinear, reduced-order, state observer

described in [81]. The state observer will be used later in this paper to reconstruct

unmeasured state variables of the process under consideration. Consider a nonlinear

process in the form of (3.1) with additional output measurements Y1 , , Yq ; that is,
 55
dx
= f ( x, u ) x(0) = x0
dt
y = h( x )
Y = K ( x)

The non-redundancy of the measured outputs ensures the existence of a locally-invertible

state transformation of the form

⎡η ⎤ ⎡ Qx ⎤
⎢ y ⎥ = T ( x ) = ⎢ h( x ) ⎥
⎢ ⎥ ⎢ ⎥
⎢⎣Y ⎥⎦ ⎢⎣ K ( x) ⎥⎦

T
where η = ⎡⎣η1 , ,ηn − m − q ⎤⎦ , and Q is a constant (n − m − q ) × n matrix which for the sake

of simplicity, is chosen such that:

c) Each row of Q has only one nonzero term equal to one; and

d) Locally

⎧ ⎡ Qx ⎤ ⎫
⎪∂ ⎪
rank ⎨ ⎢ h( x) ⎥ ⎬ = n
⎢ ⎥
⎪ ∂u ⎢ K ( x) ⎥ ⎪
⎩ ⎣ ⎦⎭

Thus, the state transformation [η y Y ] = T ( x) is locally invertible. The system of

T

(3.1), in terms of the new state variables, η1 ,… ,ηn − m − q , y, Y, takes the form

η = Fη (η , y, Y , u )
y = Fy (η , y, Y , u ) (3.7)
Y = FY (η , y, Y , u )

where

Fη (η , y, Y , u ) = Qf ⎡⎣Τ−1 (η , y, Y ), u ⎤⎦

∂h( x)
Fy (η , y, Y , u ) = f ⎡Τ−1 (η , y, Y ), u ⎤⎦
∂x x =Τ−1 (η , y ,Y ) ⎣
 56
∂K ( x)
FY (η , y, Y , u ) = f ⎡⎣Τ −1 (η , y, Y ), u ⎤⎦
∂x x =Τ−1 (η , y ,Y )

A closed-loop, reduced-order state observer is then designed in the form:

z = Fη ( z + L1 y + L2Y , y, Y , u ) − L1 Fy ( z + L1 y + L2Y , y, Y , u ) − L2 FY ( z + L1 y + L2Y , y, Y , u )

xˆ = Τ−1 ( z + L1 y + L2Y , y, Y )

(3.8)

where the constant [ (n − m − q) × m] and [ (n − m − q) × q ] matrices, L1 and L2 , are the

observer gains. The observer gains should be set such that the observer error dynamics

are asymptotically stable, which requires all eigenvalues of the [ (n − m − q ) × (n − m − q )]

matrix

∂Fη (η , y, Y , u ) ∂Fy (η , y, Y , u ) ∂FY (η , y, Y , u )

− L1 − L2
∂η ∂η ∂η

evaluated at the nominal steady-state pair, to be in the left half of the complex plane.

3.3 Nonlinear Control Method

This section presents the nonlinear control method and discusses practical issues in the

implementation of the method.

3.3.1 Lyapunov Stability Constraint

The idea of designing an input-output linearizing controller with a stability constraint has

been inspired by contractive (stability) constraints used in model-predictive control to

ensure closed-loop stability [40] and that input-output linearization is a shortest

prediction horizon continuous-time model-predictive controller [46, 80, 87]. A widely

used stability constraint in model-predictive control is:

|| x (k + 1) ||≤ α || x (k ) ||, 0<α <1 (3.9)

 57
where x = x − x . The preceding stability constraint has the general form:
N
ss

V ( x (k + 1)) − αV ( x (k )) ≤ 0

where V ( x ) is a positive definite function. The preceding inequality has the continuous-

time form:

dV ( x )
+ γV (x ) ≤ 0 (3.10)
dt

1
where γ = − ln α . The manipulated input that makes V ( x ) satisfy (3.10) is the
Δt

solution for u of:

∂V ( x )
f ( x, u ) + γ V ( x ) ≤ 0
∂x

Since

∂V ( x )
=0
∂x x = xssN

as x → xssN , (3.10) becomes ill-conditioned/singular and thus cannot be solved for u .

Several approaches/approximations have been proposed to address this singularity

problem in Lyapunov control [53, 88-90].

Consider the specific Lyapunov function:

V ( x ) = x T Px (3.11)

where P is the positive-definite symmetric matrix that satisfies the Riccati equation

AT P + PA − PBT BP = −Q

Q is a positive definite matrix,

∂f ( x, u ) ∂f ( x, u )
A= B=
∂x ( xssN ,ussN ) ∂u ( xssN ,ussN )
 58
If the Lyapunov function, V ( x ) , satisfies

d 2V dV
β2 2
+ 2β +V ≤ 0 (3.12)
dt dt

in closed-loop over X, where β is a positive constant that set the rate of decay of V ( x ) ,

then the closed-loop system is asymptotically stable over X . If V ( x ) is required to be

governed by (3.12), then

⎡ ∂V ⎧ ∂f ∂f du ⎫⎤ ∂V
β 2 ⎢ f T Pf + ⎨ f+ ⎬⎥ + 2β f +V ≤ 0 (3.13)
⎣ ∂X ⎩ ∂X ∂u dt ⎭⎦ ∂X

which is not singular at x = xssN , allowing one to solve for u at every x ∈ X .

3.3.2 Nonlinear State Feedback Design

To derive a state feedback that can achieve output tracking with guaranteed closed-loop

stability, we solve the following constrained optimization problem at each time instant, t:

m
min ∑ yˆi (τ ) − yd (τ )
2
(3.14)
u (t ) i
i =1

subject to the process model of (3.1); that is,

x = f ( x, u ), x(0) = x0
y = h( x )

with uli ≤ ui ≤ uhi , i = 1, , m and the stability constraint:

β 2V + 2βV + V ≤ 0

where

V = ( x − xssN )T P( x − xssN ) ,

xssN = F ( ysp )
 59
Assuming this optimization is feasible (has a solution u ∈ U ), let us denote the solution

by the state feedback:

u = Ψ ( x, ysp ) (3.15)

The next theorem summarizes the stability properties of the preceding state feedback.

Theorem. For a process in the form of (3.1), the closed-loop system under the state

feedback (3.15) is asymptotically stable in the region in which the state feedback is

feasible.

The tunable parameter β is suggested to be chosen such that β > max(ε1 ,… , ε m ) . This

choice of β can prevent unnecessary activation of the stability constraint when output

tracking can ensure closed-loop stability.

3.3.2.1 Implementation of Nonlinear State Feedback

The state feedback of (3.15) is the solution of the constrained optimization problem:

2
⎡ Ri −1
l ⎛ i ⎞ l
R ⎤
⎢ y spi − hi ( x ) − ∑ ε i ⎜ ⎟ i
h ( x ) − ε i
Ri Ri
hi ( x , u ) ⎥
⎝ l ⎠
m
min ∑ ⎢ l =1 ⎥
⎢ ⎥ (3.16)
u
i =1 εiRi

⎢ ⎥
⎢⎣ ⎥⎦

subject to

uli ≤ ui ≤ uhi , i = 1, ,m

β 2V + 2β V + V ≤ 0

When the process is away from the nominal steady state xssN ; that is, V ( x − xssN ) ≥ σ ,
 60
where σ is a small positive constant that is set by the controller designer, the stability

constraint of (3.12) takes the form:

⎡ ∂V ⎧ ∂f ∂f du ⎫⎤ ∂V
β 2 ⎢ f T Pf + ⎨ f + ⎬⎥ + 2β f +V ≤ 0 (3.17)
⎣ ∂X ⎩ ∂X ∂u dt ⎭⎦ ∂X

which is denoted by the compact form:

S ( x, xss , u, u ) ≤ 0 (3.18)

When the process is close to the nominal steady state xssN ; that is, V ( x − xssN ) < σ , the

stability constraint of (3.12) takes the form:

β 2 f T Pf + V ≤ 0 (3.19)

which is denoted by:

Sε ( x, u ) ≤ 0 (3.20)

The time derivative of the manipulated input vector, u , can be approximated by

u (t ) = [u (t ) − u (t − Δt )]/ Δt , where Δt is time-discretization interval. Thus, the state

feedback of (3.15) is the solution of the constrained optimization problem:

2
⎡ Ri −1
⎛R ⎞ ⎤
⎢ y spi − hi ( x (t )) − ∑ ε il ⎜ i ⎟ hil ( x(t )) − ε iRi hiRi ( x(t ), u (t )) ⎥
⎝ l ⎠
m
min ∑ ⎢ ⎥
l =1
(3.21)
u (t )
i =1
⎢ εiRi ⎥
⎢ ⎥
⎣⎢ ⎦⎥

subject to

uli ≤ ui ≤ uhi , i = 1, ,m

S ( x(t ), xssN , u (t ),[u (t ) − u (t − Δt )]/ Δt ) ≤ 0 , V ( x(t ) − xssN ) ≥ σ

Sσ ( x(t ), xssN , u (t )) ≤ 0 , V ( x(t ) − xssN ) < σ

 61
Remark 1: In the case that u ∈ℜ , the numerical solution to the constrained optimization

problem of (3.21) is obtained by using the following algorithm:

1. Solve

2
⎡ Ri −1
⎛R ⎞ ⎤
⎢ spi
y − hi ( x (t )) − ∑ ε il ⎜ i ⎟ hil ( x(t )) − ε iRi hiRi ( x(t ), u (t )) ⎥
⎝ l ⎠
m
min ∑ ⎢ ⎥
l =1
(3.22)
u (t )
i =1
⎢ εiRi ⎥
⎢ ⎥
⎣⎢ ⎦⎥

subject to

uli ≤ ui (t ) ≤ uhi , i = 1, ,m

2. If the u calculated in step 1 satisfies

⎡ ∂V ⎧ ∂f ∂f du ⎫⎤ ∂V
β 2 ⎢ f T Pf + ⎨ f + ⎬⎥ + 2β f +V ≤ 0 (3.23)
⎣ ∂X ⎩ ∂X ∂u dt ⎭⎦ ∂X

then implement the u.

3. If the u calculated in step 1 does not satisfy the inequality of (3.23), calculate u

from

S ( x(t ), xssN , u (t ),[u (t ) − u (t − Δt )]/ Δt ) ≤ 0 , V ( x(t ) − xssN ) ≥ σ

Sσ ( x(t ), xssN , u (t )) ≤ 0 , V ( x(t ) − xssN ) < σ

and then implement the u.

The algorithm indicates that the state feedback of (3.15) is witches hybrid of an input-

output linearizing state feedback and a Lyapunov-based state feedback. Indeed, at any

time instant, it is an input-output linearizing or a Lyapunov-based state feedback. It may

switch from one state feedback to the other. Figure 3.1 illustrates a typical closed-loop

output response.
 62

ysp

time
Figure 3.1 Illustration of a controlled output response achieved under the controller
system.

Example: Unstable Non-Minimum-Phase Linear System

Consider the linear system:

dx1
= x2 ,
dt
dx2
= 10 x1 + 9 x2 + u ,
dt
y = 2 x1 − x2

which is unstable and non-minimum-phase. The relative order of this system is R = 1 .

For this system, the constraint optimization problem in (3.16) takes the form:

⎡ 2 x1 − x2 − ε1 (u + 10 x1 + 7 x2 ) − ysp ⎤
2

min ⎢ ⎥
u
⎣ ε1 ⎦

subject to
 63
⎡ ∂V ⎧ ∂f ∂f du ⎫⎤ ∂V
β 2 ⎢ f T Pf + ⎨ f + ⎬⎥ + 2β f +V ≤ 0
⎣ ∂X ⎩ ∂X ∂u dt ⎭⎦ ∂X

Here ysp =1 ( x1ss = 1, x2 ss = 0, uss = −10 ). The following controller parameter values are

⎡ 20 0⎤
chosen: ε1 = 0.2 , β =0.7 and P = ⎢ ⎥
⎣ 0 1⎦

For the two sets of initial conditions, [ x1 (0), x2 (0)] = [0.8, 0.2] and [0.2, 0.8] , the

performance of the state feedback of (3.16) is shown in Figures 3.2a and 3.2b. As can be

seen, the controller successfully drives both state variables to their desired steady state

values.

3.3.3 Reduced-Order State Observer

In general, measurements of all state variables are not available. In such cases, estimates

of the unmeasured state variables can be obtained from the output measurements. Here,

we use the ‘closed-loop’, reduced-order, nonlinear, state observer of (3.8) to reconstruct

the unmeasured state variables. The state observer is applicable to processes operating at

both stable and unstable steady states.

3.3.4 Integral Action

To ensure offset-free response of the closed-loop system in the presence of constant

disturbances and model errors, the control system should have integral action. An

estimate of disturbance-free process outputs is first calculated by using the closed-loop

process model:

w = f ( w, Ψ ( w, ysp ))
(3.24)
ξ = h( w)
 64
1.2

0.8 initial condition1

x1 initial condition2
ysp
0.4

0.0
1.2
initial condition1
initial condition2
0.8
x2

0.4

0.0
2

initial condition1
initial condition2
Controller Flag

0
0 1 2

Time

Figure 3.2a State responses under the controller system and the controller flag.
 65

-5
initial condition1
initial condition2

-9
u

-13

-17

12 initial condition1
initial condition2

8
V

0
0 1 2

Time

Figure 3.2b Manipulated input response and the Lyapunov function under the controller
system.
 66
where ξ is the estimate of the disturbance-free controlled output. The difference between

this estimate and the measurement of the controlled outputs, y, is then added to the output

set-point, as in internal model control, leading to:

w = f ( w, Ψ ( w, v))
v = ysp − y + h( w) (3.25)
u = Ψ ( xˆ, v)

An interesting feature of this approach to adding integral action is that the addition of the

dynamic system of (3.24) to the state feedback of (3.14) [calculation of xssN = v

according to (3.25)] adds no additional conditions for closed-loop asymptotic stability. In

other words, if the closed-loop system is asymptotically stable under the state feedback of

(3.14) alone, it is also asymptotically stable under (3.25).

3.3.5 Controller System

Combining the state feedback of (3.14), the reduced-order observer of (3.8), and the

dynamic sub-system of (3.24), leads to the following controller system that has integral

action:

z = Fη (ηˆ , y, Y , u ) − L1 Fy (ηˆ, y, Y , u ) − L2 FY (ηˆ , y, Y , u )

ηˆ = z + L1 y + L2Y
xˆ = Τ −1 (ηˆ , y, Y )
(3.26)
w = f ( w, Ψ ( w, v))
v = F ( ysp − y + h( w) )
u = Ψ ( xˆ, v)

A parameterization of the controller system is shown in Figure 3.3.

 67

Y = K(x)
Y
ysp Θ v u = Ψ(xˆ, v) u
v = F(Θ) x = f (x,u)
y
y = h( x)
x̂

z = Fη (ηˆ, y,Y, u) − LF
1 y (η, y,Y, u) − L2FY (η, y,Y, u)
ˆ ˆ
ηˆ = z + L1y + LY
2

xˆ = T−1(ηˆ, y,Y)

ξ
w = f (w, Ψ(w, v)) ξ = h( w)

Figure 3.3 Parameterization of the controller system.

u
C Ai

CA T u
CA

Figure 3.4 Schematic of non-isothermal continuous stirred tank reactor (CSTR).

 68
3.4 Illustrative Examples

3.4.1 Single-Input Single-Output Chemical Reactor

Consider the constant-volume, non-isothermal, continuous stirred-tank reactor shown in

Figure 3.4, in which the reaction A → B takes place in the liquid phase. The reactor

dynamics are represented by the model:

dC A ⎛ −E ⎞ F
= − Z exp ⎜ a ⎟ C A + (C Ai − C A )
dt ⎝ RT ⎠ V0
dT ⎛ −E ⎞ F
= γ Z exp ⎜ a ⎟ C A + (Ti − T ) + q
dt ⎝ RT ⎠ V0

where F is the volumetric flow rate of the inlet and outlet streams, V0 is the reactor

volume, and C Ai is the concentration of A in the inlet stream. The reactor parameter

values are given in Table 3.1. This reactor has multiple steady states. It is desired to

control the reactor temperature by manipulating feed flow rate, F . All state variables are

assumed to be measured.

Here x = [C A T ] , u = [ F ] , and y = [T ] . The steady state corresponding to

T

ysp = 293.91 ( x1ss = 8.39 , x2 ss = 293.91 , uss = 0.45 ) is stable (eigenvalues of the

Jacobian evaluated at the steady state are −4.5 and −0.5 , and to ysp = 302 ( x1ss = 6.319 ,

x2 ss = 302 , uss = 0.45 ) is unstable (eigenvalues of the Jacobian evaluated at the steady

state are -4.5 and 0.309). The zero dynamics of this process are governed by:

d ζ ⎧⎪ C A − ζ ⎫⎪ ⎛ − Ea ⎞ C Ai − ζ
= ⎨ −1 − γ i ⎬ Z exp ⎜⎜ ⎟⎟ ζ − q
dt ⎪⎩ Ti − ysp ⎪⎭ ⎝ Rysp ⎠ Ti − y sp

whose Jacobian evaluated (a) at ξ ss =8.39 has an eigenvalue at -10.91 and (b) at

ξ ss =6.319 has an eigenvalue at 36.27. Thus, the steady state corresponding to the lower
 69
Table 3.1 Parameter values of the non-isothermal reactor.

Parameter Value Unit

Z 5.0 × 108 s −1
Ea / R 8100 K
γ 3.9 m3 K kmol −1
q −2.519 × 10−2 K s −1
C Ai 12 kmol m −3
Ti 300 K
V0 0.1 m3

temperature set-point is stable and minimum-phase, and the one corresponding to the

higher temperature is unstable and non-minimum-phase.

For this process, the state feedback of (3.14) takes the form:

2
⎡ ⎧⎪ ⎛ −E ⎞ x 2 − ysp ⎫⎪ Ti − x2 ⎤
min ⎢ ⎨γ Z exp ⎜ a ⎟ x1 + q + ⎬+ u⎥
u
⎢⎣ ⎪⎩ ⎝ Rx2 ⎠ ε1 ⎪⎭ V0 ⎥⎦

subject to

⎡ ∂V ⎧ ∂f ∂f ⎫⎤ ∂V
β 2 ⎢ f T Pf + ⎨ f + u ⎬⎥ + 2β f +V ≤ 0
⎣ ∂x ⎩ ∂x ∂u ⎭⎦ ∂x

The process is initially at [ x1 (0), x2 (0)] = [10, 290] , which located in the minimum-phase

region. Initially ysp = 293.91 , and it then increases to ysp = 302 . The following controller

⎡ 0.0755 0.116 ⎤
parameter values are used: ε1 = 0.2 , β = 0.4 , and P = ⎢ ⎥.
⎣0.1116 0.1524 ⎦
 70

x1 (kmol m )
-3

5
305

300 x2
x2 (K)

ysp

295

290
2
Controlller Flag

0
0 4 8 12
Time (hr)

Figure 3.5a State responses of the non-isothermal CSTR under the controller and the
controller flag.
 71

0.8

0.6
u (m3 h)

0.4

0.2

0.0
8

4
V

0
0 4 8 12
Time (hr)

Figure 3.5b Manipulated input response and the Lyapunov function, corresponding to
Figure 3.5a
 72
Figure 3.5a shows the closed-loop responses of the state variables under the control

system of (3.26); the control system successfully operates the reactor at the desired steady

state, whether stable minimum- or unstable non-minimum-phase. The figure also shows

the controller flag. Controller flag of one indicates that the stability constraint is active.

With ysp = 293.91 , the Lyapunov stability constraint is active until the input-output

linearizing state feedback satisfies the stability constraint. When the set point changes

to ysp = 302 , the constraint remains continuously active because the output tracking

cannot stabilize the closed-loop system at the non-minimum-phase steady state. The

corresponding manipulated input and Lyapunov function profiles are shown in Figure

3.5b.

3.4.2 Two-Input Two-Output Chemical Reactor

Consider the same reactor of the previous example but with feed flow rate, F , and the

rate of heating/cooling, q , as manipulated inputs to maintain the reactor temperature, T ,

and the concentration of A in the reactor, C A , at three different set points. All state

measurements are assumed to be available. Here, x = [C A T ] , u = [ F q ] , and

T T

y = [C A T ] .
T

The stability of the steady states corresponding to the set points is as follows. The

one corresponding to ⎡⎣ ysp1 = 8.39, ysp2 = 293.91⎤⎦ is stable (eigenvalues of Jacobian

evaluated at the steady state λ1 ( J ol ) = −4.5 and λ2 ( J ol ) = −0.5 ), the one corresponding

to ⎡⎣ ysp1 = 6.319, ysp2 = 302⎤⎦ is unstable ( λ1 ( J ol ) = −4.5 and λ2 ( J ol ) = 0.309 ), and the one
 73
corresponding to ⎡⎣ ysp1 = 4.57, ysp2 = 308.82⎤⎦ is stable ( λ1 ( J ol ) = −4.5 , and λ2 ( J ol ) =

−0.742 ). All of the steady states are minimum-phase.

For this process, the constraint optimization problem in (3.14) with R1 = R2 = 1 takes the

form:

u
{
min [ a1 ( x) + b1 ( x, u ) ] + [ a2 ( x) + b2 ( x, u )]
2 2
}
subject to

⎡ ∂V ⎧ ∂f ∂f ⎫⎤ ∂V
β 2 ⎢ f T Pf + ⎨ f + u ⎬⎥ + 2β f +V ≤ 0
⎣ ∂x ⎩ ∂x ∂u ⎭⎦ ∂x

where

⎛ −E ⎞ x1 − ysp1 (C Ai − x1 )u1
a1 ( x) = − Z exp ⎜ a ⎟ x1 + , b1 ( x, u ) =
⎝ Rx2 ⎠ ε1 V0

⎛ −E ⎞ x − ysp 2 (T − x )u
a2 ( x) = γ Z exp ⎜ a ⎟ x1 + 2 , b2 ( x, u ) = i 2 1 + u2
⎝ Rx2 ⎠ ε2 V0

The initial conditions [ x1 (0), x2 (0)] = [10, 290] are in the minimum-phase region. At

first, ysp1 = 8.39 and ysp2 = 293.91 . They are then changed to ysp1 = 6.319 and ysp2 = 302

and finally to ysp1 = 4.57 and ysp2 = 308.82 . The following controller parameter values

⎡0.0069 0 ⎤
are uses: ε 1 = 0.3 , ε 2 = 0.3 , β = 0.6 , and P = ⎢
0.0016 ⎥⎦
.
⎣ 0

As Figures 3.6a and 3.6b show, the control system of (3.26) successfully operates the

reactor at the three steady states. In Figure 3.6a shows that the Lyapunov constraint is

inactive when the state variables are close to their steady state values. The decay of the
 74
10
x1
ysp1

x1 (kmol m )
8
-3

4
310

305
x2 (K)

x2

300
ysp2

295

290
2
Controller Flag

0
0 1 2

Time (hr)

Figure 3.6a State responses of the non-isothermal MIMO CSTR under the control
system and the controller flag.
 75
0.5

0.4

u1 (m h )
-1
3
0.3

0.2

0.1
0

-20
u2 (K s )
-1

-40

-60

-80

-100
0.15

0.10
V

0.05

0.00
0 1 2
Time (hr)

Figure 3.6b Manipulated input responses and the Lyapunov function under the control
system, corresponding to Figure 3.6b.
 76
Lyapunov function with time confirms the asymptotic approach of the state variables to

their desired steady-state values.

3.5 Conclusions

A control method that can be used to operate nonlinear processes at stable and unstable

steady states, whether non-minimum- or minimum-phase, was presented. The control

method has advantages of both input-output linearization and Lyapunov control. Input-

output linearization performs output tracking while the Lyapunov stability constraint

ensures asymptotic closed-loop stability when the tracking is incapable of ensuring the

stability. The feasibility of the control system of (3.14) implies asymptotically stability

of the closed-loop system, and the feasibility region depends on the choice of matrix P.
 77
3.6 Notation

A Reactant
B Product
C Ai Inlet concentration of the reactant, kmol m-3
CA Outlet concentration of the reactant, kmol m-3
D Differential operator, D = d / dt
Z Reaction rate constant, s −1
m Number of manipulated inputs
n Number of state variables
Ri Relative order (degree) of output yi
t Time, s
T Reactor outlet temperature, K
Ti Reactor inlet temperature, K
u Vector of manipulated inputs
V Reactor volume, m3
x Vector of state variables
y Vector of controlled outputs
ysp Vector of set-points
F Reactor feed flow rate m3 h −1
Ea Activation energy of the reaction kJ kmol −1
R Universal gas constant kJ kmol −1 K −1

Greek

ε 1 ,… , ε n Adjustable controlled parameters

γ Reactor model parameter, K m3 kmol −1
β Adjustable controlled parameter
σ Adjustable controlled parameter
 78
Subscripts

A, B Chemical species
ss Steady State
sp Set-point
0 Initial value

Math Symbols

⎛a⎞ a!
⎜ ⎟ b !(a − b)!
⎝b⎠
 79
Chapter 4: Software for Analytical Model-Based Controller Design

4.1 Introduction

The differential-geometric control has not been widely implemented in the process

industries due to the difficulties and complexities inherent in the controller design. It

requires taking analytical partial derivatives and performing symbolic manipulations,

which become cumbersome as the relative order and/or the level of complexity of the

model increases. Furthermore, many steps are involved in the design and implementation

of differential controllers including process modeling, process analysis, controller design,

controller verification, and code implementation. The design tasks are very elaborate and

error prone. Today, computers play a significant role in solving complex scientific and

engineering problems. Development of software simplifying the design of the model-

based controller is of great interest in the field of process control [65-74].

This chapter presents the development of a software package motivated by the

deficiencies of the existing analytical model-based controller design software. A new

software package that has a user-friendly interface and fully automates the design of

differential-geometric controllers is introduced herein for general nonlinear

processes. The software package generates controller equations in C, FORTRAN, and

MATLAB formats and carries out closed-loop simulations. The MATHEMATICA

program is selected as the symbolic computational engine, because it has built-in

functions that provide adequate precision and accuracy in computation, allowing the

creation of specific programming packages. In the package, the users can develop their

own functions that support their individual calculating needs based on the available
 80
functions in MATHEMATICA. Furthermore, it does not require the declaration of

symbolic variables in an expression, and it supports the communication between an

external program (front-end interface) and its kernel. The Visual Basic language is used

for creating a stand-alone application that communicates with the MATHEMATICA

kernel to perform symbolic manipulations and calculations using the MathLink program.

The differential-geometric control methods presented in [91-95] are programmed as

MATHEMATICA packages, to design controllers for general nonlinear processes

whether stable, unstable, minimum-phase, and/or non-minimum-phase. The software

automates the design of analytical model-based controllers for both continuous and batch

processes, avoiding laborious analytical calculations and manipulations when the process

model is complex or has high relative order(s).

The organization of this chapter is as follows. The scope of the study and some

mathematical preliminaries on differential-geometric, model-based control are given in

Section 4.2. Section 4.3 presents the software environment and its details. Finally, the

application and implementation of the software are illustrated in Section 4.4. Concluding

remarks are given in Section 4.5.

4.2 Mathematical Preliminaries

The control package introduced herein synthesizes the differential-geometric control

methods presented in [91-94, 96], always including a state feedback. It may also include

a state observer to reconstruct unmeasured state variables and/or another dynamic system

(compensator) to add integral action. In this section, the state feedback design is

reviewed briefly. For the observer design, the reader can refer to [95].
 81
Consider the general class of multivariable processes having mathematical models

in the form:

dx
= f ( x, d , u ) x(0) = x0
dt (4.1)
y = h( x )

where x = [ x1 xn ]T ∈ R n is the vector of state variables, u = [u1 um ]T ∈ R m is the

vector of manipulated inputs, d = [d1 d q ]T ∈ R q is the vector of measured disturbances,

y = [ y1 ym ]T ∈ R m is the vector of controlled outputs, and f ( x, u ) =

[ f1 ( x, u ) f n ( x, u )]T and h( x) = [h1 ( x) hm ( x)]T are smooth functions.

A) Minimum-phase processes. For this sub-class of processes, one can

implement the input-output linearization method described in [91, 96] for processes with

measured disturbances. For a process in the form of (4.1), closed-loop output responses

of the following form are requested:

(ε1 D + 1) R1 y1 = ysp1
(4.2)
(ε m D + 1) Rm ym = yspm

where D = d / dt , R1 ,… , Rm are relative orders of the process outputs, y1 ,… , ym , and

ε1 , , ε m are positive constants that set the speed of the process outputs, y1 ,… , ym ,

respectively. If equation (4.2) is solvable for u (after substituting for the time derivatives

of the outputs from the process model), a state feedback of the following form can be

calculated when the process has no measured disturbances:

u = ψ R ( x, ysp ) (4.3)
 82
B) Stable Non-minimum-phase processes. Kanter et al. [92] extended the

method in [96] to stable, non-minimum-phase, nonlinear processes, leading to an

approximate input-output linearizing control method. For the processes in the form of

(4.1), with no measured disturbances, linear responses of the closed-loop process outputs

are requested, having the form:

(ε1 D + 1) P1 y1 = ysp1
(4.4)
(ε m D + 1) Pm ym = yspm

where P1 ≥ R1 , , Pm ≥ Rm . This leads to a dynamic state feedback in the compact form:

Φ P ( x, u, u (1) , u (2) , ) = ysp (4.5)

After setting all the time derivatives of u to zero, if (4.5) is solvable for u, a static state

feedback in the general form:

u = ψ P ( x, ysp ) (4.6)

is calculated. The tunable parameters ε 1 , , ε m are chosen such that for every

l = 1, , m , all of the eigenvalues of [ I + ε l J ol ( xss , uss )] lie inside the unit circle.

C) General processes. The application of the method described in [92] is limited

to stable processes. To design controllers for unstable, non-minimum-phase processes,

Panjapornpon et al. [93] and Panjapornpon and Soroush [94] propose two controller

design methods.

Approximate input-state linearizing control method. This approach is an

extension of [92]; it involves requesting higher-order linear responses for the state

variables [93]. The following constrained optimization problem is solved at each time

instant:
 83
2
n ⎡ (ε i D + 1) pi xi − xssi ⎤
min ∑ wi ⎢ ⎥ (4.7)
u
i =1 ⎣⎢ ε ipi ⎥⎦

subject to:

u (l ) = 0, l ≥1

where xssi is the steady-state value of state, xi , corresponding to a given output set-point,

ysp , w1 ,..., wn , are adjustable, positive, scalar weights, whose values are set according to

the relative importance of the state variables; the higher the value of wi , the smaller the

xi response time.

Input-output linearizing control method with stability constraint. This

approach involves tracking linear output trajectories subject to a hard stability constraint

at each time instant [94]:

2
m ⎡ (ε i D + 1) Ri yi − ysp ⎤
min ∑ ⎢ i
⎥ (4.8)
u
i =1 ⎢
⎣ εiRi
⎥⎦

subject to:

β 2V + 2βV + V ≤ 0

where V is a Lyapunov function given by:

V = ( x − xss )T P( x − xss )

ε1 , , ε m are positive constants that set the response speed of the process outputs,

y1 ,… , ym , P is a positive-definite matrix, and a tuning parameter, β , is chosen such that

β > max(ε1 ,… , ε m ) .
 84

Figure 4.1 Flow diagram of the major components of the model-based controller
design software.

4.3 Software Environment

Given a process model, the software allows one to synthesize easily differential-

geometric, model-based controllers and perform closed-loop numerical simulations. The

software has three principal features: a user interface, MATHEMATICA routines, and a

simulation algorithm. A flow diagram of the software, showing the interactions among

these features, is shown in Figure 4.1. The software receives the process model and

parameter values through the front-end window, and then sends a set of commands to the

MathLink program. The latter invokes the MATHEMATICA kernel to execute the
 85
control packages. The kernel handles the controller design task based upon the process

information and the control method chosen by the user. The designed controller

equations and the tuning parameter values are used to carry out numerical simulations.

The performance of the closed-loop control system is presented in the form of tables and

graphs. Each feature is described next.

4.3.1 User Interface

The front-end window interacts with the user in five steps: process information

acceptance, process analyses, controller equation generation, closed-loop simulation, and

graph selection. The user enters the requested information in each step before proceeding

to the next step. A flow diagram linking the steps and the tasks involved is shown in

Figure 4.2. In addition, the front-end window has an input command step that shows all

of the commands sent to the MATHEMATICA kernel. The user enters the process model

equations and identifies the controlled outputs, manipulated inputs, process parameters,

and available measurements in the process information step shown in Figure 4.3. The

process model is entered as a set of ordinary differential and algebraic equations in the

MATHEMATICA format. The developed software has been tested successfully with the

process models that contain state variables less than or equal to 10. It also aids the user

to design both continuous and batch processes controllers. However, the controller design

for the batch processes is limited to stable minimum-phase processes and cannot perform

a closed-loop simulation. The following steps after the process information acceptance

are only valid for the controller design of the continuous processes. In the step of process

analyses, the software calculates all feasible steady-state pairs and allows the user to

select a desired pair to perform local stability analysis, as illustrated in Figure 4.4.
 86

Figure 4.2 Flow diagram of the tasks in the model-based controller design software
(numbers correspond to the steps in the user interface).
 87

Figure 4.3 The system information window (step 1) of model-based controller design
software.

Depending on the analysis results (stability and minimum-phase-ness), the software

allows the user to select an appropriate controller method. According to the logic shown

in Figure 4.5, the software directs the user to the applicable control methods. Figure 4.6

shows the code viewer window in the controller equations step, which presents the

model-based controller equations either in C, FORTRAN, or MATLAB formats.

Subsequently, a file containing the code for the control (closed-loop) system is loaded in
 88
the simulation setup step. The user enters initial values for the state variables, controller-

tuning parameter values, and the simulation time, as shown in Figure 4.7. The simulation

method in the dynamic library link (DLL) is selected according to the type of the control

method. The dynamic library link is a library of executable functions that can be used by

a Windows application. By creating a simulation function as DLLs, the new software

package can be installed on any Windows-based computer without requiring the

installation of MATLAB. The closed-loop simulation results are presented in the form of

tables and graphs, as depicted in Figure 4.8. The user selects the plot variables in the plot

selection step.

Figure 4.4 The process model window (step 2) of model-based controller design
software.
 89

Figure 4.5 Diagram showing how the software directs the user to the controller method
that are applicable to the continuous process under consideration.
 90

Figure 4.6 The code viewer window showing controller equations in the FORTRAN
language format.
 91

Figure 4.7 The simulation setup (step 4) and the plot selection (step 5) windows of the
software.
 92

Figure 4.8 A simulation plot generated by plot selection step of the software.

4.3.2 MATHEMATICA Packages

Two types of MATHEMATICA packages, that are controller design and process analysis

packages, were developed. The differential-geometric control methods in [91-94, 96] are

programmed into the controller design packages for application to stable and unstable

processes, whether non-minimum- or minimum-phase. Each controller package designs

 93
the state feedback, compensator, state observer, and generates controller code in a

programming language, such as C, FORTRAN, or MATLAB. The input expressions for

the five controller design methods are shown in Table 4.1.

The process analysis packages include two main support functions; the stability

analysis function and the state equilibrium function. The stability analysis function

determines the stability of the process, the zero dynamics, and the zero dynamics

stability. The state equilibrium function computes all feasible steady-state pairs within

given operating ranges of the state variables. The input expressions for both functions are

described in Table 4.2.

Table 4.1 Controller design packages.

Package Name Input Expression

[Reference]

StableNonMinimum.m StableNonminimum[InputEq, OutputEq, StateEq, UEq,

[92] OrderP, ProcPar, SSsp, Measured]

Minimum.m Minimum[InputEq, OutputEq, StateEq, UEq, ProcPar,

[96] SSsp, Measured]

MinimumDm.m MinimumDm[InputEq, OutputEq, StateEq, UEq, DmEq,

[91] ProcPar, SSsp, Measured]

UnstableNonminimum.m MIMOunstableNonmin[InputEq, OutputEq, StateEq,

[93] UEq, OrderP, ProcPar, SSsp, Measured]

MIMOLyapunov.m MIMOLyapunov[InputEq, OutputEq, StateEq, UEq,

[94] ProcPar, SSsp, Measured]

BatchMinimum.m BatchMinimum[InputEq, OutputEq, StateEq, UEq,

[96] ProcPar, Measured]
 94
Table 4.2 Design analysis packages.

Package Input expression

Name

analyse.m analyse[InputEq, OutputEq, UEq, dmEq, StateEq, SSsp, ProcPar]

StateEquilibrium[InputEq,StateEq, OutputEq, OutputSp, ProcPar,

IpRange, IpSpan ]

4.3.3 Simulation Algorithm

While MATHEMATICA has symbolic manipulation capabilities, its numerical

algorithms are weak. For this purpose, the simulation algorithms of the model-based

controller design software are created as Dynamic Library Link (DLL) functions using

the COM Builder of MATLAB. The DLL functions provide a flexible environment for a

stand-alone application. The controller equations in the MATLAB format, with given

process parameters, are integrated by appropriate functions that are developed

specifically for the control methods in [91-94, 96]. These simulation functions are

OdeAna, OdeFminunc and OdeFmincon. The user adjusts the tuning parameters and

simulation time to evaluate the controller performance.

4.4 Examples

Next, for several chemical and biochemical reactors, the software is applied to design

differential-geometric, model-based controllers.

 95
4.4.1 Continuous Bioreactor

Consider a constant-volume, continuous bioreactor described by:

x2
x1 = − x1u + x1 (1 − x2 ) exp( )
γ1
x2 1 + α1
x2 = − x2u + x1 (1 − x2 ) exp( )
γ 1 1 + α1 − x2
y = x1

where x1 is the dimensionless outlet cell mass concentration, x2 is the dimensionless

outlet substrate concentration, and u is the dimensionless substrate feed rate. The

parameters α1 = 0.02 and γ 1 = 0.48 . It is desired to maintain the cell mass concentration,

x1 , at 0.1448, by manipulating u . It is assumed that only the cell mass concentration is

measured.

With x = [ x1 x2 ] , u = [ u ] , and y = [ x1 ] , the inputs to the software in step 1 are

T

as follows:

Step 1
Input1:
x1'[t_]=-x1[t]*u[t]+x1[t]*(1-x2[t])*Exp[x2[t]/gamma1],
x2'[t_]=-x2[t]*u[t]+x1[t]*(1-x2[t])*Exp[x2[t]/gamma1]*(1+alpha1)/(1+alpha1-x2[t]),
y1=x1[t]
Input 2: alpha1=0.02, gamma1=0.48
Input 3: x1[t], x2[t] Input 4: u[t]
Input 5:
Input 6: y1 Input 7: 0.1448
Input 8: {x1[t],0,1},{x2[t],0,1},{u[t], 0.2, 1.5}
Input 9: 5
Input 10: Observer design Input 11: Continuous process
 96
The software calculates all feasible steady state pairs within given operating ranges. The

user chooses the pair ( x1ss = 0.1448, x2 ss = 0.8455, uss = 0.9) to perform stability analysis.

At this steady state pair, the process is unstable and non-minimum-phase; the eigenvalues

of the Jacobians of the process and its zero dynamics are in the right-half-plane (RHP)

( λ1 ( J ol ) = 0.061 + 1.731 i , λ2 ( J ol ) = 0.061 − 1.731 i and λ1 ( J zd ) = 1.374). When the

approximate input-state linearizing control method [93] is selected as the controller

design method and the orders of the requested state responses are p1 = 1 and p2 = 1 , the

inputs for step 2 are given below. The controller equations are generated in Step 3.

Step 2
Input 1: 0.1448, 0.8455, 0.9 Input2: 1, 1
Input 3: Approximate Input-State Linearization

Initial conditions are [ x1 (0), x2 (0)] = [0.1448, 0.8455], and it is desired to maintain the

outputs at ysp1 = 0.1448, which is in the non-minimum-phase region. The following

tuning parameter values are selected: ε 1 = 0.1 , ε 2 = 0.35 , and L2 = 0.7, and the simulation

time is 4 hours. Details of inputs in Step 4 are shown below, with the closed-loop

responses in Figure 4.9.

Step 4
Input 1: [0, 4]
Input 2: {Epsilon1, Epsilon2, L11} Input 3: 0.1,0.35,0.7
Input 4: x1, x2, Zeta1, Xi11, Xi21 Input 5: 0.1,0.75,0.05,0.1,0.75
Input 6: u Input 7: 0.45
 97

Figure 4.9 Simulated closed-loop response of the state variables and manipulated input
of the bioreactor.

4.4.2 Single-Input Single-Output Chemical Reactor

Consider a constant-volume, non-isothermal, continuous stirred-tank reactor, in which the

exothermic reaction, A → B , takes place in the liquid phase. The reactor dynamics are

represented by the model:

 98

dC A ⎛ E ⎞ F
= − k0 exp ⎜ − a ⎟ C A + (C A0 − C A )
dt ⎝ RT ⎠ V
dT ⎛ E ⎞ F
= γ k0 exp ⎜ − a ⎟ C A + (Ti − T ) + q
dt ⎝ RT ⎠ V
y =T

where F is the inlet volumetric flow rate of pure A , V is the reactor volume, and C A0 is

the concentration of A in the feed stream. The process parameter values are given in

Table 4.3.

Table 4.3 Process parameters of the non-isothermal reactor.

Parameter Value Unit

k0 5.0 × 108 s −1
Ea / R 8100 K
γ 3.9 m3 K kmol −1
q −2.519 × 10−2 K s −1
C A0 12 kmol m −3
Ti 300 K
V 0.1 m3

The reactor temperature, T , is measured and it is desired to maintain it at ysp = 302 by

manipulating the feed rate, F . Here, x = [C A T ] , u = [ F ] , and y = [T ] . The inputs to

T

the software in step 1 are given below.

 99
Step 1
Input 1:
x1'[t_] = u1[t]*(Ca0 - x1[t])/V - k1*x1[t]*E^(-E1R/x2[t]),
x2'[t_] = u1[t]*(Ti - x2[t])/V+ gamma*k1*x1[t]*E^(-E1R/x2[t]) +Q,
y1 = x2[t]
Input 2: Ca0=12, Ti=300, k1= 1.8 × 1012 , E1R=8100, gamma=3.9, Q=-90.7, V=0.1
Input 3: x1[t], x2[t] Input 4: u1[t]
Input 5:
Input 6: y1 Input 7: 302
Input 8: {x1[t], 0, 12}, {x2[t], 290, 325}, {u1[t], 0.1, 3.5}
Input 9: 5
Input 10: Observer design Input 11: Continuous process

The steady state pair ( x1ss = 6.319 , x2 ss = 302 , and uss = 0.45 ) is selected to perform

stability analysis. Eigenvalues of the Jacobians of the process and its zero dynamics are

in the RHP ( λ1 ( J ol ) = −0.45 , λ2 ( J ol ) = 0.309 , λ1 ( J zd ) = 36.271). Thus, the desired

steady state is unstable and non-minimum-phase. The method of input-output linearizing

control with the stability constraint [94] is selected as the controller design method. The

software inputs in step 2 are given below.

Step 2
Input 1: 6.319,302,0.45 Input 2: 2,2
Input 3: Input-Output Linearization with Stability Constraint

The desired set point is ysp = 302, and the process is initially at [ x1 (0), x2 (0)] =

[9.394, 293.906] , which is in the non-minimum-phase region. The tuning parameters are

ε 1 = 0.1 , β1 = 0.25 , a11 = 0.0755 , a12 = 0.105 , a21 = 0.105 , and a22 = 0.1524 , and the
 100
simulation time is two hours. The inputs in step 4 are given below. Finally, the

performance of the controller is shown in Figure 4.10.

Step 4
Input 1: [[0,2]
Input 2: {Epsilon1, Beta1, a11, a12, a21, a22}
Input 3: 0.1, 0.25, 0.0755, 0.105, 0.105, 0.1524
Input 4: x1, x2 Input 5: 8.394, 293.906
Input 6: u1 Input 7: 0.45

Figure 4.10 Simulated closed-loop response of the controlled output and manipulated
input of the non-isothermal CSTR.
 101
4.4.3 Multi-Input Multi-Output Jacketed Chemical Reactor

Consider a non-isothermal, continuous chemical reactor in which the reactions

A → B → C take place in the liquid phase. The process dynamics are represented by the

following model:

dC A F E
= (C A0 − C A ) − k1 exp(− 1 )C A2
dt V RT
dCB F E E
= − CB + k1 exp(− 1 )C A2 − k2 exp(− 2 )CB
dt V RT RT
dT F (−ΔH1 ) E (−ΔH 2 ) E US
= (Ti − T ) + k1 exp(− 1 )C A2 + k2 exp(− 2 )CB + (T − T )
dt V ρcp RT ρcp RT ρ c pV j
dT j F US
= (T ji − T j ) − (T j − T )
dt V ρ jcp jV j
y1 = CB
y2 = T

It is desired to maintain CB at 5.233 mol/l and T at 443.92 K by manipulating F and

F j . Only measurements of CB and T are available. The process parameter values are

given in Table 4.4. Let x = [C A CB T T j ]T , u = [ F Fj ]T and y = [CB T ]T . The process

dynamics can be expressed as:

u1 E
x1 = (C A0 − x1 ) − k1 exp(− 1 ) x12
V Rx3
u1 E E
x2 = − x2 + k1 exp(− 1 ) x12 − k2 exp(− 2 ) x2
V Rx3 Rx3
u1 (−ΔH1 ) E (−ΔH 2 ) E US
x3 = (Ti − x3 ) + k1 exp(− 1 ) x12 + k2 exp(− 2 ) x2 + (T − x )
V ρcp Rx3 ρcp Rx3 ρ c pV j 3
u2 US
x4 = (T ji − x4 ) − ( x4 − x3 )
V ρ j c p jV j
y1 = x2
y2 = x3
 102

Table 4.4 Process parameters of the non-isothermal jacketed reactor.

Parameter Value Unit

C A0 12 mol l −1
k1 2.5 × 1010 lmol −1hr −1
k2 1.5 × 1010 hr −1
E1 / R 8000 K
E2 / R 9100 K
Ti 320 K
T ji 298.15 K

ΔH1 −20 kJ mol −1

ΔH 2 −80 kJ mol −1
ρ 1 kg l −1

ρj 1.1 kg l −1

cp 2.25 kJ kg −1 K −1

cp j 3 kJ kg −1 K −1

U 3825 kJ m −2 K −1s −1
S 0.225 m2
V 5 l
Vj 5 l
 103
The inputs to the software in step 1 are given below.

Step 1
Input 1:
x1'[t_]= -((k1*x1[t]^2)/E^(E1R/x3[t])) + ((Ca0 - x1[t])*u1[t])/V,
x2'[t_]= (k1*x1[t]^2)/E^(E1R/x3[t]) - (k2*x2[t])/E^(E2R/x3[t]) - (x2[t]*u1[t])/V,
x3'[t_]= -((k1*H1*x1[t]^2)/(Cp*E^(E1R/x3[t])*Rho)) -
(k2*H2*x2[t])/(Cp*E^(E2R/x3[t])*Rho) + (S*U*(-x3[t] + x4[t]))/(Cp*Rho*V) + ((Ti -
x3[t])*u1[t])/V,
x4'[t_]= -((S*U*(-x3[t] + x4[t]))/(Cpj*Rhoj*Vj)) + ((TJi - x4[t])*u2[t])/Vj,
y1=x2[t],
y2=x3[t]
Input 2: k1=2.5*10^10, k2=1.5*10^10, H1=-20, H2=-80, Ca0=12, V=5, Vj=5,
E1R=8000, E2R=9100, Rho=1, Rhoj=1.1, Cp=2.25, Cpj=3, Ti=320, TJi=298.15,
U=3825, S=0.225
Input 3: x1[t], x2[t], x3[t], x4[t] Input 4: u1[t], u2[t]
Input 5:
Input 6: y1, y2 Input 7: 5.233, 443.92
Input 8: {x1[t],0,1.5}, {x2[t],0,10}, {x3[t],400,450}, {x4[t],400,440}, {u1[t],50,120},
{u2[t],50,120}
Input 9: 5
Input 10: Observer design Input 11: Continuous process

The steady-state pair corresponding to the desired set-point, ysp1 = 5.233 , and

ysp2 = 443.92 ( x1ss = 0.7 , x2 ss = 5.233 , x3 ss = 443.92 , x4 ss = 403.24 , u1ss = 80.95 , and

u2 ss = 100.94 ), is selected. The process is unstable [ λ1 ( J ol ) = −491.38 , λ2 ( J ol ) = −90.71 ,

λ3 ( J ol ) = 86.28 , λ4 ( J ol ) = −15.36 ] and non-minimum-phase [ λ1 ( J zd ) = 589.59]. The

approximate input-state linearizing control method [93] is selected as the controller

 104
design method. The orders of the linear responses of state variables, x1 , x2 , x3 , and x4 ,

are requested to be p1 = 2 , p2 = 2 , p3 = 2 , and p4 = 1 , respectively. The inputs to the

software in step 2 are given below.

Step 2
Input 1: 0.700, 5.233, 443.92, 403.24, 80.95, 100.94 Input 2: 2,2,2,1
Input 3: Approximate Input-State Linearization

The process is initially at [ x1 (0), x2 (0), x3 (0), x4 (0)] = [0.7, 5.233, 443.92, 403.24], ysp1 =

5.233, and ysp2 = 443.92. The set-point is in the non-minimum-phase region. The

following values are selected: ε1 = 0.09 , ε 2 = 0.09 , ε 3 = 0.08 , ε 4 = 0.08 , L1 =0.5,

L2 =0.5, with a simulation time of three hours. The inputs to step 4 are given below.

The simulation results showing the controller performance are shown in Figure 4.11.

Step 4
Input 1: [0,3]
Input 2: Epsilon1, Epsilon2, Epsilon3, Epsilon4, L11, L22
Input 3: 0.09, 0.09, 0.08, 0.08, 0.5 ,0.5
Input 4: x1, x2, x3, x4, Zeta1, Zeta2, Xi11, Xi12, Xi21, Xi22, Xi31, Xi32, Xi41
Input 5: 1.8138, 9.4456, 387.69, 380.67, -2.909, 186.83, 1.8138, -5.8085, 9.4456,
2.8518, 387.69, 25.457, 380.67
Input 6: u1, u2 Input 7: 40, 10
 105

Figure 4.11 Simulated closed-loop response of the controlled outputs and manipulated
input of the MIMO chemical reactor.
 106
4.4.4 Continuous Chemical Reactors in Series

Consider two isothermal, CSTRs in series, in which the reactions A → B → C take place

in the liquid phase. The process model is:

dC A1 F1
= (C A0 − C A1 ) − k AC A21
dt V1
dCB1 F
= − 1 CB1 + k AC A21 − k BCB1
dt V1
dC A2 F1 F
= (C A1 − C A 2 ) + 2 (C A0 − C A2 ) − k AC A2 2
dt V2 V2
dCB 2 F1 F
= (CB1 − CB 2 ) − 2 CB 2 + k AC A2 2 − k BCB 2
dt V2 V2
y1 = CB1
y2 = C B 2

where F1 and F2 are the inlet volumetric flow rates of pure A , V1 =0.01, V2 =0.01, C A0

=7, k A =6, and k B =1. The concentrations, CB1 and CB 2 , are measured and it is desired to

maintain them at ysp1 = 2 , and ysp2 = 3 by manipulating the feed rates, F1 and F2 .

With x = [C A1 CB1 C A 2 CB 2 ]T , u = [ F1 F2 ] , and y = [CB1 CB 2 ] , the inputs to the

T T

software in step 1 are given below.

Step 1
Input 1:
x1'[t_]= -(ka*x1[t]^2) + ((Ca0 - x1[t])*u1[t])/V1,
x2'[t_]= ka*x1[t]^2 - kb*x2[t] - (x2[t]*u1[t])/V1,
x3'[t_]= -(ka*x3[t]^2) + ((x1[t] - x3[t])*u1[t])/V2 + ((Ca0 - x3[t])*u2[t])/V2,
x4'[t_]= ka*x3[t]^2 - kb*x4[t] + ((x2[t] - x4[t])*u1[t])/V2 - (x4[t]*u2[t])/V2,
y1=x2[t],
 107
y2=x4[t]
Input 2: ka=6, kb=1, Ca0=7, V1=0.01, V2=0.01
Input 3: x1[t], x2[t], x3[t], x4[t] Input 4: u1[t],u2[t]
Input 5:
Input 6: y1, y2 Input 7: ysp1=2, ysp2=3
Input 8: {x1[t],0,7},{x2[t],0,7},{x3[t],0,7},{x4[t],0,7},{u1[t],0,0.2},{u2[t],0,0.2}
Input 9: 5
Input 10: Observer design Input 11: Continuous process

The steady state pair ( x1ss = 0.699 , x2 ss = 2 , x3 ss = 1.101 , x4 ss = 3 , u1ss = 0.005 ,

u2 ss = 0.0013 ) that corresponds ysp1 = 2 and ysp2 = 3 is selected. The pair is stable

[ λ1 ( J ol ) = −14.90 , λ2 ( J ol ) = −8.85 , λ3 ( J ol ) = −2.72 , λ4 ( J ol ) = −1.46 ] and non-

minimum-phase [ λ1 ( J zd ) = 17.57, λ2 ( J zd ) = 11.03]. The approximate input-output

linearizing control method [9], with P1 = 2 and P2 = 2 , is used. The inputs to step 2 are

detailed below.

Step 2
Input 1: 0.699, 2.0, 1.101, 3.0, 0.005, 0.013 Input 2: 2, 2
Input 3: Approximate Input-Output Linearization

With [ x1 (0), x2 (0), x3 (0), x4 (0)] = [0.75, 2.192, 0.75, 2.192], ε1 = 0.4 , ε 2 = 0.3 , and six

hours of simulation time, the inputs to step 4 are given below. Figure 4.12 shows the

closed-loop responses.
 108
Step 4
Input1: [0, 6]
Input2: {Epsilon1, Epsilon2} Input3: 0.4, 0.3
Input4: {x1, x2, x3, x4, Eta11, Eta12, Eta21, Eta22}
Input5: 0.75,2.192,0.75,2.192,0.75,2.192,0.75,2.192
Input6: {u1, u2} Input7: 0.1, 0.1

4.5 Conclusions

A user-friendly, integrated, software package for the design of nonlinear model-based

controllers is presented. Given a process model in the form of ordinary differential and

algebraic equations, the software derives an analytical model-based controller, generates

the controller equations in C, FORTRAN, or MATLAB format, and carries out closed-

loop simulations. A MathLink program provides an interface for communication with

the MATHEMATICA kernel. The software is applied for the control of chemical and

biochemical reactors. The software allows control engineers to design differential-

geometric, model-based controllers with ease. With the software, differential-geometric,

model-based controllers can be designed for processes involving complex models, as

implemented in industry.
 109

Figure 4.12 Simulated closed-loop response of the controlled outputs and manipulated
input of the isothermal chemical reactors in series.
 110
4.6 Notation

A, B, C Chemical species
cp Heat capacity of feed and product, kJ kg −1 K −1

cp j Heat capacity of jacket fluid, kJ kg −1 K −1

C A0 Inlet concentration of the reactant A , mol l −1

C A , C A1 , C A2 Outlet concentration of the reactant A , mol l −1
CB , CB1 , CB 2 Concentration of B , mol l −1
D Differential operator, D = d / dt
Ea , E1 , E2 Activation Energy, kJ mol −1
F Reactor flow rate m3 h −1 or l h −1
Fj Jacket coolant flow rate, l h −1
J ol Open loop Jacobian
J cl Closed-loop Jacobian

J zd Jacobian of the zero dynamics

L Matrix of observer gain

m Number of manipulated inputs
n Number of controlled outputs
Vector of the orders of the requested output
p
responses
Vector of the orders of the requested state
P
responses
q Cooling/heating rate, kJ s −1
ri Relative order of state variable xi

Ri Relative order of output variable yi

R Universal gas constant, kJ kmol −1 K −1

S Heat transfer surface area, m 2
t Time, s
 111
T Reactor outlet temperature, K
Ti Reactor inlet temperature, K
Tj Jacket temperature, K

T ji Jacket inlet temperature, K

u Vector of manipulated inputs
U Overall heat-transfer coefficient, kJ m −2 K −1 s −1
V , V1 , V2 Reactor volume, m
3

Vj Jacket Volume, l
x Vector of state variables
y Vector of controlled outputs
ysp Vector of set-points

k0 , k1 , k2 Reaction rate constant, h −1 , l mol −1 h −1 , h −1

−ΔH1 , − ΔH 2 Heat of reaction, kJ mol −1
InputEq Process model equation
OuputEq Output variables
StateEq State variables
UEq Output variables
DmEq Measured disturbances
OrderP Request output order
SSsp Select steady state pair
ProcPar Process parameters
Measured Complete or incomplete state measurements
OutputSp Output set point
IpRange Input ranges of state variables
IpSpan Input span ranges of state variables
 112
Greek

ε 1 ,… , ε n Adjustable parameters of controller

β Adjustable parameter of stability constraint

γ Reactor model parameter,
η State for integral action
λ1 , λ2 , Eigenvalues of Jacobian
ρ Density of mixture in reactor,
ρj Density of jacket fluid,

Subscripts

A, B Chemical species
ss Steady State
sp Set-point
0 Initial value, process
 113
Chapter 5: Conclusions and Future Research Directions

5.1 Conclusions

This research project was motivated by the inadequacy of the existing controller design

methods for general nonlinear processes and the difficulty of designing and implementing

differential-geometric model-based controller design. The specific contributions of this

research are as follows:

ƒ Two new control methods that are applicable to input-constrained general

multivariable processes, whether minimum- or non-minimum-phase, were

developed. Both control methods address limitations of I-O linearization.

ƒ The nonlinear state feedback of the first control method was obtained by

requesting state response as close as possible to a set of desired linear state

responses. The integral action ensures the offset-free of closed-loop response in

the presence of process-model mismatch. A closed-loop, reduced order observer

was used to estimate the unmeasured state variables. Compared to a general,

multivariable model-predictive controller, this control system has less tuning

parameters to achieve closed-loop, asymptotically stability. The advantages of

this control system over a general multivariable model-predictive controller will

be explained further at the end of this subsection.

ƒ The control system of the first control method does not have the limitations of the

control method presented by Kanter et al. [1]. However, because of the

optimization form (numerical nature) and shortest prediction horizon form of the
 114
control system, an analytical proof of the closed-loop stability for the proposed

control system is not available now.

ƒ The second control method has a hybrid control structure that is a combination of

I-O linearization and Lyapunov control. The I-O linearization performs optimal

tracking while the Lyapunov stability constraint ensures asymptotic closed-loop

stability within an assessable domain of attraction. The stability is ensured by

requiring state variables to evolve within a shrinking state-boundary. A feasible

solution to optimization problem in (3.15) implies asymptotically stability. The

size of the feasibility region depends on the adjustable P and β .

ƒ A new stability Lyapunov constraint was presented. It does not suffer from the

singularity problem that the existing one does. In addition, the proposed method

has a larger successive domain of attraction than one designed based on only

Lyapunov-based control.

ƒ A new software package for analytical model-based controller design and closed-

loop simulation was developed. With its user-friendly environment, the software

package designs the controller on the basis of a process model in a set of

differential and algebraic equations and generates the analytical model-based

controller equations in multiple formats. The model-based software integrates the

simulation tools to simulate the closed-loop response and display the results. This

prototype software package was developed to simplify industrial implementation

and testing of the model-based controllers. The new design software for nonlinear

model-based control would allow control engineers to more conveniently design

nonlinear controllers. This will lead to an increase in industrial implementations

 115
of the nonlinear model-based controllers. The developed software has already

been tested successfully on the process models with the number of state variables

less than or equal to 10.

The new control methods have less number of tuning parameters in comparison with a

typical MPC. For example, consider the general class of multivariable processes in the

form of (2.1) under the following MPC [97] :

min J = (Yr (k + 1) − Y (k + 1))T Q(Yr (k + 1) − Y (k + 1)) + ΔU (k )T RΔU (k ) (5.1)

ΔU ( k )

where

Yr (k + 1) [ yr (k + 1), yr (k + 2),… , yr (k + P )]T

Y (k + 1) [ y (k + 1), y (k + 2),… , y (k + P )]T

ΔU ( k ) [(u (k ) − u (k − 1)), (u (k + 1) − u (k )),… , (u (k + M − 1) − u (k + M ))]T

Yr (k + 1) is a vector of reference trajectory outputs over prediction horizon P . Y (k + 1)

is a vector of outputs over the prediction horizon P . ΔU (k ) is a vector of control action

changes over control horizon M . Q is an mP × mP matrix and R is an mM × mM

matrix weighting on the predicted error vectors of outputs and inputs respectively. Figure

5.1 illustrates the MPC concept and the control horizon. The control action is obtained by

carrying out open-loop optimization locally. The model-predictive control is seldom

applied to the unstable processes because the controller performance quickly deteriorates

in the presence of disturbance or process-model mismatch. Therefore, the unstable

process needs to be stabilized before implementing model-predictive control. The MPC

needs a large prediction horizon to handle the non-minimum-phase processes. To achieve

asymptotic stability, the MPC in (5.1) has ( mM × mM + mP × mP ) parameters to tune

 116

Figure 5.1 Illustration of the output prediction and control action in model-predictive
control with prediction horizon P and control horizon M.

while the feedback controllers in both approximate input-state linearization and I-O

linearization with stability constraint have n and m + n × n parameters, respectively. For

example, consider the process with 4 state equations, 2 manipulated inputs, 2 outputs, the

prediction horizon P =4, and control horizon M =4. The optimization in (5.1) requires 32

tuning parameters. In contrast, the approximate input-state linearization and the I-O

linearization with stability constraint require only 4 and 18 tuning parameters,

respectively.

To further illustrate the advantage of proposed control methods, consider the application

to the continuous bioreactor in section 2.4.2, where two different model-predictive

controllers have also been applied to this example (Brengel and Seider [7] and

Ramaswamy et al. [98]). The optimization problem in [7] was formulated by using two

outputs ( x1 , x2 ) and one manipulated input ( u ). It requires five tuning parameters (two

weights on the outputs, the prediction horizon, the control horizon, and the relinearization

instants per sampling interval). Alternatively, the optimization problem in [98]

 117
formulated by using only one output ( x1 ) and one manipulated input ( u ). It requires six

tuning parameters ( M , P , Q , R , S and sampling periods Δt ). The state feedback of

approximate input-state linearization has only two tuning parameters. Even though the

state feedback of input-output linearization with stability constraint requires five

parameters, it is applicable to a wide operating range. In the developed control methods,

the values of tuning parameters are initially determined by selecting the tuning

parameters of each state variable equal to its time constants. The tuning parameters are

adjusted based on the input and output response: the smaller the values of tuning

parameters, the faster the state responses.

5.2 Future Research Directions

As a consequence of carrying out this research project, the open problems listed below

were encountered.

The control method with guaranteed closed-loop stability

The first proposed control method is capable of guaranteeing the asymptotic stability of

the closed-loop system. However, the analytical proof of closed-loop stability under this

control method is difficult and still open due to the ‘short prediction horizons’,

optimization form of the control method. The shortness of the prediction horizons can be

easily seen, when one derives the controller by using the model-predictive approach

described in 3.2.1.

Feasibility of control action

In the design of stability constraint, the second derivative of the Lyapunov function was

used to avoid the singularity problem at the equilibrium point, which led to a more
 118
complex stability constraint. A simpler stability constraint without the singularity

problem is of interest. In addition, the existence of a control action satisfying the

constraint is a major problem in the design of the domain of attraction around the

reference steady state. The trajectories that are both optimal and feasible are complicated

by the fact that the space of possible control action is large and possibly non-convex. The

domain of attraction should not contain the singular point in the state trajectory because it

hinders achieving the requested response. To improve the feasibility of the control action,

some techniques such as the multiple Lyapunov functions can be used to enlarge the

domain of attraction. To guarantee feasibility, an efficient algorithm for computing time

variant subsets of piecewise quadratic Lyapunov function is required.

Deadtime in the input and output

Deadtimes in outputs and/or manipulated inputs can deteriorate the closed-loop stability.

Controllers that are capable of handling deadtimes are needed. Controllers that can

handle deadtimes in the outputs can be designed by including the time delays in the

desired closed-loop response. Controllers that can handle deadtimes in the manipulated

inputs are more difficult to design.

Robustness and optimality

In this work, the controllers were developed under the assumption that unmeasured

disturbances are piece-wise constant. However, many processes present different classes

of uncertainty, such as time-varying uncertainty. If the uncertainties are not accounted

for, they can lead to serious degradation of closed-loop performance. To achieve robust

stability, techniques such as the dual mode or hybrid controllers can be deployed. The I-O

linearization is generally used as an inner loop controller to maintain stability. The

 119
Lyapunov direct method is used as an ‘outer’ controller to handle the uncertainty and

force the state variables to the desired steady state to guarantee the stability of the ‘inner’

controller.

Software for nonlinear model-based control

The prototype software for analytical model-based controller design can be improved

further to make it more convenient for engineering practice. Suggestions for the software

improvement are as follows:

- The software should be flexible to add, remove or update the control methods into the

software environment.

- When the process has a highly nonlinear complex model, an efficient algorithm that

can speed up the calculation of the steady state pairs is required.

- The algorithm of zero dynamics calculation should not only be applicable to non-

redundant equations.

- The software should support controller design for broader classes of processes

- Currently, the software needs Mathematica as the symbolic calculation engine. Future

controller design software should be completely stand-alone applications.

- The software that integrates the controller design into real-time implementation is

also needed. Software such as LABVIEW is generally used to acquire real-time data

and supports the real-time implementation of PID controllers on the process.

Controller design software that can interface with LABVIEW will be of interest to

control engineers.
 120
List of References

[1] Kanter, J. M., M. Soroush, and W. D. Seider, "Nonlinear controller design for
input-constrained, multivariable processes," Ind. Eng. Chem. Res., 41 (16), 3735
(2002).

[2] Kravaris, C., P. Daoutidis, "Nonlinear state feedback control of second-order

nonminimum-phase nonlinear systems," Comp. Chem. Eng., 14 (4-5), 439 (1990).

[3] Ball, R., B. F. Gray, "Transient thermal behavior of the hydration of 2,3 epoxy-
1propanol in a continuously stirred tank reactor," Ind. Eng. Chem. Res., 34, 3726
(1995).

[4] Heemskerk, A. H., J. M. H. Fortuin, "The hysteresis of steady states in gradientless

liquid-phase reaction system," Chem. Eng. Sci., 38, 1261 (1983).

[5] van de Vusse, J. G., "Plug-flow type reactor versus tank reactor," Chem. Eng. Sci.,
19 (12), 994 (1964).

[6] Agrawal, P., C. Lee, H. C. Lim, and D. Ramkrishna, "Theoretical investigations of

dynamic behavior of isothermal continuous stirred tank biological reactors," Chem.
Eng. Sci., 37 (3), 453 (1982).

[7] Brengel, D. D., W. D. Seider, "Multistep nonlinear predictive controller," Ind. Eng.
Chem. Res., 28 (12), 1812 (1989).

[8] Dochain, D., "Adaptive control algorithms for nonminimum phase nonlinear
bioreactors," Comp. Chem. Eng., 16 (5), 449 (1992).

[9] Henson, M. A., D. E. Seborg, "Nonlinear control strategies for continuous

fermenters," Chem. Eng. Sci., 47 (4), 821 (1992).

[10] Jacobsen, E. W., "On the dynamics of integrated plants--non-minimum phase

behavior," J. Process Control, 9 (5), 439 (1999).

[11] Prokopakis, G. J., W. D. Seider, "Azeotropic distillation towers: steady-state and

dynamic simulation," Comp. Chem. Eng., 9 (5), 416 (1985).
 121
[12] Kumar, A., P. Daoutidis, "Modeling, analysis and control of ethylene glycol
reactive distillation column," AIChE J., 45 (1), 51 (1999).

[13] Ogunnaike, B. A., W. H. Ray, Process dynamics, modeling and control. New York:
Oxford, 1994.

[14] Mayne, D. Q., J. B. Rawlings, C. V. Rao, and P. O. M. Scokaert, "Constrained

model predictive control: Stability and optimality," Automatica, 36 (6), 789 (2000).

[15] Isidori, A., Nonlinear control systems. New York: Springer Verlag, 1995.

[16] Dubljevic, S., N. Kazantzis, "A new Lyapunov design approach for nonlinear
systems based on Zubov's method," Automatica, 38 (11), 1999 (2002).

[17] El-Farra, N. H., P. D. Christofides, "Coordinating feedback and switching for

control of hybrid nonlinear processes," AIChE J., 49 (8), 2079 (2003).

[18] El-Farra, N. H., P. Mhaskar, and P. D. Christofides, "Uniting bounded control and
MPC for stabilization of constrained linear systems," Automatica, 40 (1), 101
(2004).

[19] Kravaris, C., J. C. Kantor, "Geometric methods for nonlinear process-control.1.

Background," Ind. Eng. Chem. Res., 29 (12), 2295 (1990).

[20] Kravaris, C., J. C. Kantor, "Geometric methods for nonlinear process-control.2.

Controller synthesis," Ind. Eng. Chem. Res., 29 (12), 2310 (1990).

[21] Niemiec, M. P., C. Kravaris, "Nonlinear model-state feedback control for

nonminimum-phase processes," Automatica, 39 (7), 1295 (2003).

[22] Chen, D. G., B. Paden, "Stable inversion of nonlinear non-minimum phase

systems," Int. J. Contr., 64 (1), 81 (1996).

[23] Devasia, S., "Approximated stable inversion for nonlinear systems with
nonhyperbolic internal dynamics," IEEE Trans. Auto. Contr., 44 (7), 1419 (1999).

[24] Devasia, S., D. Chen, and B. Paden, "Nonlinear inversion-based output tracking,"
IEEE Trans. Auto. Contr., 41 (7), 930 (1996).
 122
[25] Doyle, F. J., F. Allgower, and M. Morari, "A normal form approach to approximate
input-output linearization for maximum phase nonlinear SISO systems," IEEE
Trans. Auto. Contr., 41 (2), 305 (1996).

[26] Hunt, L. R., G. Meyer, "Stable inversion for nonlinear systems," Automatica, 33
(8), 1549 (1997).

[27] Isidori, A., A. Astolfi, "Disturbance attenuation and H∞ control via measurement
feedback in nonlinear systems," IEEE Trans. Auto. Contr., 37 (9), 1283 (1992).

[28] Isidori, A., C. I. Byrnes, "Output regulation of nonlinear systems," IEEE Trans.
Auto. Contr., 35 (2), 131 (1990).

[29] Kravaris, C., M. Niemiec, and N. Kazantzis, "Singular PDEs and the assignment of
zero dynamics in nonlinear systems," Syst. Control Lett., 51 (1), 67 (2004).

[30] McLain, R. B., M. J. Kurtz, and M. A. Henson, "Habituating control for nonsquare
nonlinear processes," Ind. Eng. Chem. Res., 35 (11), 4067 (1996).

[31] Mickle, M. C., R. Huang, and J. J. Zhu, "Unstable, nonminimum phase, nonlinear
tracking by trajectory linearization control," Proc. IEEE Int. Conf. Control Appl., 1,
812 (2004).

[32] Tomlin, C. J., S. S. Sastry, "Bounded tracking for non-minimum phase nonlinear
systems with fast zero dynamics," Int. J. Contr., 68 (4), 819 (1997).

[33] van der Schaft, A. J., "L2-gain analysis of nonlinear systems and nonlinear state-
feedback H∞ control," IEEE Trans. Auto. Contr., 37 (6), 770 (1992).

[34] Scokaert, P. O. M., D. Q. Mayne, and J. B. Rawlings, "Suboptimal model

predictive control (feasibility implies stability)," IEEE Trans. Auto. Contr., 44 (3),
648 (1999).

[35] Chen, H., F. Allgower, "A quasi-infinite horizon nonlinear model predictive control
scheme with guaranteed stability," Automatica, 34 (10), 1205 (1998).

[36] Zheng, A., M. Morari, "Stability of model predictive control with mixed
constraints," IEEE Trans. Auto. Contr., 40 (10), 1818 (1995).
 123
[37] Mayne, D. Q., "Nonlinear model predictive control: An assessment," in Fifth
International Conference on Chemical Process Control, CACHE, AIChE, J. C.
Kanter, C. E. Garcia, and B. Carnahan, Eds., 1997, pp. 217.

[38] Scokaert, P. O. M., J. B. Rawlings, "Infinite horizon linear quadratic control with
constraints," Proceedings of the 13th World Congress, International Federation of
Automatic Control, M, 109 (1996).

[39] Rawlings, J. B., K. R. Muske, "The stability of constrained receding horizon

control," IEEE Trans. Auto. Contr., 38 (10), 1512 (1993).

[40] de Oliveira Kothare, S. L., M. Morari, "Contractive model predictive control for
constrained nonlinear systems," IEEE Trans. Auto. Contr., 45 (6), 1053 (2000).

[41] Gobin, F., L. C. Zullo, and J. P. Calvet, "Model-predictive control of an open-loop

unstable train of polymerization reactors," Comp. Chem. Eng., 18, 525 (1994).

[42] Nagrath, D., V. Prasad, and B. W. Bequette, "A model predictive formulation for
control of open-loop unstable cascade systems," Chem. Eng. Sci., 57 (3), 365
(2002).

[43] Valluri, S., M. Soroush, and M. Nikravesh, "Shortest-prediction-horizon non-linear

model-predictive control," Chem. Eng. Sci., 53 (2), 273 (1998).

[44] Soroush, M., C. Kravaris, "MPC formulation of GLC," AIChE J., 42 (8), 2377
(1996).

[45] Hernandez, E., Y. Arkun, "Study of the control-relevant properties of

backpropagation neural network models of nonlinear dynamic-systems," Comp.
Chem. Eng., 16 (4), 227 (1992).

[46] Soroush, M., C. Kravaris, "A continuous-time formulation of nonlinear model

predictive control," Int. J. Contr., 63 (1), 121 (1996).

[47] Soroush, M., K. Muske, "Analytical model predictive control," in Nonlinear model
predictive control, vol. 26, Progress in systems and control theory series, F.
Allgower and A. Zheng, Eds. Basel: Birkhauser-Verlag, 2000, pp. 163.
 124
[48] Slotine, J.-J. E., W. Li, Applied nonlinear control. Englewood Cliffs, N.J.: Prentice
Hall, 1991.

[49] Lin, Y. D., E. D. Sontag, "A universal formula for stabilization with bounded
controls," Syst. Control Lett., 16 (6), 393 (1991).

[50] Freeman, R. A., J. A. Primbs, "Control Lyapunov functions: new ideas from an old
source," Proc. IEEE Conf. Decis. Control, 4, 3926 (1996).

[51] Sepulchre, R., M. Jankovic, and P. kokotovic, Constructive nonlinear control:

Springer, 1996.

[52] Freeman, R. A., P. V. Kokotovic, Robust nonlinear control design: state-space and
Lyapunov techniques. Boston: Birkhauser, 1996.

[53] Li, Z.-H., M. Krstic, "Maximizing regions of attraction via backstepping and CLFs
with singularities," Syst. Control Lett., 30 (4), 195 (1997).

[54] Kazantzis, N., C. Kravaris, "Energy-predictive control: a new synthesis approach

for nonlinear process control," Chem. Eng. Sci., 54 (11), 1697 (1999).

[55] Kazantzis, N., C. Kravaris, "Singular PDEs and the single-step formulation of
feedback linearization with pole placement," Syst. Control Lett., 39 (2), 115 (2000).

[56] DeCarlo, R., S. Drakunov, "Sliding mode control design via Lyapunov approach,"
Proc. IEEE Conf. Decis. Control, 2, 1925 (1994).

[57] Lin, W., C. Qian, "New results on global stabilization of feedforward systems via
small feedback," Proc. IEEE Conf. Decis. Control, 1, 873 (1998).

[58] Jankovic, M., R. Sepulchre, and P. V. Kokotovic, "Global adaptive stabilization of

cascade nonlinear systems," Automatica, 33 (2), 263 (1997).

[59] Mazenc, F., L. Praly, "Adding an integration and global asymptotic stabilization of
feedforward systems," Proc. IEEE Conf. Decis. Control, 1, 121 (1994).

[60] Larsen, M., P. Kokotovic, "Passivation design for a turbocharged diesel engine
model," Proc. IEEE Conf. Decis. Control, 2, 1535 (1998).
 125
[61] Kazantzis, N., C. Kravaris, "Synthesis of state feedback regulators for nonlinear
processes," Chem. Eng. Sci., 55 (17), 3437 (2000).

[62] Wu, W., "Adaptive-like control methodologies for a CSTR system with dynamic
actuator constraints," J. Process Control, 13 (6), 525 (2003).

[63] Michalska, H., D. Q. Mayne, "Robust receding horizon control of constrained

nonlinear systems," IEEE Trans. Auto. Contr., 38 (11), 1623 (1993).

[64] Mhaskar, P., N. H. El-Farra, and P. D. Christofides, "Hybrid predictive control of

process systems," AIChE J., 50 (6), 1242 (2004).

[65] Barker, H. A., P. W. Grant, and M. Zhuang, "Control system analysis with
Mathematica," IEE Colloquium on Symbolic Computation for Control, 2/1 (1996).

[66] Barker, H. A., M. Zhuang, "Control system analysis using Mathematica and a
graphical user interface," Comput. Control Eng. J., 8 (2), 64 (1997).

[67] Bemporad, A., M. Morari, "Optimization-based hybrid control tools," Proc. of

ACC, 2, 1689 (2001).

[68] Blankenship, G. L., H. G. Kwatny, C. LaVigna, and V. Polyakov, "Integrated

modeling and design of nonlinear control systems," Proc. of ACC, 3, 1395 (1997).

[69] Campbell, S. L., F. Delebecque, and D. von Wissel, "A mixed symbolic-numeric
software environment," Proc. IEEE Int. Symp. Comput. Aid Control Syst. Des., 436
(1996).

[70] Datta, B. N., D. Sarkissian, "Numerical linear control library;-a Mathematica-based

integrated environment with the modern control algorithms," Proc. IEEE Int. Symp.
Comput. Aid Control Syst. Des., 91 (1999).

[71] de Jager, B., "The use of symbolic computation in nonlinear control: is it viable?"
IEEE Trans. Auto. Contr., 40 (1), 84 (1995).

[72] Kitamoto, T., M. Saeki, and K. Ando, "CAD package for control system on
Mathematica," IEEE International Conference on Systems Engineering, 448
(1992).
 126
[73] Ohtani, T., M. Fukuzawa, and M. Masubuchi, "A CAD system for nonlinear control
system design using Mathematica," Proceedings of IEEE/IFAC Joint Symposium
on Computer-Aided Control System Design, 197 (1994).

[74] Sack, J., T. Singh, "Automated design of model predictive controllers," Proc. of
ACC, 6, 3758 (2000).

[75] MATHEMATICA, 4 ed. Champaign, IL: Wolfram Research, Inc., 1999.

[76] Maple6. Waterloo, Canada: Waterloo Maple, Inc., 2000.

[77] MATLAB, 7.0 ed. Natick, MA: The Mathworks, Inc., 2002.

[78] www.scilab.org.

[79] Mickle, M. C., R. Huang, and J. J. Zhu, "Proceedings of the IEEE International
Conference on Control Applications," Proc. IEEE Int. Conf. Control Appl., 1, 812
(2004).

[80] Soroush, M., H. M. Soroush, "Input-output linearizing nonlinear model predictive

control," Int. J. Contr., 68 (6), 1449 (1997).

[81] Soroush, M., "Nonlinear state-observer design with application to reactors," Chem.
Eng. Sci., 52 (3), 387 (1997).

[82] Wright, R. A., C. Kravaris, "Nonminimum-phase compensation for nonlinear

processes," AIChE J., 38 (1), 26 (1992).

[83] Kravaris, C., M. Niemiec, R. Berber, and C. B. Brosilow, "Nonlinear model-based

control of nonminimum-phase processes," in Nonlinear Model-based Control, R.
Berber and C. Kravaris, Eds. Dordrecht: Kluwer Academic, 1998.

[84] Niemiec, M., C. Kravaris, "Controller synthesis for multivariable nonlinear non-
minimum-phase processes," Proc. of ACC, 2076 (1998).

[85] Allgower, F., F. J. Doyle, "Approximate I/O-linearization of nonlinear systems," in

Nonlinear Model-based Control, R. Berber and C. Kravaris, Eds. Dordrecht:
Kluwer Academic, 1998.
 127
[86] Kanter, J. M., M. Soroush, and W. D. Seider, "Continuous-time, nonlinear feedback
control of stable processes," Ind. Eng. Chem. Res., 40 (9), 2069 (2001).

[87] Soroush, M., C. Kravaris, "A continuous-time formulation of nonlinear model

predictive control," Proc. of ACC, 1561 (1992).

[88] Artstein, Z., "Stability in the presence of singular perturbations," Nonlinear

Analysis, 34, 817 (1998).

[89] Artstein, Z., "Relaxation in singularly perturbed control systems," Proc. IEEE Conf.
Decis. Control, 4, 4330 (2002).

[90] Fridman, E., "Exact slow-fast decomposition of the nonlinear singularly perturbed
optimal control problem," Syst. Control Lett., 40 (2), 121 (2000).

[91] Daoutidis, P., M. Soroush, and C. Kravaris, "Feedforward/Feedback Control of

Multivariable Nonlinear Processes," AIChE J., 36 (10), 1471 (1990).

[92] Kanter, J. M., M. Soroush, and W. D. Seider, "Nonlinear feedback control of

multivariable non-minimum-phase processes," J. Process Control, 12 (6), 667
(2002).

[93] Panjapornpon, C., M. Soroush, and W. D. Seider, "A model-based control method
applicable to unstable, non-minimum-phase, nonlinear processes," Proc. of ACC, 4,
2921 (2004).

[94] Panjapornpon, C., M. Soroush, and W. D. Seider, "Nonlinear output-tracking with

guaranteed asymptotic stability," submitted to Automatica (2005).

[95] Soroush, M., "State and parameter estimations and their applications in process
control," Comp. Chem. Eng., 23 (2), 229 (1998).

[96] Kravaris, C., M. Soroush, "Synthesis of multivariable nonlinear controllers by

input/output linearization," AIChE J., 36 (2), 249 (1990).

[97] Seborg, D. E., T. F. Edgar, and D. A. Mellichamp, Process dynamics and control,
second ed: John Wiley & Sons, Inc, 2004.
 128
[98] Ramaswamy, S., T. J. Cutright, and H. K. Qammar, "Control of a continuous
bioreactor using model predictive control," Process Biochem., 40, 2763 (2005).
 129
Vita

Name : Chanin Panjapornpon

Nationality : Thai

Date of Birth : May 20, 1973

Education : Ph.D. in Chemical Engineering, Department of Chemical and

Biological Engineering, Drexel University (2002-2005)

MS. in Chemical Engineering, Department of Chemical Engineering,

Drexel University (2000-2002)

B.Sc. in Chemical Engineering, Department of Chemical

Technology, Faculty of Science Chulalongkorn University, Bangkok
Thailand (1991-1995)

Experience : Vinyl Chloride Monomer Quality and Development coordinator,

Vinylthai Public Co. Ltd. , Thailand (1999-2000)

Vinyl Chloride Monomer Assistant Operation Manager, Vinylthai

Public Co. Ltd. , Thailand (1997-1998)

Vinyl Chloride Monomer Shift Superintendent, Vinylthai Public Co.

Ltd. , Thailand (1995-1997)

Awards : Scholarships from Royal Thai Government to study in MS.- PhD.

Program in Chemical Engineering

B.Sc.(2nd Honor) in Chemical Engineering at Chulalongkorn

University, Thailand

Publications:

1. Panjapornpon, C., M. Soroush, and W.D. Seider, “Model-Based Controller Design for
Unstable, Non-Minimum-Phase, Nonlinear Processes,” submitted to Ind. Eng. Chem.
Research (2005)

2. Panjapornpon, C., and M. Soroush, “Asymptotic Output Tracking with Guaranteed

Asymptotic Stability,” to be submitted to Automatica (2005)
 130
3. Panjapornpon, C., M. Soroush, and W.D. Seider, “Software for Analytical Model-
Based Controller Design,” submitted to Computers and Chemical Engineering (2005)

4. Panjapornpon, C., N. Fletcher, and M. Soroush, “A Flexible Pilot-Scale Setup for

Real-Time Studies in Process Systems Engineering,'' Chem. Eng. Education, in press
(2005)

5. Panjapornpon, C., and M. Soroush, “Real-Time Experiments in Model-Based Control

and On-Line Parameter Estimation,’’ submitted to Chem. Eng. Education (2005)

6. Mehranbod, N., M. Soroush, and C. Panjapornpon, “A Method of Sensor Fault

Detection and Identification,'' J. of Process Contr., 15(3), 321-339 (2005)