Xiong Qiang 2007

This
document is downloaded from DR‑NTU (https://dr.ntu.edu.sg)
Nanyang Technological University, Singapore.
Multivariable control system design based on
relative interaction energy
Xiong, Qiang
2007
Xiong, Q. (2007). Multivariable control system design based on relative interaction energy.
Doctoral thesis, Nanyang Technological University, Singapore.
https://hdl.handle.net/10356/39072
https://doi.org/10.32657/10356/39072
Downloaded on 11 Nov 2022 04:04:24 SGT

ATTENTION: The Singapore Copyright Act applies to the use of this document. Nanyang Technological University Library
Multivariate Control System Design

Based on
Relative Interaction Energy
Xiong Qiang
School of Electrical & Electronic Engineering
A thesis submitted to the Nanyang Technological University

in fulfillment of the requirement for the degree of
Doctor of Philosophy
2007
Acknowledgments i
Acknowledgments
I would like to acknowledge all of the people who guide, help and support me
during the undertaking of this research.
First and foremost. I would like to express my sincere gratitude to my

supervisor, Associate Professor Cai Wen-Jian, for his professional guidance and
invaluable advices throughout the past four years. His insightful comments and
thoughtful discussion have been an inspiration for my research.
I would also like to thank all of my colleagues, friends and the technicians in
Process and Instrumentation Laboratory. Their friendship and help will always be
treasured.
Finally, my deep thanks go to all of my family members for their

understanding and encouragement. I save the last special thanks for my wife. She
stands by me through the most difficult time of this period, offers her
unconditional love and support.
Nanyang Technological University

Table of Contents i!
Table of Contents
Acknowledgments i
Table of Contents ii
Summary iv
List of Figures vi
List of Tables viii
Chapter 1 1
Introduction 1
1.1. Motivations 1
1.2. Objectives 2
1.3. Major contributions of the thesis 2
1.4. Organization of the thesis 3
Chapter 2 5
Literature Review 5
2.1. Introduction 5
2.2. Preliminaries 8
2.3. Interaction analysis and loop pairing 10
2.3.1. RGA based loop paring 12
2.3.2. RIA based loop paring 15
2.3.3. Dynamic RGA (DRGA) based loop paring 18
2.3.4. GRDG based loop paring 20
2.3.5. SSV based loop paring 21
2.4. Multi-loop controller design 23
2.4.1. Detuning methods 24
2.4.2. Sequential loop closing methods 27
2.4.3. Independent design methods 29
2.4.4. Simultaneous equation solving methods 31
2.5. Decoupling controller design 34
2.5.1. Ideal/dynamic decoupling methods 37
2.5.2. Static decoupling methods 39
2.5.3. Partial/block decoupling methods 41
References 43
Chapter 3 49
Interaction Analysis and Loop Pairing for Multivariate Processes 49
3.1. Introduction 49
3.2. Preliminaries 51
3.3. A modified loop pairing rule 52
3.4. Case studies 58
3.5. Conclusions 61
References 61
Chapter 4 63
Decentralized Control System Design 63

Table of Contents ill
4.2. Equivalent transfer function 64
4.3. SISO controller design 71
4.5. Conclusions 87
References 87
Chapter 5 .' 89
Decoupling Control System Design 89
5.2. General formulation of MIMO control 90
5.3. Full decoupling control 92
5.4. Block decoupling control 94
5.6. Conclusions 109
References 109
Chapter 6 112
Application to an Experimental HVAC System 112
6.2. Four-room air temperature control system 115
6.3. Modeling and simulations 117
6.4. Experiment results 123
6.5. Conclusions 126
References 126
Chapter 7 128
Conclusion and Recommendations 128
7.1. Conclusion 128
7.2. Recommendations 129
Author's Publications 131

Summary IV
Summary
For the design of multi-input multi-output (MIMO) system, two major tasks
have to be resolved, i.e., loop pairing and controller design. For the first task, one
needs to pair the input and output variables in such a way that the interactions
among different loops are smallest. The other task is control system design
according to paired variables. Generally, a good pairing results in relatively easy
controller design. Unfortunately, both loop pairing and controller design are
difficult since there are complicated interactions among different loops. Especially
for controller design, adjusting controller parameters of one loop will affect the
performance of others, sometimes to the extent of destabilizing the entire system.
To ensure stability, many industrial controllers for multivariable processes are
tuned loosely, which causes inefficient operation and higher energy costs.
In this thesis, relative energy array (REA) is defined by using steady-state
gains and critical frequencies of the transfer function matrix elements, and used to
provide a simple yet comprehensive description of loop interactions for MIMO
processes. The value of REA can be conveniently calculated by control engineers
since the critical frequencies can be easily obtained from given transfer function
matrix. Furthermore, REA is used to derive simple equivalent transfer functions
(ETFs) for each transfer function element under closed-loop control, which is
fundamental in control system design. Consequently, the decentralized, full
decoupling and block decoupling controllers can be easily designed by simply
using single loop design approaches. The method is very simple, straightforward
and easily understood by control engineers. The advantage of the proposed
method is even more significant when applied to higher dimensional processes
with complicated interaction modes. As a matter of fact, higher dimension does
not increase the design difficulty. The simulation results for a variety of industrial
processes show that the performance of the proposed controller design method is
much better as compared with the existing approaches. The proposed method has

Summary _
also been implemented in an experimental HVAC system to illustrate its

effectiveness.

List of Figures \i
List of Figures
Figure 2.1. Hierarchy of control system functions in a typical processing plant

Figure 2.2. Closed-loop multivariate control system
Figure 2.3. Definition of dynamic interaction in uJ-yi loop
Figure 3.1. Frequency response curves and the energy, et), of g^jco)
Figure 4.1. Interaction mode with An < 1, yn < 1
Figure 4.2. Interaction mode with An < 1, yu > 1
Figure 4.3. Interaction mode with An > 1. yn < 1
Figure 4.4. Interaction mode with kn > 1, yn > 1
Figure 4.5 Closed-loop responses for Example 4.1

Figure 4.6. Closed-loop responses for Example 4.2
Figure 4.7. Closed-loop responses for Example 4.2 (model mismatched)
Figure 5.1. Column dominance for Example 5.1
Figure 5.2. Closed-loop responses for Example 5.1 (nominal process)
Figure 5.3. Singular value based robustness index for Example 5.1
Figure 6.1. Scheme of a typical building HVAC system
Figure 6.2. Experimental centralized HVAC system
Figure 6.3. Four-room HVAC system (Top View)
Figure 6.4. Closed-loop responses of four-room system

List of Figures Ml
Figure 6.5. Closed-loop responses of four-room system (model mismatched)

Figure 6.6. Real-time closed-loop responses of four-room system
Figure 6.7. Real-time manipulated variables of four-room system

List of Tables
List of Tables
Table 4.1. Typical gain and phase margin values

Table 4.2. ETFs and PID controllers for gain and phase margin method
Table 4.3. Controllers for Example 4.1
Table 6.1. Decentralized controllers for HVAC system

i
Introduction
Chapter 1
Introduction
1.1. Motivations
Most industrial operating units are naturally characterized by MIMO

processes. However, MIMO controllers are more difficult to design even if the
controllers themselves are specified with simple structures. The efficient tuning of
decentralized PID controllers for multivariable processes is a very complex and
time-consuming procedure. To relieve the pain of tuning MIMO systems, there
has been strong interest in developing effective control system design approaches
which are easy to use and effective in solving real process control problems.
Even though considerable effort has been dedicated to this problem and a
great many design techniques have been proposed over recent years, there is no
simple and yet systematic approach for the design of MIMO control systems. A
unified and generally accepted definition is still lacking. Important issues, such as
the effects of interaction on closed-loop performance, the measure of the severity
of interaction, and the impact of variable pairing and controller tuning on
interaction and subsequently on closed-loop performance, are far from being
resolved.
Consequently, multivariable control system design and implementation

continues to present a challenge to researchers as well as practical engineers alike.
As a matter of fact, only very few of them have been successfully applied to solve
real world problems.

Introduction :
1.2. Objectives
Adjusting controller parameters of one loop affects the performance of others,

sometimes to the extent of destabilizing the entire system. To ensure stability,
many industrial multi-loop controllers are tuned loosely, which causes inefficient
operation resulting in higher energy costs and shorten the equipment life. The task
of developing a satisfactory design procedure for multivariate process control is
a big challenge and still remains as an open topic. This research is focused on
developing a method for simple interaction measure and the
decentralized/decoupling controller design for MIMO systems.
In summary, this thesis attempts to develop a practical and engineering

oriented design methodology for the control of interactive MIMO systems.
1.3. Major contributions of the thesis
In this thesis, a systematic design approach for decentralized/decoupling

control systems is proposed for general MIMO processes. Through intensive
investigations on interaction analysis and controller tuning, several significant
results are obtained. The major contributions of the thesis are summarized as
follows:
(1) A novel interaction measure and loop pairing criterion for decentralized
control of multivariate processes is proposed. Both steady-state gain and
critical frequency of the elements in the transfer function matrix are used
to provide an effective description of loop interactions. The criterion can
be conveniently applied by control engineers since the information needed
can be easily obtained from the process model. It gives accurate interaction
assessment even for some complex processes while the conventional
relative gain array (RGA) based loop pairing criterion may result in a
wrong conclusion.
(2) A novel method for decentralized control system design based on the
proposed interaction measure is proposed. The loop interactions are

Introduction 3
quantified by relative gain and relative critical frequency. The interactions

for a particular loop from all other closed loops are then analyzed through
both steady-state gain and critical frequency variations. Consequently, the
equivalent transfer function is constructed and the decentralized controllers
can be independently designed by employing single loop tuning techniques.
Differing from existing equivalent transfer functions, the proposed
equivalent transfer function provides both gain and phase information for
decentralized controller design in a simple and straightforward manner.
(3) A novel full decoupling/block, decoupling design approach based on the

equivalent transfer function matrix for closed-loop system is proposed.
Based on this matrix, both off-diagonal controllers and main-loop diagonal
controllers can be easily designed using the existing PI/PID tuning rules.
The decoupling controller obtained in this way has good set-point tracking
and disturbance rejection properties. It still works with satisfactory
performance even if the primary process parameters have significant
variations as compared with existing decoupling designs.
1.4. Organization of the thesis
The rest of thesis is organized as follows:
Chapter 2 gives a review in the area of multivariable control system,

especially in control system configuration, decentralized and decoupling control
system design.
Chapter 3 presents an interaction measure and loop pairing criterion based on

the concept of relative energy.
Chapter 4 presents a decentralized controller design approach based on the

closed-loop equivalent transfer functions.
Chapter 5 presents a framework for full decoupling and block decoupling

controller design approaches.
Chapter 6 describes the application of the proposed approach to the room

4
Introduction
temperature control of an experimental HVAC system.
Chapter 7 concludes the thesis and suggests some possible future research
topics in this field.

5
Literature Review
Chapter 2
Literature Review
2.1. Introduction
In modern processes, there are mainly two multi-level hierarchies of control

functions as shown in Figure 2.1 (Qin and Badgwellb, 2003). A conventional
control structure is shown in Figure 2.1(a) for n operating units, while Figure 2.1
(b) shows a MPC structure for the same n units. In both hierarchy structures, a
plant-wide optimizer determines optimal steady-state settings for each unit in the
plant at the top of the structure. These may be sent to local optimizers at each unit
which run more frequently or consider a more detailed unit model at the plant-
wide level. The unit optimizer computes an optimal economic steady-state and
passes this to the dynamic constraint control system for implementation. In
conventional structure (Figure 2.1(a)), it is accomplished by using a combination
of PID algorithms, lead-lag (L/L) blocks and high/low select logic. While in MPC
structure (Figure 2.1(b)), this combination of blocks is replaced by a single MPC
controller which using dynamic constraint control to move the plant from one
constrained steady-state to another while minimizing constraint violations along
the way. In any hierarchy structure, however, the bottom level regulating loops are
PID controller based DCS systems. As the performance of the bottom level is
instrumental in plant-wide optimization, it is an important part of a multi-level
hierarchy of control functions.

6
Literature Review
Global Economic
3
lant •Vide Optimizat on Optimization
(every day)
Local Economic
Unit 1 Local Optimizer Unit n Local Optimize' Optimization
(eve*> nojr)
T
h g h ' t o * Select Logi: H g h ' L o * Select _og c
Dynamc
FIC PID PC L'L PIO Constraint
Control
(eve-y minute)
SUM SUM SUM S'JM

T
Bas c Dynamic
Un11 DCS- PID Controls Unit n D C S - P D Controis Control
(every second;
© ^) S / fc) $£) <xB)
(a) Conventional structure
Unitl Unitn
Global Econonvc
Optimization
=lant-' .Vide Optimizat on
leveny day;
Local Economic
Unit 1 Local Optimizer Unit n Local Cotimize* Optimization

(eve<Y nojr)
T
Dynamic
Model Preofctive Control Model Predictive Control Constraint
IMPC) (MPC) Control
;eve*y minute)
Bas c Dynamc
Unit" DCS- PID Controls Unit n D C S - P D Contras Control
(every second j
©©
(b) MPC structure
Figure 2.1. Hierarchy of control system functions in a typical processing plant

7
Literature Review
Most industrial operating units are naturally characterized by MIMO

processes. MIMO controllers are more difficult to design even if the controllers
themselves (PID) are specified with simple structures. The efficient tuning of
decentralized PID controllers for multivariable processes is a very complex and
time-consuming procedure. Generally, a multivariable process can be controlled
either by a multivariable (centralized), a decoupling or a multi-loop
(decentralized) controller. Each control structure has advantages and
disadvantages. For industrial multivariable processes, the choice of an appropriate
control structure depends on the process characteristics, the economic value of
improved control, and the available computing and process control resources. The
various control structures are evaluated by taking into account the complexity of
implementation, the robustness of the algorithm for parametric variations, and the
performance for set-point changes and disturbance rejection.
Indeed, decentralized PI/PID controllers either in the simple multi-loop form

or decoupling based multi-loop form remain the standard in process control
applications (Campo and Morari, 1994). especially in large scale industrial
processes, even though the performance may be inferior owing to the constraint
on the controller structure. The underlying advantages responsible for the wide
acceptance of decentralized control structures are:
(1) Hardware simplicity: The cost of implementing of a decentralized control

system is significantly less than that of a centralized controller. A
centralized control system for an n-input, ^-output plant usually consists of
«x« individual single-input single-output (SISO) transfer functions, which
significantly increases the complexity of the controller hardware.
Furthermore, if the controlled and/or manipulated variables are physically
far apart, a full controller could require numerous expensive
communication links. Decentralized control structure, on the other hand,
allows the use of already existing equipment and easy to implement and be
understood.
(2) Design and tuning simplicity: Decentralized controllers involve far fewer
parameters, resulting in a significant reduction in the time and cost of
tuning. For PID control of a two-input two-output (TITO) process, for

Literature Review s
example, only 6 parameters need to be tuned instead of possible 12 for

decoupling control and possible even more for centralized control.
(3) Robustness: The failure tolerance is usually easier to achieve in

decentralized control systems (Morari and Zafiriou, 1989). In many
systems, it is important that stability and performance are preserved to
some degree when some sensors or actuators fail.
Decentralized and decoupling control design consist of two stages: loop

pairing and multi-loop or decoupling controller design. Considerable effort has
been dedicated to this problem and a great many design techniques have been
proposed in recent years. The main achievements will be briefly reviewed in the
following sections after some preliminaries.
2.2. Preliminaries
A multivariate control system with n inputs and n outputs can be typically

shown as in Figure 2.2, where rl,i = \,2,---n , are the reference inputs,
u„ i = l,2,•••« , are the manipulated variables. di,i = l,2,---n, are the output
disturbances, yi,i = l,2,---n, are the system outputs, G(s) is the process transfer
function matrix expressed by

'gu(s) gn(s) ... gln(s)~
gll(s) g2l(s) - g2n(s)
G(s) = (2.1)
JAS) gn2(S) - gm(s)

Gc(s) and Gd(s) are the controller and disturbance transfer function matrix with
compatible dimensions whose structures are determined by the chosen control
schemes, respectively.

9
Literature Review
di d2 dn
I
ri +^ e, 111
»
+1 yi
»
—^v +
V2 - 1 e: 112 y2.
-£<5* * PQ9
-i Gc(s) G(s)
rn
y
•"6 P^ Un r i yn.
-ry
Figure 2.2. Closed-loop multivariable control system

In order to make use of SISO techniques, a multivariable process can be
controlled by either decentralized, decoupling or block decoupling control
schemes. The choice of the appropriate control scheme depends on the process
characteristics, the economic value of improved control, and the available
computing and process control resources. The pros and cons as well as the bases
of selecting different control schemes may be stated as follows.
Decentralized control: It has a diagonal controller structure
"*ci(*) 0 - 0
0 gc2(s) - 0
GM = (2.2)
0 0 ... gcn{s)_
and it is often used for processes with mild interactions. The control structure
simplicity, easiness to design and implement, and loop failure tolerance (i.e.,
without the need for readjustment to the other parts of the control system to
preserve stability in the case of a sensor failure and/or actuator failure) are the
main advantages.
Full decoupling control: It has a full controller structure
8cAs) 8c,n(s) - gc.i„(*)~

Sc.2l(J) Sc,22(*) - Sc,2„(s)
Ge(s) = (2.3)
_gcAS) Sc.n2(S) " 8c,nn(S)_
and it is often adopted for two-dimensional closely coupled processes or systems

which require higher control quality. By eliminating the undesirable cross-
couplings responsible for the control loop interactions through decoupling

Literature Review 10
controller, the control system performance and robustness can be improved. The
drawbacks may be the sacrifice of structure simplicity and control system integrity.
Block decoupling control: It has a block diagonal controller structure
"G cl (s) 0 ... 0

0 Gc2(s) ... 0
Gc(s) = (2.4)
0 0 ... Gcm(s)
where m<n , and Gcl(s), l = \...m , are controller sub-matrices. It is a

compromise between the above two schemes for high dimensional processes. It
only decouples those closely coupled subsystems while the whole structure is still
decentralized. The integrity may be kept intact between the sub-blocks.
Define the sensitivity and complementary sensitivity functions as
S(S)±(I + G(s)Gc(s))]
and
T(s)^G(s)Gc(s)(I + G(S)Gc(s)yi
respectively, where, T(s) is the closed-loop transfer function from the reference
inputs, r, to the system outputs, v, and S(s) is the closed-loop transfer function
from the output disturbances, d, to the system outputs, y.
Using the sensitivity and complementary sensitivity functions, the closed-loop
transfer function of Figure 2.2 can be written as
Y(s) = T(s)R(s) + S(s)D(s) (2.5)
The ultimate performance goals for the closed-loop control system to achieve
are tracking set-point changes, rejecting disturbances, and robustness to parameter
variations. For physical realization, Gc(s)must be selected such that all poles of
T(s) and S(s) lie in the open left halfplane (LHP). Furthermore, for the feedback
control system to have no offset for set-point or disturbance changes, we need
lim T{s) = 1 and lim S(s) = 0.

s->0 j-»0
2.3. Interaction analysis and loop pairing
In addition to ys, the assigned corresponding output variable, the manipulated

11
Literature Review
variable u, also influences other outputs in a closed-loop multivariate control

system. This interaction problem is recognized as one of the main problems for
the control of multivariable processes. For a system with several input and output
variables, it can be configured in several different ways. Each u} to yK or pairing,
constitutes a control configuration. For a TITO system there are two such
configurations:
Configuration 1 Configuration 2
ui—yi u,—y2
u2—yi U2—yi
These two configurations are mutually exclusive, i.e., only one of them can be
used at any particular time. For n*n dimensional systems, there are n! possible
input-output configurations.
Interaction analysis and structure selection play an important role in

decentralized control systems. The major objective in selecting pairing is
primarily to minimize the loop interaction so that the resulting multivariable
control system mostly resembles its SI SO counterparts and the subsequent
controller tuning is largely facilitated by independent design (Seborg and Edger,
1989). This objective is usually accomplished by utilizing various interaction
measures as tools to screen possible pairing alternatives (Bristol, 1966; Mc Avoy,
1983; Zhu and Jutan, 1993). The following aspects are important in the
development of a measure for loop interactions:
(1) Interaction should be defined on the basis of the dynamic framework, since
a rigorous definition of interaction must be valid at more than just steady-
state. This is clear by noticing that there is no interaction at steady-state if
integral action is used and the system is stable;
(2) Interaction should be extracted from the true dynamics of the closed-loop
system without making any overly simplifying assumption about the
system concerned and should have physical significance;
(3) Interaction should be defined in terms of individual control loops in

relation to corresponding interaction-free subsystems, so that direct
implications for controller tuning and variable pairing can be better
explored.

Some popular interaction measures and paring rules for process control will be
summarized below.
2.3.1. RGA based loop paring

Consider the nxn plant G(s) with input w(s)and output y(s) as shown in
Figure 2.2, define relative gain and relative gain array (RGA) as (Bristol, 1966:
Grosdidier and Morari, 1985;Niederlinski, 1971)
J,
\ l J /all loops
tops op
open
A..=
' ( * . / * , all
) *other loops close except for loop y, -u.
and
A=
A, A7 ... A
nl nl m
respectively. Since
y(s) = G(s)u(s),
which implies
u(s) = G-\s)y(s).
Obviously, if all ykk±J are perfectly controlled by the use of uIUl then the gain
from w, to yJ is given by l/[G~'] . The relative gain is actually the ratio of these
"open-loop" and "closed-loop'' gains while 5 = 0 , and can be computed by

(Grosdidier and Morari, 1985)
A = G(0)®G" r (0)
where the operator <8> is the hadamard product and G 7 represents the transpose of
inverse of G.

13
Literature Review
The primary use of the RGA is to determine control loop pairings in multi-
loop control systems, and in assessing the extent of interaction present in such
systems. Further more, Niederlinski (1971) introduced Niederlinski index (NI).
The NI is defined, also using steady-state gain matrix of a process, as below:
det(G(0))_ |G(0)|
NI
det(G(0))
fWo)
where det(G(0)) and |G(0)| denote the determinant of matrix G(0) and
G(0) = diag(G(0)). Under closed-loop conditions in all n loops, the system will
be unstable for all possible values of controller parameters if the NI is negative
(Mc Avoy, 1983; Seborg and Edger. 1989; Skogestad and Lundstrom, 1990). For
NI, the following are worth noticing:
(1) A negative NI is a necessary and sufficient condition for instability only
for TITO systems; for higher dimensional systems, it is only a sufficient
conditions: if it holds then the system is definitely unstable; otherwise, the
system may, or may not be unstable depending on the values taken by the
controller parameters;
(2) For 2><2 systems the NI becomes:
NI = \-%
i—\
where £ = . If £ > 1, the NI is always negative; hence 2*2 systems
paired with negative relative gains are always structurally unstable.

(1) The NI is for systems with rational transfer function elements, thereby
excluding time-delay systems. However, since it depends only on steady-
state gains, the results of this theorem provide useful information about
time-delay systems as well, although the analysis is no longer rigorous. It
should be applied with caution when time delays are involved.
In summary, the rules for using the RGA and NI for loop pairing are:
(1) Given G(s), obtain steady-state gain matrix G(0) and the RGA;
(2) Tentatively choose a loop pairing based on the RGA by pairing on positive
elements that are closest to 1.0;
(3) Use NI to verify the stability status of the control configuration resulting
from step 2: if the pairing is unacceptable, i.e., NI < 0, try another paring;

14
Literature Review
(4) Variables should be paired in such a way that the resulting pairing
corresponds to an NI closest to 1.0. The NI interaction rule is based on
empirical observations of the definition of the NI and is justified largely
from the relationship between the sizes of the RGA. The NI interaction
rule has been found to be capable of avoiding ambiguities in using the
RGA interaction rule.
The RGA based paring is further enhanced by taking into account integrity
(Chiu and Arkun, 1990; Zhu and Jutan, 1993) as well as robustness (Zhu and
Jutan, 1993) considerations, together with the use of the RGA and the NI.
The use of the RGA was popularized by Shinskey (1988). who applied it to
numerous examples, including blending, energy conservation, and distillation. Mc
Avoy (1983) published a monograph in which many of the applications of the
RGA up to that date were presented. From a cost benefit point of view, the RGA
method is highly effective due to the following facts:
(1) It requires minimal process information, and it is in the form of a ratio so

that an approximate process model can give useful results;
(2) It is independent of control system tuning and process disturbances and it

is simple to calculate.
Despite the availability of vast amount of literature on RGA, there are some
common misconceptions. Kariwala and Hovd (2006), in their distinguished work,
pointed out a number of incorrect claims and disproved some conjectures related
to RGA. For example, they showed that RGA alone cannot be used as a definite
measure of two-way interactions in a system, and pointed out that the RGA based
results are time-tested heuristics, and that the theoretical basis for some of the
results is poor. According to the authors, cross-checking of the findings using
alternate methods is essential. In the absence of verification, the chosen control
structure may not be optimal in the sense of missed opportunities and the closed-
loop system may spring many surprises at the practicing engineer.
Indeed, the RGA based rules have been found to have deficiencies, and in
some cases they lead to incorrect conclusions on how control loops should be
paired and how much loop interaction exists. The possible situations are:

15
Literature Review
(1) In MIMO systems disturbances can have a profound effect on the transient
performance (Stanley et ah, 1985). However, the RGA does not provide
any insight on this aspect;
(2) In the case of one-way interaction, RGA does not apply even though the
effect could be severe;
(3) RGA does not consider dynamics, which may lead to incorrect loop
pairings.
2.3.2. RIA based loop paring

This is a RGA-like approach. The relative interaction for loop yl - uJ is
defined as the ratio of two elements: the increment of the process gain after all the
other n-1 control loops are closed and the apparent gain in the same loop when all
the other control loops are open (Bristol. 1967), i.e.,
\ si J 'all other loops close except for loop y - » , ^ St t > all loops open
= :
<P, ' ' •
1
i ^ , ^ , ) all loops open
This concept was further developed by Zhu and Jutan (1996), where loop
interaction can be exposed by investigating the response of an output to its input
in the presence of interaction from other loops through the hidden loops. This
situation can be best depicted by Figure 2.3. Focusing on an arbitrary control loop
with uj paired withy,, output in response to any change in the input consists of
two parallel paths: the direct forward path via the interaction free process and the
parallel path via the hidden loops due to off-diagonal elements of the process.

16
Literature Review
h.
w
av(s)
u
i h.
w SU(*) - ^ - _V!
W
(a) absolute interaction
& - *
(b) relative interaction

Figure 2.3. Definition of dynamic interaction in «,->', loop
Here, 0y(s) is the absolute interaction, which can be obtained by the signal flow
technique (Zhu and Jutan. 1996), the relative interaction can be then represented
as the ratio of the absolute interaction and the independent interaction free process
as
Q„Q)
Vv(s) =
At steady-state, assuming 1) the closed-loop system possesses integrity against

any single-loop failure and 2) each controller in each control loop contains
integral action, we obtain (Shinskey, 1988)
a (0) = — e (0).
,J }
G;/(0) ^
Hence, at steady-state we obtain the relative interaction
1
<P 1.
* U g y (0)G;'(0)
All the individual relative interaction terms constitute a relative interaction array
(RIA) as
Compared with the definition of the RGA, it is easy to give

17
Literature Review
f l =-L-i.
" \
Clearly, the RIA provides an equivalent interaction measure to the RGA. Loop
interaction and its implications can be briefly characterized by the RIA as follows:
(1) <p = 0 => no interaction;
(2) q>v > 0 => interaction acts in the same direction as interaction-free process;
a > 1.0 => interaction dominates over interaction-free process gain:
(3) <p < 0 => interaction acts in the reverse direction as interaction-free
process gain;
(4) <p, <-1.0=> reverse interaction dominates over interaction-free process
gain.
Under RIA based pairing criterion, variables should be paired in such a way so
that:
(1) All the RIA elements are closest to 0;
(2) NI is positive;
(3) All the RIA elements are greater than -1;
(4) RIA elements close to -1 are avoided.
It is important to notice that both very large positive and very large negative
RGA elements drive loop interaction toward the inverse direction of the original
interaction-free process gain and that both very small positive and very small
negative RGA elements are also, actually even more, harmful to process control.
Apparently, in contrast to the RIA, the RGA exhibits discontinuous behavior
when used to characterize loop interaction.
Although the above interaction rule seems to be equivalent to the RGA
interaction rule, the closeness to a value (0 in RIA rule, 1.0 in RGA rule) has
different implications. Specifically, the distance of the RGA elements from 1.0
may not realistically reflect the amount of interaction as noted previously, whereas
the distance of the RIA does. In particular, the RGA interaction rule mainly
targets at dismissing large RGA elements, which makes the robustness rule
redundant, whereas the RIA interaction rule essentially aims at eliminating small
RGA elements (large interaction), which complements the robustness rule.

Literature Review is
However, the RIA does not contain any more information on the process than
the RGA, although they provide valuable alternative viewpoints and smoother
interpretation. The common feature of those indices is that they only use the
system model at zero frequency.
2.3.3. Dynamic RGA (DRGA) based loop paring
DRGA is another approach to evaluate the interactions, which may overcome

some limitations from the RGA based approaches. The first paper to define and
propose an approach to solve DRGA was published by Witcher and Mc Avoy
(1977). They used a transfer function model in place of the traditional steady-state
model used for the RGA calculation. The denominator of the DRGA involves
achieving perfect control at all frequencies, while the numerator is simply the
open-loop transfer function. Other methods to calculate DRGA were also
proposed (Bristol, 1979; Tung and Edgar, 1981; Gagnepain and Seborg, 1982).
Since the DRGA is most valuable for screening alternate control system designs,
the requirement of an extensive controller design tends to defeat the utility of
these methods. A better approach to define a useful DRGA should involve
relatively little user interaction in the controller design aspect of the analysis.
Recently, Mc Avoy et al. (2003)_presented a new approach to define a DRGA.

It is assumed that a dynamic model of the process is available. If the model is a
transfer function model then it is converted to a state space model. The optimal
proportional output feedback controller is used to define the DRGA. The scaled
linear state space dynamic model is given by
\x = Ax + Bu
\y = Cx
The objective function to be minimized is defined as
J= ^[(/Qy+uTRu)dt.
The optimal control problem to be solved is a linear quadratic regulator with

output feedback (Levine and Athans. 1970). Here, Q and R are typically taken as

19
Literature Review
identity matrices, since u and y have already been scaled by their operating
ranges. Such a choice for Q and R treats all measurements and manipulated
variables equally. An output feedback matrix gives the admissible controls as
u =-Kx
where K is an optimal feedback controller.
By solving the expected value of J for the case where the initial state forcing
lies on a unit sphere around the origin, the DRGA which is based on the optimal
proportional controller Kcan be defined for the (i,j)'h element of the DRGA as
"" ty/ty L4=o.*» '

where both terms give the gain of ut to y during a transient in which the process
is controlled using the optimal output proportional gain matrix, K. Furthermore,

from u = -Kx, the partial derivatives can be calculated to give the following
result
—K
where K is equal to AT1. The following interpretation of XUj and paring rule is
proposed. If XDlJ is close to 1, then the optimal gain between «, and yy remains
the same regardless of whether the remaining manipulated variables are changing
aggressively or moving very little. If Aaj =1.0, then one can interpret this result
to mean that the manipulated variable under consideration does not interact with
the other manipulated variables so far as yf is concerned. Thus, pairing y. with
ut would be one candidate for SISO control. If a manipulated variable has a
negative value of Xn , then its behavior switches sign depending on how
aggressively the other manipulated variables are moving. Such a pairing should be
avoided since interaction is likely to be high. Large values of X^ also indicate
large changes in the behavior of a manipulated variable, which in turn indicate

that interaction is high.
Even though in all examples above, the new DRGA results in the best pairings
to use and it accurately assesses the extent of interaction present, its disadvantages
are also obvious:

2(1
Literature Review
(1) This pairing rule is still heuristic;

(2) Complex computing: a conversion from a transfer function model to a
state space model and a linear quadratic optimal control solution are
required;
(3) The value of ADlJ is sensitive to the weighting matrices. Q and R.
2.3.4. GRDG based loop paring
Another interesting dynamic interaction measure is the generalized relative

dynamic gain (GRDG) (Huang et al„ 1994). The GRDG also takes into account
the dynamics of the closed-loops. The element /4,,(s) of GRDG for a TITO
process is defined as:
g,,(5)g : : (5)——
4,W = J J ^ -
JL
g\\^)g^S) —-gn{s)glx{s)
v2(*)
where the variables r,(s)and y2(s) are the set-point and the process output of the
r (s)
other loop, and -= is the desired dynamics of the second loop. The transfer
y2(s)
functions gu(s) , g]2(s) , g2\(s) ar
*d fe(5) are me
elements of the process
transfer matrix G(s).
The loop paring rules are as below:
(1) Eliminate the infeasible pairings by normal RGA procedure;
(2) Calculate interaction potentials to obtain the most feasible pairings.
Pomerleau and Pomerleau (2001) gave a practical representation of the GRDG
as a function of both closed-loop bandwidths. For easier control and tuning, the
specifications on the closed-loop set-point responses have to be chosen in a
frequency band where interaction is reduced so the system behaves more like
SISO systems. In order to do so, the closed-loop response specifications are
chosen in frequency band where the GRDG is close to one since, as for RGA, it
means that the interaction is low.

21
Literature Review
Because the zeros in a transfer function affect the process dynamic, the
GRDG, which takes into account the dynamic part of the transfer function, is
preferred to the RGA, which only considers the steady-state. However, as can be
seen from ^,(5), the complexity of representation and computation of the RDGA
will dramatically increase if the system dimension increases.
2.3.5. SSV based loop paring
A shortcoming of many currently available interaction measures is their

limited theoretical basis. Using the notion of structural singular value (SSV),
Grosdidier and Morari (1986) presented a dynamic interaction measure for
multivariate systems under feedback with diagonal or block diagonal controllers.
Let G(s) be the process, and G{s) be a matrix consisting of the block
diagonal elements of G(s). A block diagonal controller
C(s) = diag(Cl(s),C2(s),...,Cm(s))
is to be designed for the system
G(s) = diag(Gu(s),G22(s),...,Gmm(s))
such that block diagonal closed-loop system with the transfer matrix
H(s) = G(s)C(s)(I + G(s)C(s)Y[
is stable. An interaction measure expresses the constraints imposed on the choice

of the closed-loop transfer matrix H(s) for the block diagonal system, which
guarantee that the full closed-loop system
H(s) = G(s)C(s)(I + G(s)C(s)yl

is stable.
Using relative error.

E(s) = (G(s)-G(s))G-\s),
which arises from the "approximation" of the full system G(s) by the block
diagonal system G(s). the general interaction measure can be stated as:
Assume that G(s) and G(s) have the same RHP poles and that H(s) is
stable. Then the closed-loop system H(s) is stable if
^mJH(jw))<M-](E(jw)) Vw, (2.6)
where a(A) is the spectral radius of A, fi is the SSV suggested by Doyle (1982).
Equation (2.6) implies that for stability, the magnitude of the diagonal blocks
H,(s) has to be constrained by the reciprocal of the SSV n of the relative error
matrix. The value of /j. depends on the structure assumed for H(s).
Although all popular interaction measures are either explicitly or implicitly

based on this definition and thus it has a rigorous theoretical basis, this definition
has its limitations and therefore the results should be interpreted with caution. The
reason is that interaction measure is based on the block diagonal H(s) which
might not be indicative of the actual full closed-loop transfer function matrix
H(s). Though the definition of the interaction measure guarantees H(s) to be
stable it can be very badly behaved. The interaction measure might indicate
"small" interactions but the performance could be arbitrarily poor. Furthermore,
the following disadvantages also need to be noticed as they are obstacles for it to
be understood and accepted by control engineers:
(1) When applied to loop paring, all pairing possibilities need to be screened
and calculated to find the best one;
(2) The computation of SSV /u needs complicated digital optimization;
(3) The SSV // gives equal preference to all the loops. In some cases this may
impose an early and perhaps unnecessary roll-off for some of the H^s).
The introduction of a weighting matrix with block diagonal structure equal

23
Literature Review
to that of H(s) can circumvent this problem. Nevertheless, this makes the
application more difficult.
2.4. Multi-loop controller design
A decentralized control system consists of a set of independent SISO

controllers. In process control industries, more than 95% of the control loops are
of P1D type and most controllers only use the proportional and integral actions (PI)
(Astrom and Hagglund, 1995). The popularity of these controllers comes from
their simplicity of implementation, their availability in distributed control systems
(DCS) and programmable logic controllers (PLC), their flexibility, and their easy
acceptance by operators. According to Skogestad and Morari (1987), they have
fewer tuning parameters, are easier to understand, and are more easily made fault
tolerant than decoupling controllers. In practice, decentralized control structure for
MIMO systems are also preferred for ease of start-up, soft automatic/manual
transfer, and fault tolerance in the event of actuator or sensor failures
(Lakshminarayanane/a/., 1997).
However, compared to SISO systems, efficient tuning of decentralized PID

controllers for multivariable processes is much more complex and many
controllers work without satisfactory performance. Many researchers, for
example, Vinante and Luyben (1972), Wood and Berry (1973) and Ogunnaike and
Ray (1979) have studied intensively on a large number of genuine decentralized
control systems, which are made up of SISO controllers acting in a multi-loop
fashion. Owing to inner interactions in a multivariable controlled system, a change
of controller parameters of one loop will affect the performance of others. The
design of SISO controllers for a MIMO process is more difficult because of these
interactions and has been an open research topic for years. In particular, the design
of SISO controllers for highly coupled multivariable processes often leads to poor
performance because of a bad choice of manipulated variables, poor specifications
and/or poor tuning of the controllers. The decentralized control scheme should not
suppress those interactions, only takes them into account in controller design.
Therefore, it is preferable to have a design technique which is simple, can be
frequently repeated, to give good, though not optimal, performance in all the

conditions. Many design methods have been reported in the literature, which may
be roughly divided into four categories:
(1) Detuning methods;

(2) Sequential loop closing methods;
(3) Independent methods;
(4) Simultaneous equation solving methods.
2.4.1. Detuning methods

In detuning methods, each controller of the single-loop control system
composed of diagonal element of transfer function matrix is tuned, ignoring the
process interactions from other loops. Then interactions are taken into account and
each controller is detuned using some interaction measures until some stability
criterion is met. If a MIMO process model is available, several detuning schemes
for decentralized PI/PID controllers are available based on frequency domain
analysis.
The following practical procedure is often recommended for 2*2 processes
(Mc Avoy, 1983):
(1) Use any single loop tuning rules to obtain starting values for the individual
controllers based on the corresponding main process, let the controller
gains be k°pi.
(2) These gains should be reduced using the following expressions that
depend on the relative gain parameter X:
{A-Â2-A)k°pi A>\.0
k
,,=
A + v A1 - A k°pj A<1.0
It may still be necessary to retune these controllers after they have been put in
operation, however, this will not require as much effort as if one were starting
from scratch. The disadvantage of the method is that this nicely analytic formula
is only suitable for 2*2 systems, no analytic solution for higher dimensional
systems.

Another practical work in this field is the biggest log modulus tuning (BLT)
method proposed by Luyben (1986). The system studied is supposed as an nxn
multivariate process given in the transfer function form
Y = G(s)U
where Y and U are vectors of the controlled variables and the manipulated
variables, respectively, and G is a matrix of the open-loop process transfer
functions gt) relating controlled variables yt to manipulated variables ur These
transfer functions are typically of the form
k ^ s + De-""'
g (s) = — .
The structure of the control system variable pairing has been determined such that
yt is controlled by «, using a PI controller ga(s). Thus, the feedback controller
matrix Gc(s) is diagonal since we are using n SISO controllers with
ga(s) = kpj(l+—).
The first step in the proposed procedure is to calculate the Ziegler-Nichols setting
for each individual loop. The ultimate gain ku,, and ultimate frequency cou, of
each diagonal transfer function g„(s) are calculated in the classical SISO way.
Next, a factor / is assumed. Typical values vary from 2 to 5. The gains of all
feedback controllers (k ) are calculated by dividing the Ziegler-Nichols gain
(kZN)by the factor /
Pl=
f
where
*», - 22 •
The reset times (r ; / ) of all controllers are calculated by multiplying the Ziegler-
Nichols reset times (r Z V j ) by the same factor /
x
,, = fTzs,
where

26
Literature Review
In
\.2o)UJ
The / factor can be considered a "detuning" factor which is applied to all loops.
The larger the value is, the more stable the system will be but the more sluggish
the set-point and load responses will be. Based on intuition and empirical grounds,
a multivariable closed-loop log modulus is defined as
where
W(s) = -\ + det(I + GB)(s).
The proposed tuning method is based on varying the factor / until the "biggest
log modulus" (Lcm)max, which is a measure of how far the system is from closed-
loop instability, is equal to some reasonable number. Thus, this tuning method is
called the biggest log modulus tuning. By a larger number of tests, the author
found that the following tuning criterion appears to give good response:
(4J max =2»-

Similar methods taking into accounts the interaction by detuning have also
been addressed by Chien et al. (1999, 2000) and Chen and Seborg (2003). In the
work of Chien et al, a straightforward tuning method is derived from a controller
synthesis method with a control performance specification of 5% overshoot of the
servo response. Depending on the interaction natures of the multi-loop systems,
controller tuning based on the diagonal elements of the model and further
detuning may be necessary. The RGA is employed to detune the parameters. For
systems with a relative gain array [/?G^(/l H )<l], a detuning factor based on this
information is proposed. The information needed for controller tuning purposes

includes the dynamic model parameters of the diagonal elements and the process
gain information of the off-diagonal elements.
The simplicity of this method is its major advantage. Nevertheless, the
disadvantage lies in the fact that loop performance and stability can not be clearly
defined through the detuning procedures. Generally, this class of method only
provides reasonable preliminary controller settings with guaranteed closed-loop
stability.

27
Literature Review
2.4.2. Sequential loop closing methods

In sequential loop closing methods, a controller is designed for a selected
input-output pair and this loop is closed. Then a second controller is designed for
a second pair while the first controller has been closed. The dynamic interaction of
current loop is considered in the closing of next loop, and so on. In this manner, a
series of controllers are designed sequentially. Usually the fastest loop is designed
first. Examples of such methods are those of Mayne (1973), Chiu and Arkun
(1992), and Hovd and Skogestad (1994). If the process model is not available, the
relay auto-tuning approach (Astrom and Hagglund, 1995; Yu, 1999) is a useful
and simple method to obtain system frequency information for PI controller
design. Because MIMO systems have an infinite number of ultimate points
(Halevi et al, 1997). the relay auto-tuning approach for SISO systems cannot be
directly applied. Loh et al. (1993), Loh and Vasnani (1994) proposed sequential
relay-feedback tests to locate the ultimate points of a MIMO system and to design
PI controllers. Palmor et al. (1995), Halevi et al. (1997) used similar simultaneous
relays in all of the control loops to obtain the ultimate point information. In both
the sequential and simultaneous relay auto-tuning approaches, after the frequency
response information has been obtained, the Z-N (Ziegler and Nichols, 1942) or
modified ZN tuning rules (Shen and Yu, 1994; Tan et ai, 1999; Hang et al, 1991)
can be used to tune the PI/PID controllers. The trial-and-error methods such as
iterative continuous cycling (ICC) method (Lee et al, 1998) determine the PI
(only) parameters sequentially by driving the system to have continuous cycling.
Additional constraints are imposed to compute the controller settings so as to
guarantee their nominal stability.
A typical autotuning approach proposed by Shen and Yu (1994) is

summarized here. Consider an n*n system G(s) with the entry gv(s) under
decentralized control with a known pairing as shown in Figure 2.2. Initially, a

relay is placed between y\ and M, while all other loops are on manual. The closed-
loop relationship between yy and H, becomes a bit more complicated than its open-
loop one (g,,):
£l.,Ci(5) =
V = gy. 1 -
gr.Sl&g,
V "l Ja

where the subscript CL stands for closed-loop, k2 is the sub-matrix of the rest of
the system matrix G(s) , and h2 = g22k2[l + guk2] is the complementary
sensitivity function for the remaining loops. Following the relay-feedback test, a
controller can be designed from the ultimate gain and ultimate frequency for loop
1. The next step is to perform relay-feedback test between y^ and u2 while loop 1
is on automatic and other loops are still on manual. A controller can also be
designed for loop 2 following the relay-feedback test. Once the controller on the
loop 2 is put on automatic, a relay-feedback experiment is performed between y3
and M3 . In this way, all loops will be tested and closed. Then, another relay-
feedback experiment is performed between yt and M, again. Generally, a new set
of tuning constants is found for the controller in loop 1. This procedure is repeated
until the controller parameters converge. Typically, the controller parameters
converge in 3 ~4 relay-feedback tests for 2 ><2 systems. At each stage, a modified
Ziegler-Nichols formula, with a detuning factor / = 2.5, is employed to tune the
controller parameters. It is worth noticing that the proposed MIMO autotuning
concept repeats the "identification-design" procedure on equivalent SISO transfer
functions.
This class of approaches has several advantages. First, it makes the problem
simple by treating the MIMO system as a sequence of SISO systems which the
relay-feedback system is proven useful and reliable. Second, it operates in an
efficient manner. Third, it can be applied for nonlinear multivariable systems.
However, the method has some fundamental disadvantages: 1) it cannot guarantee
convergence even if starting in a neighborhood of the true solution; 2) loop failure
tolerance is not automatically guaranteed when loops designed early fail; 3) the
performance of the method depends strongly on which loop is designed first and
how the first controller is designed; and 4) the iteration procedure is essential
because closing the subsequent loops may alter the response of the previously
designed loops. Hence, conservative design may result due to the RHP zero on the
diagonal which may not be the RHP transmission zeros of the MIMO process. In
some scheme like the work of Shen and Yu (1994), after all loops have been
closed, the controller will be re-tuned one after another with all other loops closed
with the controllers obtained in the previous step. This procedure will go on until

2')
Literature Review
they converge. For higher dimensional systems, the number of relay-feedback

tests will be very large for the solution to converge. Finally, the tie between the
tuning procedure and the loop performance is weak.
2.4.3. Independent design methods

In independent design methods, the design procedure is to find each controller
based on the paired transfer function while satisfying some constrains due to
process interactions (Grosdidier and Morari, 1986; Lee et ai, 2001; Skogestad and
Morari, 1989; Hovd and Skogestad, 1993). In this approach the issue of
interaction between loops is examined first, and sufficient conditions are derived
for the individual loops that guarantee stability and robust performance.
Subsequently, the SISO controllers are designed such that each loop satisfies these
sufficient conditions. Most of design methods are based on pre-specifying the
form of closed-loop transfer function. While this kind of approach is systematic
and with stability guaranteed, a disadvantage is that this method is potentially
conservative since the information on other controllers is not exploited in the
design of a particular controller (Hovd and Skogestad, 1994). Some recent
progress is summarized as below.
Desbiens et al. (1996) presented a practical frequency based tuning approach
for TITO systems. For a 2*2 process G(s) with a decentralized controller
Gc(s) as shown in Figure 2.2. first select the desired tracking closed-loop
dynamics, which can be minimal-phase or nonminimal-phase, for each loop as
yi(s) = \-rms
r,(s) (l + rn5)(l + r 2 r 0
and
y2(s) _ l-r o : 5
r2(s) (1 + r,,s)(l + rns)
Then the open-loop characteristics are deduced from tracking the closed-loop
specification as
Gou W = gci (s)gi, (5) = ^ ^ (2.7)

5(r 11 +r :i +r 01 +r 11 r 2 ,5)
and

]
Gau(') = gA')g*W = ~t —^ :• (2-8)
s(r,, + r,, + r02 + r l2 r 2l s)
By solving Equations (2.7) and (2.8). the magnitude and phase of the SISO
controllers leading to the desired performances are obtained for any frequency:
gîjo)) and gc2(jaj). The last step is to define the transfer functions gcl(s) and
gr,(.s)that best approximate the frequency responses gcl(j(o) and gc2(jco). The
following controller structure is used:
K((\ + T,s)(\ + Tdls + Td2s2)
T,s(\ + Tfls + Tf2s )
An optimization method is then used to get an approximation of each

corresponding transfer function of the controller while a constraint based on
singular value stability robustness theorem is implemented to ensure a stable
closed-loop system.
The settings obtained by this approach can lead to performances
corresponding closely to the selected tracking closed-loop specification. However,
the resulting controllers are not of standard PID structure and an optimization
procedure is included, so it is difficult to implement and accepted by control
engineers. In addition, it is not straight forward to extend the approach to 3*3 and
higher dimension systems.
Lee and Edgar (2000) proposed a phase stability condition imposed on the
individual loop transfer function to reduce conservatism of the independent design
method with the fj. -interaction measure alone. The fi -interaction measure is
based on the small gain theorem. Although the // -interaction measure is the
tightest norm bound of the individual closed-loop transfer function, it is
conservative because the phase information is not utilized. However, by
combining various linear fractional transformations and the small gain theorem, a
phase stability condition can be obtained. This condition can be used to design
multi-loop control systems to reduce the conservatism of the fi -interaction
measure considerably. This approach is straightforward and may result in stable
multi-loop control systems that cannot be obtained by the p -interaction measure
alone. Nevertheless, some relatively complicated plots are needed during design
procedure.

31
Literature Review
Chen and Soberg (2002) proposed a forward multi-loop controller design

method based on Gershgorin bands. A new definition of the ultimate point is
introduced for diagonal dominant MIMO systems. Analytical formulas are
developed for new ultimate gains and ultimate frequencies based on the system
frequency response and Gershgorin bands, a decentralized PI/PID is then designed
by a modified version of the Z-N relations. This approach is systematic with
classical theoretical foundation. But it cannot be directly applied to processes
without diagonal dominance and the design procedure is rather complicated for
practical engineers.
2.4.4. Simultaneous equation solving methods

In Simultaneous equation solving methods, the equations related to controller
parameters are created according to certain performance criterion and solved by
some optimization technique. Generally, design of multi-loop controller by way of
simultaneous equation solving is numerically difficult. Some important works are
summarized below.
Wang et al. (1998) presented a design method for multi-loop PI/PID
controllers. Consider a stable 2*2 process as shown in Figure 2.2,
yt(s) gU(S) g\l(S) «,(5)
M'\>J L&iC?) gn(s). u2(s)

Assume that proper input-output pairing has been made to the process. The
process is to be controlled in a negative feedback configuration by the multi-loop
controller:
gjs) 0
Gc(s) =
0 gc2(*)
The controller design objective is to find Gt(.s)such that both loops achieve
satisfactory performance. The equivalent transfer functions g,(s) between input
H,(s)and outputyt(s), and g,(s)between input M,(S) and output y2(s) can be
obtained as
g\2(S)gll(S)
gi(s) = gu(s)-
g~c2](S) + g22(S)
and

gc>(s) + gu(s)
respectively. It is thus clear that in order to achieve a good loop performance,
gc](s) and gc2(s) should be designed for g,(s)and g,(s) instead of the plant
diagonal element gn(s) andg„,(s), as is done in many other methods. In this

work, the modified Ziegler-Nichols is applied to the equivalent transfer function
g,(s) and g2(s), i.e., the controller ga(s),i = \,2 is designed such that a given
point on the Nyquist curve of g,(s),/' = l,2 is moved to a desired point. In this
way, the optimum settings for the multi-loop controller can be obtained with all
the benefits derived from the well-developed SISO tuning. Let
4=gl{m) = raei(-"'-)
be the two given points on the Nyquist curves g,(s),i = 1,2. They are to be moved
respectively to the desired points
Bi=gl(M)gAJ<»i) = rkie'{-*+'>).
Let g c; (s),/ = l,2be of PID type, i.e.,
1 +
Sc,(-0 = M + rd,s),i = 1,2-
Applying the modified Ziegler-Nichols method to each ga(s) yields
P,s a,
where
tan(0k -<PJ + -j^ + t a " 2 lV* ' <?<••) a > 0
P=
,«. =0
tan(^,-^)
It follows from the phase part of the representation of B, that
g a ( » = ^ , 0 + y ' t a n ( ^ - ^ ) ) , / = l,2.
Then, the representation of B, can be written as
gXM)kpj=cos{9h,-(pa,)rhle^'^. (2.9)

33
Literature Review
Bringing g:(s),i = 1,2 and gcl(s),i = 1,2 into Equation (2.9) gives two
equations which are complex and can be broken down to four real equations.
Because the number of unknowns equals the number of real equations, intuitively,
it is expected that the problem can be easily solved. Unfortunately this is not the
case. The main difficulty lies in the nonlinearities of the equations which are
further complicated by the coupling among the equations. Even though the authors
presented a novel graphic iterative procedure to solve the these equations with
Nyquist stability constraints, there is no guarantee for the existence of solutions.
In addition, the graph based iterative optimization technique may be complicated
for control engineers and the extension to 3><3 or higher dimensional systems is
far from straight forward if not impossible at all.
Another work of Bao et al. (1999) formulated the multi-loop design as a
nonlinear optimization problem with matrix inequality constraints. As has been
illustrated, the formulation does not include the systems that have different input
delays, which unfortunately happens to be very common in MIMO process control.
Simultaneous optimization for solving multi-loop controllers is also numerically
difficult. The result is very much dependent on the conditions defined in the
objective function. The controllers may result in instability, in case of loop failure
or where loops are closed in different orders. Zhuang and Atherton (1994)
investigated a PID tuning approach using an integral performance optimization
procedure for a TITO system. First of all, represent the error matrix as
combination of the process model, reference inputs and controller parameters.
Then some integral optimum indices related to the error matrix were introduced.
The optimal PID settings of decentralized controller are therefore obtained by
minimizing an integral performance criterion with some constrains. Even if only
for TITO systems, the calculation is relatively complicated and the existence of
the optimal solution is not guaranteed. In addition, Vlachos et al. (1997) studied a
genetic algorithm (GA) approach for optimal PI controller design. After
evaluating all n objective function elements for the j * set-point pattern, the
results can be weighted and added together to form a single number that
represents the quality of the controllers for the j * set-point pattern. Finally, when
all n set-point patterns have been applied to the closed-loop system, the maximum
of all resulting objective function elements (indicating the worst performance) can

be selected as the final objective function value for the genetic algorithm. The
crossover operator improves the accuracy of the GA according to the authors. The
main advantage of the proposed method is the ability to handle arbitrary
performance objectives that can be different for each of the outputs of the system.
A disadvantage of the proposed method is that n simulations are required for the
evaluation of the objective function for a single set of controller parameters. This
significantly reduces the speed of the genetic algorithm. In addition, only a TITO
simulation example is given.
2.5. Decoupling controller design
For processes with significant interactions, the performance by

decentralized control may be too poor to meet specifications, or the processes
may not even be stabilized due to degraded performances caused by process
interactions. To improve the closed-loop performances of multi-loop control
systems, decouplers are often used such that the interaction effects are
eliminated, or at least significantly reduced.
Generally, decoupling is achieved by introducing an additional transfer

function block or interaction compensator between the single loop controllers, and
the process. The decoupler causes the control loops to act as if totally independent
of each another, thus reducing the controller tuning task to that of tuning several
non-interacting controllers. The design of ideal decouplers can be either
simplified decoupling or generalized decoupling.
For simplified decoupling, the structure of decoupling matrix is in the form of
1
gmW ••• £/i»
G,(s) =

35
Literature Review
the decoupler design is to find g,,j(s\ '*j such that the cross coupling is
eliminated and G(s) becomes diagonal.
For general decoupling, a diagonal matrix, GR(s), is specified, usually with the
form of
GR(s) = Diag[G(s)].
Then the decoupler Gt (s) is determined by
G(S)GI(S) = GR{S)_
Each method has advantages and disadvantages. For simplified decoupling,

the "equivalent" open-loop decoupled system is much more complicated than the
GR(s) matrix specified in the generalized decoupling design. Therefore, the tuning
and closed-loop performance for the generalized decoupling is much better than
for simplified decoupling. However, the price for the improved closed-loop
performance in the generalized decoupling is complicated decoupler. Because of
complexities in implementation or the lack of a high-quality dynamic model, it is
often necessary to make use of simplifications of the ideal decoupler. In order to
make the best balance between complexity and closed-loop controller
performance, steady state decoupling and partial/block decoupling are often
employed in practice.
Steady-state decoupling requires only the matrix of steady-state gains as a

model. If each transfer function element g,j(s), has a steady-state gain term kt] and
the matrix of steady-state gains is represented by K, then the simplified steady-
state decoupler for an n X/J system is

1 kll2 ... kIln
G,(s) = . . .
_kM ... ktm_l 1 _
which will take simple, constant numerical values, and will always be realizable,
as well as implementable.
In generalized steady-state decoupling, the decoupler matrix is given by
G (s) = K'K
where KR is the steady-state version of GR(s). Since the inversion concerns only a
matrix of numbers, it will always be realizable and easily implemented.
The main advantages of steady-state decoupling are that its design involves
simple numerical computations, and the resulting decouplers are always realizable.
As steady-state decouplers are very easy to design and implement, it is normally
recommended that steady-state decoupling is the first technique to try in industry.
Only if the dynamic interactions prove to be persistent will dynamic
considerations be entertained. Especially, if the dynamic aspects of the transfer
function elements in each row of the transfer function matrix are similar, the
dynamic decoupler will be very close to its steady-state version.
Partial/block decoupling reduces the dimensionality of the problem by

requiring decoupling in only a few of the control loops. When some of the loop
interactions are weak or if some of the loops need not have high performance, one
may consider partial decoupling. In this case, attention is focused only on a subset
of the control loops where the interactions are important and high performance
control is required. It is typical to consider partial decoupling for 3 x 5 or higher
systems, as the main advantage is reduction of dimensionality.
In summary, designing a decoupler is to select elements of the compensator to

satisfy one of the following objectives:

37
Literature Review
(1) Ideal/dynamic decoupling: eliminate interactions from all loops, at every

instant of time. This is an ideal decoupling objective. However, perfect
decoupling is possible only if the process model is perfect which is
impossible in practice. In addition, ideal dynamic decouplers may have
realization problems, particularly when time delays are involved in the
transfer function elements.
(2) Steady-state decoupling: eliminate only steady-state interactions from all
loops, and dynamic interactions are tolerated. The difference between
dynamic decoupling and steady-state decoupling is that the former uses the
complete, dynamic transfer function elements in the decoupler, and the
latter uses only the steady-state gain portion of these transfer function
elements. If the dynamic aspect of the transfer function elements in each
row of the transfer function matrix are similar, the dynamic decoupler will
be very close to its steady-state version.
(3) Partial/block decoupling: this objective focuses attention only on the
critical loops in which the interactions are strongest, leaving those with
weak interactions intact. When some of the loop interactions are weak or
some of the loops do not need high performance, one may consider partial
decoupling. It is typical to consider partial decoupling for 3x3 or higher
systems, as the main advantage is to reduce dimensions.
Some existing decoupling design approaches are summarized as below.
2.5.1. Ideal/dynamic decoupling methods

Ideal/dynamic decouplers compensate for process interactions and result in
decoupled responses of controlled variables (Luyben, 1970; Waller, 1974). The
solution to this problem has been studied by several researchers and various
approaches, configurations and techniques have been used to achieve decoupling
(see, for example, Wonham, 1979; Hautus and Heymann, 1983; Hammer and
Khargonekar, 1984; Desoer and Gundes, 1986; Vardulakis, 1987; Wang, 1992).
Especially, decouplers have been well established in the field of distillation
control and almost all textbooks for process control deal with them (Shinskey,
1988; Seborg and Edgar, 2004).

Among all the dynamic decoupling methods, Wang et al. (2003) proposed a
full dimensional decoupling control design for an n*n process described, as shown
in Figure 2.2, via a square and non-singular transfer matrix G(s) as
*(*) g>.(5) giAs) ",(•*)
-0 v
where gtJ(s) = g,j0(s)e '' , and glj0(s) are strictly proper, stable scalar rational
functions and 6IJ are non-negative constants. The process is to be controlled in a
conventional negative feedback configuration by a full cross coupled

ljS
multivariable controller G (s) = \gc As)] , where g<. As) = g 0(s)e , d are
non-negative constants, and gclj0(s) are scalar proper rational functions. The task
here is to find a Gc(s) such that the closed-loop transfer matrix between the
output vector v and the reference vector r:

H(s) = G(s)GAs)[l + G(s)Gc(s)}\
is decoupled, that is, H(s) is diagonal and nonsingular. By analyzing the
properties of the decoupler and the achievable performance of the decoupled
system, the authors gave the following decoupling design procedure:
(1) Based on the process characteristics, derive hn, the objective closed-loop
transfer functions from the i* set-point rt to the /** output yt for each
decoupled closed-loop. The i* objective open-loop transfer function qrn is
then given by
<ln,=
\-K
(2) Match the actual loop gnku to the objective loop: g„(.s)gCi„(s) = 9 m (5), to
find the ideal diagonal elements of the controller as
gT'(s) = g-\s)qrU(s) =«"(*> ^;qrtl(s),
where G" is the cofactor corresponding to g in G , and the
corresponding ideal decoupling off-diagonal elements gt Jt(s) of the
controller as

(3) Approximate the ideal controller g'fîs) wisely with a low-order
transfer function plus possible delay so that a set of specified performance

requirement is satisfied.
Even though a systematic design procedure for MIMO system design with
good performance is provided, the overall design is complicated for control
engineers as the model reduction technique is used in both desired performance
and controller design. Furthermore, although the dynamic decoupling can work
well with exact process model, it is not robust to modeling errors. This has been
confirmed by Arkun et al. (1984) using a robustness analysis procedure based on
the singular value to distillation control systems with decouplers. In addition,
dynamic decouplers also require detailed process models.
2.5.2. Static decoupling methods

To avoid detailed process models, static decouplers only use process steady-
state gains. Astrom et al. (2002) showed that static decouplers are very effective
for some processes when very fast controls are not required. Astrom et al. (2002)
described a simple multivariate controller consisting of a static decoupler and
two single loop PI controllers. A method for finding the parameters of the
controller is developed in the paper. For static decoupled transfer function matrix,
the authors introduced two quantitative interaction indices kx and k2 for very
small 5 (very low frequencies). Possible detuning of the PID controllers is then
quantified in terms of these indices. The interaction indices kt and k2 depend on
both the process and the controller and take dynamics into account. The indices
introduced in the paper capture the fact that interaction can be reduced by
detuning the loops. Nevertheless, the proposed interaction indices and controller
design depend on the assumption that the interaction is not too severe. In addition,
the approach is only developed for TITO systems.
Recently, Lee et al. (2005) showed that static decouplers should be applied
only to the integral modes for such processes since static decouplers can cause
undesirable effects on high frequency responses for some processes. Frequency

411
Literature Review
dependent relative gain arrays (Bristol, 1966) are used to determine whether static
decouplers applied only on the integral modes are recommended. The main results
are listed here. Consider a process of n inputs and n outputs, as shown in Figure
2.2, whose transfer function is
G(s) = {gIJ(s),i = \,2,...,n,j = \,2,...,n}.
To improve control performances of multi-loop controllers Gc(s) = lK/s + K },
where K, and K are diagonal, static decouplers designed from steady-state
process gains can be used as follows
' •
v.'
GC(5) = G'(0) -K+Kp = -G\0) + G-'(0)
s
0 p."
The static decoupler G '(0) reduces process interactions at frequencies near zero.
However, for some processes, it can introduce undesirable effects at high
frequencies. Typically, diagonal elements of the process transfer function are
faster than off-diagonal elements. As frequency increases, the magnitudes of off-
diagonal elements decrease faster than those of diagonal elements and,
consequently, interactions for such processes become smaller. However, when
static decouplers are applied to such processes, off-diagonal elements of the
decoupled process G(s)G'\0) contain faster elements and their magnitudes
relative to those of diagonal elements will not decrease, resulting in higher process
interactions at high frequencies. There are many realistic process models, such as
Wood-Berry column model, for which the static decouplers can have negative
effects on process interactions at high frequencies. To avoid the disadvantage that
a static decoupler causes undesirable effects on reducing process interactions at
high frequencies, the authors proposed a control system where the static decoupler
is applied for only the integral mode, i.e..
0 V I
0
]
Gc(s) = -G (0)K,+K =-G-'(0)
Qualitatively, the integral mode G'](0)Kjs is dominant at low frequencies and

its magnitude decreases as the frequency increases. Thus the static decoupler is

41
Literature Review
effective only at low frequencies and its undesirable behavior at high frequencies
is avoided.
A static decoupler implemented on the integral mode can improve the closed-
loop performances at low frequencies with good robustness. Generally speaking,
although static decouplers are simple to design and adjust in the field, they do not
always provide better closed-loop performances.
2.5.3. Partial/block decoupling methods

Since diagonal decoupling generally comes at a cost, there has also been
interest in related strategies which achieve some form of partial/block decoupling
(Hautus and Heymann 1983, Williams and Antsaklis 1986, Linneman and Wang
1993). The other reason why block decoupling is popular is that in many
applications decoupling is practiced only in the direction of the most important
loop to avoid the increased sensitivity of two-way decoupling to process
uncertainty (Shinskey, 1996).
Recently, Gilbert et al. (2003) presented a practical one-way decoupling
problem using finite frequency response data. For a 2*2 system as shown in
Figure 2.2, the overall controller structure in matrix notation is given by
k.
+ 1 *,,+ 0
1 k, W
DUo>,)Gc(ja),). m
0 1 o
m
where D(jco) is a decoupling matrix with a lead/lag compensator in the 1. 2
position, and Gc{jco) is a diagonal PI controller. First, a frequency response

estimate is obtained from two closed-loop tests using an identification technique
based on bandpass filters. The open-loop frequency response is then calculated
and used to tune PI controllers and lead-lag compensators for decoupling. For this
purpose, a desired closed-loop transfer function is specified as
a>;,e 0
r+2C,<y„,5 + < 1
U*) = (10)
0 <eT"
2
s + ^zeomZs + iol2
The authors suggested that (on] = conZ and £, = £, = 1.0 . The optional delay

4:
Literature Review
parameters Tdl and rrf,can be specified as zero since the advantage of working
directly with frequency response data, as opposed to transfer functions, is that

delay estimation is not important because its effect is included in the frequency
response data. Then, perform tuning by making the real closed-loop performance
Tt approach the desired frequency response Td defined by Equation (10) at each
of the harmonic frequency using the following index:
^(K'-r-'fKf)
7 =1
where the indices k and / refer to the outputs and inputs, respectively, and i refers
to the i'h harmonic frequency of the Ith set-point excitation, Wkl are weights
applied to the k" output and /' input. A' are the weights applied to each frequency,
which should be chosen to correspond to the harmonic amplitudes of a trapezoidal
waveform of initial slop K and period P:
4K . i 2n
A,=-—sin ,i = \.\5...det(l + GDGc).
K I K PJ
While the optimization is mainly carried out for user-specified performance
specifications, two constraints are proposed to check. They are the stability checks
for violations of the multivariable Nyquist stability margins on det[l + GDGC),
and the robustness maximization by maximizing the minimum related singular
values ofdet(l+{GDGc)~x\, and det{l + (DGcG)~'y
Using this approach, it is not necessary to approximate the process with a

FOPDT or SOPDT model, and hence avoiding the need to treat time delays since
the time delay is indirectly represented in frequency response data. However, the
developed design is not so easy to apply to higher dimensional systems since the
reliability of the frequency estimates may become a limitation. In addition, the
existence of the acceptable solutions may not be guaranteed.
Usually, the block decoupling is easier to realize for those processes which
can be grouped into some sub-systems with weak interaction among them. How to
make the groups is, however, still empirical.

References
Arkun, Y.; Manousouthakis, B; Palazoglu, A. Robustness analysis of process

control systems: A case study of decoupling control in distillation. Ind Eng Chent
Process Des Dev. 1984, 23, 93-101.
Astrom, K. J.; Hagglund, T. PID Controllers: Theory, Design, and Tuning (2nd
edition); Instrument Society of America: Research Triangle Park, NC, 1995.
Astrom, K. J.; Johansson, K. H.; Wang, Q.-G. Design of decoupled PI controllers
for two-by-two systems. IEE Proc.-Control Theoty Appl. 2002, 149.
Bao, J.; Forbes. J. F.; McLellan, P. J. Robust multiloop PID controller design: a
successive semidefinite programming approach. Ind. Eng. Chem. Res. 1999, 38,
3407-3419.
Bristol, E. H. On a new measure of interactions for multivariate process control.
IEEE Trans. Autom. Control 1966, 11, 133-134.
Bristol, E. H. On a philosophy of interaction in a multiloop world. ISA Chemical

& Petroleum Division Chempid Symposium, St. Louis, Missouri, USA, May, 1967.
Bristol, E. H. Recent results on interactions in multivariable process control. In
Proceedings of the 71s' Annual AIChE Meeting, Houston, TX, USA; AIChE, 1979.
Campo, P. J.; Morari, M. Achievable closed-loop properties of systems under

decentralized control: conditions involving the steady-state gain. IEEE Trans. AC
1994, 39,932-943.
Chen. D.; Seborg, D. E. PI/PID Controller Design Based on Direct Synthesis and
Disturbance Rejection. Ind. Eng. Chem. Res. 2002, 41, 4807-4822.
Chen, D.; Seborg, D. E. Design of decentralized PI control systems based on
Nyquist stability analysis. J. Proc. Cont. 2003, 13, 27-39.
Chien, I.-L.; Huang, H-P.; Yang, J.-C. A simple multiloop tuning for PID
controllers with no proportional kick. Ind. Eng. Chem. Res. 1999. 38, 1456-1468.
Chien, I.-L.; Huang, H.-P.; Yang, J.-C. A simple TITO method suitable for
industrial applications. Chem. Eng. Comm. 2000, 182, 181-196.
Chiu, M. S.; Arkun, Y. Decentralized control structure selection based on integrity
considerations. Ind. Eng. Chem. Res. 1990, 29, 369-373.
Chiu, M. S.; Arkun, Y. A methodology for sequential design of robust

44
Literature Review
decentralized control systems. Automatica 1992, 28, 997-1001.

Cutler, C. R.; Ramaker, B. L. Dynamic matrix control—a computer control
algorithm. In Proceedings of the 71s' Annual AIChE Meeting, Houston, TX, USA;
AIChE, 1979.
Desbiens, A.; Pomerleau, A.; Hodouin, D. Frequency based tuning of SISO
controllers for two-by-two processes. IEE Proc.-D. 1996, 143, 49 - 56.
Desoer, C. A.; Gundes, A. N. Decoupling linear multiinput multioutput plants by
dynamic output-feedback-an algebraic-theory. IEEE Trans. AC 1986, 31, 744-750
Gagnepain, J. P.; Seborg, D. E. Analysis of process interactions with application
to multiloop control system design. Ind Eng Chem Process Des Dev. 1982, 21,5-
11.
Gilbert, A. F.; Yousef, A.; Natarajan, K.; Deighton, S. Tuning of PI controllers

with one-way decoupling in 2x2 MIMO systems based on finite frequency
response data. J. Proc. Cont. 2003, 13, 553-567.
Grosdidier, P.; Morari, M. Closed-loop properties from steady-state gain

information. Ind. Eng. Chem. Fund. 1985, 24, 221-235.
Grosdidier, P.; Morari, M. Interaction measures under decentralized control.

Automatica, 1986, 22. 309-319.
Halevi, Y.; Palmor, Z. J. Efrati, T. Automatic tuning of decentralized PID
controllers for MIMO processes. /. Proc. Cont. 1997, 7. 119-128.
Hang, C. C ; Astrom, K. J.; Ho, W. K. Refinements of the Ziegler-Nichols tuning
formula. IEE Proc.-D. 1991, 138, 111-117.
Hammer, J.; Khargonekar, P. P. Decoupling of linear-systems by dynamic output-
feedback. Math. Systems Theory 1984, 17, 135-157
Hautus, M. L. J.; Heymann, M. Linear feedback decoupling-transfer-function
analysis. IEEE Trans. AC 1983, 28, 823-832
Hovd, M.; Skogestad, S. Improved independent design of robust decentralized
controllers. J. Proc. Cont. 1993, 3, 43-51.
Hovd, M.; Skogestad, S. Sequential design of decentralized controllers.
Automatica 1994, 30, 1601-1607.
Huang, H.-P.; Ohshima, M.; Hashimoto, I. Dynamic interaction and multiloop
control system design. J. Proc. Cont. 1994, 4, 15-27.

45
Literature Review
Kariwala, V.; Hovd, M. Relative Gain Array: common misconceptions and

clarifications. In Proceedings of 71 Symposium on Computer Process Control,
Lake Louise, Canada; 2006.
Lakshminarayanan, S.; Shah, S. L.; Nandakumar, K. Modeling and control of
multivariable processes: dynamic PLS approach. AIChE Journal 1997, 43, 2307-
2322.
Lee, J.; Cho, W.; Edgar, T. F. Multiloop PI controller tuning for interacting
multivariable processes. Comp. Chem. Eng. 1998, 22, 1711-1723.
Lee, J.; Edgar, T. F. Phase conditions for stability of multi-loop control systems.
Comp. Chem. Eng. 2000, 23. 1623-1630.
Lee, J.; Kim, D. H.; Edgar, T. F. Static decouplers for control of multivariable
processes. AIChE J. 2005, 51, 2712-2720.
Lee, T. K..; Shen, J.; Chiu, M. S. Independent design of robust partially
decentralized controllers. J. Proc. Cont. 2001, 11, 419-428.
Levine, W.; Athans, M. On the determination of the optimal constant feedback
gains for linear multivariable systems. IEEE Trans. AC 1970, 15, 44^18
Linneman, A.; Wang, Q.-G. Block-decoupling with stability by unity output-
feedback- solution and performance limitations. Automatica 1993, 29, 735-744.
Loh. A. P.; Hang, C.C.; Quek, C. X.; Vasnani. V. U. Autotuning of multiloop
proportional-integral controllers using relay feedback. Ind. Eng. Chem. Res. 1993,
32,1102-1107.
Loh. A. P.: Vasnani, V. U. Describing function matrix for multivariable systems
and its use in multiloop PI design. J. Proc. Cont. 1994, 4, 115-120.
Luyben, W. L. Simple method for tuning SISO controllers in multivariable
systems. Ind. Eng. Chem. Proc. Des. Dev. 1986, 25, 654-660.
Luyben, W. L. Distillation decoupling. AIChE J. 1970, 16, 198-203.
Mayne, D.Q. The design of linear multivariable systems. Automatica 1973, 9,

201-207.
Mc Avoy, T. Interaction Analysis: Principles and Applications; Inst. Soc. of
America: North Carolina, 1983.
Mc Avoy, T.; Arkun, Y.; Chen. R.: Robinson, D.: Schnelle, P. D. A new approach
to defining a dynamic relative gain. Control Eng. Practice 2003, 11, 907-914.
Morari, M.; Baric, M. Recent developments in the control of constrained hybrid

46
Literature Review
systems. Comp. Chem. Eng. 2006. 30. 1619-1631.

Morari, M.; Zafiriou, E. Robust Process Control; Prentice-Hall: Englewood Cliffs,
NJ, 1989.
Niederlinski, A. A heuristic approach to the design of linear multivariable
interacting subsystems. Automatica 1971, 7, 691-701.
Ogunnaike, B. A.; Ray, W. H. Multivariable controller design for linear systems

having multiple time delays. AIChEJ. 1979, 25, 1043.
Palmor, Z. J.; Halevi, Y.; Efrati, T. A general and exact method for determining
limit cycles in decentralized relay systems. Automatica 1995, 31, 1333-1339.
Pomerleau, D.; Pomerleau, A. Guide lines for the tuning and evaluation of
decentralized and decoupling controllers for processes with recirculation. ISA
Trans. 2001,40, 341-351.
Qin, S. J.; Badgwellb, T. A. A survey of industrial model predictive control
technology. Control Eng. Practice 2003, 11, 733-764.
Richalet, J.; Rault, A.; Testud, J. L.; Papon. J. Model predictive heuristic control:
Applications to industrial processes. Automatica 1978, 14, 413-428.
Seborg. D. E.; Edgar. T. F.; Mellichamp. D. A. Process Dynamics and Control.
John Wiley and Sons: New York. 1989.
Seborg, D. E.; Edgar, T. F.; Mellichamp, D. A. Process Dynamics and Control,

(2nd edition). John Wiley and Sons: New York, 2004.
Shen, S. H.; Yu, C. C. Use of relay-feedback test for automatic tuning of

multivariable systems. AIChE Journal 1994, 40, 627-646.
Shinskey, F. G. Process Control Systems; McGraw-Hill: New York, 1988.
Shinskey, F. G. Process control systems: Application, design and tuning;

McGraw-Hill: New York, 1996.
Skogestad, S.; Morari, M. The dominant time constant for distillation columns.
Comp. Chem. Eng. 1987, 11, 607-617.
Skogestad, S.; Lundstrom, P. Mu-optimal LF-control of distillation columns.

Comp. Chem. Eng. 1990, 14. 401-413.
Skogestad, S.; Morari, M. Robust performance of decentralized control systems

by independent designs. Automatica 1989. 25, 119-125.

47
Literature Review
Stanley, G.; Marinogalarraga M.; Mc Avoy, T. J. Shortcut operability analysis .1.

the relative disturbance gain. Ind. Eng. Chem. Des. Dev. 1985, 24, 1181-1188.
Tan, K. K.; Wang, Q.-G.; Hang, C. C. Advances in PID Control; Springer-Verlag:

New York, 1999.
Tung, L.; Edgar, T. Analysis of control-output interactions in dynamic systems.
AIChEJ. 1981,27,690-693.
Vardulakis, A. I. G. Interal stabilization and decoupling in linear-multivariable

systems by unity output-feedback compensation. IEEE Trans. AC 1987, 32, 735-
739
Vinante, C. D.; Luyben, W. L. Kern. Teollisuus 1972, 29, 499.
Vlachos, C.; Evans, J. T.; Williams, D. PI controller tuning for multivariate
process using genetic algorithms. In Genetic Algorithms in Engineering Systems:
Innovations and Applications; 1997; 446, 43-49.
Waller, K. V. T. Decoupling in distillation. AIChEJ. 1974, 20, 592-594.
Wang, Q.-G. Decoupling with internal stability for unity output feedback systems.
Automatica 1992, 28, 411-415.
Wang, Q.-G.; Lee, T. H.; Zhang. Y. Multi-loop version of the modified Ziegler-
Nichols method for two input two output process. Ind. Eng. Chem. Res. 1998, 37,
4725-4733.
Wang, Q.-G.; Zhang, Y: Chiu, M. S. Non-interacting control design for
multivariable industrial processes. J. Proc. Cont. 2003, 13, 253-265.
Williams, T. W. C; Antsaklis, P. J. A unifying approach to the decoupling of

linear-multivariable systems. Int. J. Control 1986, 44, 181-201.
Witcher, M.; Mc Avoy, T. J. Interacting control systems: steady state and dynamic
measurement of interaction. ISA Trans. 1977, 16, 83-90.
Wonham, W. M. Geometric state-space theory in linear multivariable control: A

status report. Automatica 1979, 15, 5-13.
Wood, R. K.; Berry, M. W. Terminal composition control of a binary distillation
column. Chem. Eng. Sci. 1973, 28, 1707-1717.
Yu, C. C. Autotuning of PID Controllers: Relay Feedback Approach; Springer-
Verlag: New York, 1999.

Zhu, Z. -X.; Jutan, A. A new variable pairing criterion based on Niederlinski

Index. Chem. Eng. Commun. 1993, 121, 235.
Zhu. Z. -X.; Jutan, A. Loop decomposition and dynamic interaction analysis of
decentralized control systems. Chem. Eng. Sci. 1996, 51, 3325-3335.
Zhuang, M.; Ather ton, D. P. PID controller design for a TITO system. IEE
Proc.-D. 1994, 141, 111-120.
Ziegler, J. G.; Nichols. N. B. Optimum settings for automatic controllers. Trans.
ASME 1942, 64, 759-768.

Interaction Analysis and Loop Pairing for M u l t i v a r i a t e Processes 12
Chapter 3
Interaction Analysis and Loop Pairing

for Multivariable Processes
3.1. Introduction
Usually, the interactive multivariable systems can be either controlled by 1) a

multivariable or centralized MIMO controller or by 2) a set of SI SO decentralized
controllers. Algebraic decoupling methods or optimal multivariable control theory
are usually applied to obtain centralized MIMO controllers. Centralized
controllers are complex and weak in control system integrity. In contrast, the
decentralized control systems enjoy certain advantages: 1) the structure is easier to
understand, and implement, and requires fewer parameters to tune; 2) loop failure
tolerance of the resulting control system can be assured during the design phase;
and 3) both hardware and software realization is simple. Therefore, they are more
often used in process control applications (Grosdidier and Morari, 1987; Chiu and
Arkun, 1990). However, the potential disadvantage of using the limited control
structure is the deteriorated closed-loop performance caused by interactions
among loops as a result of the existence of nonzero off-diagonal elements in the
The main part of this chapter was published in Journal of Process Control 15 (2005) 741-747.

Interaction Analysis and Loop Pairing for Multivariable Processes —
transfer function matrix (Grosdidier and Morari, 1986; He and Cai, 2004). Thus,
the primary task in the design of decentralized control systems is to determine
loop configuration, i.e., pair the manipulated variables and controlled variables to
achieve the minimum interactions among loops so that the resulting multivariable
control system mostly resembles its single-input single-output counterparts and
the subsequent controller tuning is largely facilitated by SISO design techniques
(Seborg and Edgar, 1989).
As mentioned in Chapter 2, the RGA based techniques for control loop

configuration have been widely used in industry, including blending, energy
conservation, and distillation columns, etc (Bristol, 1966; Shinskey, 1988; Wolff
and Skogestad, 1995; Hansen et al, 1998). The most important advantages of
RGA based techniques are computational simplicity (since only process steady-
state gains are involved) and scaling independence (due to its ratio nature
(Grosdidier and Morari, 1985)). However, using steady-state gain alone may
result in incorrect interaction measures and consequently incorrect loop pairing
decisions since no dynamic information of the process is taken into consideration.
Even through several pairing methods have later been proposed by using the
transfer function model instead of the steady-state gain matrix to calculate RGA to
consider the effects of process dynamics to overcome the limitations of RGA
based loop pairing criterion, they are often controller dependent, making it more
difficult to calculate and to understand by practical control engineers. (Witcher
and Mc Avoy, 1977; Bristol, 1979; Tung and Edgar, 1981; Mc Avoy et al, 2003).
To combine the advantages of both RGA and DRGA. the steady-state gain and
bandwidth of the process transfer function elements are used in this chapter to
provide a more comprehensive description for loop interactions. Then, a new loop
pairing criterion based on the new interaction measure which minimizes loop
interactions is proposed in terms of relative energy array (REA) of the process.
The main advantages of the method are: 1) as compared with DRGA methods, it
requires much less computation, and does not require the specification of the
controller type; 2) it results in less conservative controllers when detuning factor
design methods are used; and 3) it is very simple for field engineers to understand
and work out pairing decision in practice. Several examples, for which the RGA

Interaction Analysis and Loop Pairing for Multivariate Processes li
based loop pairing criterion gives an inaccurate interaction assessment, are

employed to demonstrate the effectiveness of the proposed interaction measure
and loop pairing criterion.
3.2. Preliminaries
Consider a multivariate system with n inputs and n outputs as shown in

Figure 2.2, where G(s) and Gc.(s) are given in Equation (2.1) and (2.2),
respectively.
Assume that the system is open-loop stable and the process matrix G(s) is
nonsingular at steady-state. The loop pairing problem defines the control system
structure, i.e., which of the available plant inputs is to be used to control which
plant output. The most popular loop pairing method is the RGA and NI based
pairing rules as below (Bristol, 1966; Grosdidier and Morari, 1985: Niederlinski,
1971).
(1) The paired RGA elements are closest to 1.0;

(2) All the paired RGA elements are positive;
(3) Large RGA elements should be avoided;
(4) Niederlinski index is positive, i.e., NI > 0.
One of the main advantages of these methods is that the interaction depends
on only the steady-state gains. This information is easily obtained from simple
identification experiments or steady-state design models. A potential weakness of
these methods, however, is the same fact that they only use the steady-state gains
on which the assumption of perfect loop control to determine loop pairing is
based. We use the following example to illustrate this point.
Example 3.1. Consider a process given by (Mc Avoy et al, 2003)
5e
G(5) = 100s+ 1 \0s + l
-5e
105 + 1 100s+ 1
The RGA is
f0.8333 0.1667
A(G(0)) =
0.1667 0.8333

Interaction Analysis and Loop Pairing for M u l t i v a r i a t e Processes —
This result implies the diagonal pairing 1-1/2-2 is a good choice. Hence,
according to the RGA based loop-pairing criterion, the pairing of 1-1/2-2 should
be preferred for the smaller interaction to any one loop from other closed-loop.
However, Mc Avoy et al. used DRGA and optimal decentralized PI controllers for
various configurations, and found that the diagonal pairing resulted in a poor
closed-loop performance (Mc Avoy et al, 2003). The off-diagonal pairing 1-2/2-1
takes advantage of the fast g2l(s) and gl2(s) transfer functions to achieve better
response for y2:
(1) there is very little interaction from y{ to y2;
(2) the off-diagonal pairing v, response is significant by better than that of y2

with diagonal pairing;
(3) the y\ response for the diagonal pairing is somewhat better than that of the
off-diagonal pairing when y2 is given a step set-point change.

The main reason for the poor performance of the diagonal pairing is the
dynamic properties of the transfer functions. It can be easily seen that the time
constants and delays of 10 and 4 of the off-diagonal elements are 10 times smaller
than diagonal ones of 100 and 40. In such a case, pairing the faster loops (even
with smaller steady-state gains) takes the advantage of the time scale decoupling
such that seriousness of the interactions from the slower loop would be reduced.
3.3. A modified loop pairing rule
In designing decentralized controllers for multivariable processes, it is desired

that the interaction measures and loop pairing will address the following issues:
(1) The interaction measure should consider the finite bandwidth control, since
the assumption of perfect control is only valid for very low frequency range.
(2) The loop pairing decision should be controller independent such that any
controller type could be designed after loop pairing.
(3) The pairing results in minimal interaction within the interested frequency
range not only statically but also dynamically.
(4) It should be simple and easy to use for practical engineers.

53
Interaction Analysis and Loop Pairing for M u l t i v a r i a t e Processes
In decentralized control system design, individual loop is tuned around the critical
frequency region of the transfer function which is the region around the control
system bandwidth. Thus, this is the frequency region that should be focused upon
when considering the effect of interactions. Therefore, two factors in the open-
loop transfer functions will affect the loop pairing decision:
(1) Steady-state gain: the steady-state gain gu(0) of the transfer function
reflects the effect of manipulated variable u; to controlled variable v,.
(2) Response speed: response speed is accountable for the sensitivity of the
controlled variable y, to manipulated variable uf and, consequently, the
ability to reject the interactions from other loops.

In order to develop a new interaction measure which can reveal both gain and
phase information, let us factorize an arbitrary element in frequency domain.
g y ( » , i n G(ja) as
.0/ • x--#//;"
g.,Ua>) = g,j<f>)gl(ja>)e »' (3.1)
u$
where g„(0), g"{j(o) and e are steady-state gain, normalized transfer function
(g°(0) = 1) ar>d time-delay of g,j(jo)), respectively.
To use both steady-state gain and response speed information for interaction
measure and loop pairing, we now define the energy of gv(ja)) as the area in the
amplitude plot of g^ja) in the frequency domain, i.e.,
*,=MO) f>:(>>*"v dco (3.2)
where cocl] is the critical frequency of gv(ja).
As g,j(j(o)e ' represents the magnitude of the transfer function at various
frequencies, the physical implication of Equation (3.2) represents the effective

energy transmission ratio of gtl{j(o) from u] to yt.
In calculating etJ, the critical frequency, COLIJ can be defined as either
(1) 0)ciJ =(oh,j, where cociJ for i,j = l,2,--n is the bandwidth of the transfer
function &U*)e and determined by the frequency where the
magnitude plot of frequency response reduced to 0.707 , as shown in

Figure 3.1 (a), i.e.,

0/ • \ ~"n> •• 0 . 7 0 7 ,
or
(2) a)ciJ -CDUIJ, where aUJ) for i,j = \,2,-n is the ultimate frequency of the
transfer function g°Ua)e -0..'' ar

>d determined by the frequency where the
phase plot of frequency response across -n, as shown in Figure 3.1 (b)
i.e.,
arg -71 .
The value of etJ is either the area covered by the curve gt] (0) - P - Q - 0)b I; in
Figure 3.1(a) (when using a>hlJ as critical frequency), or by the curve
SA^)~Q~<0u4j m
Figure 3.1(b) (when using a>UIJ as critical frequency).
g*(0)
0.707gv(0)
«U
(a) Energy based on bandwidth a>b
e-Ttn Q
s e
i
I
Q.
Frequency
(b) Energy based on ultimate frequency 0)u tJ
Figure 3.1. Frequency response curves and the energy, et), of g^jco)

Interaction Analysis and Loop Pairing for Multivariable Processes 55
Applying Equation (3.2) to all transfer function elements, the overall process
energy an be written in the form of energy matrix and expressed as
"11
c.-, ... c,
'12 'In
E =
e
e i •>
To simplify the calculation, we approximate the integration of ev by a rectangular
area, i.e.,
ev * g,j(0)a>cy i,j = 1,2,—n
which is the area of the rectangle gv(0)-Q'-eohJJ in Figure 3.1(a) (when using
(oblJ as the critical frequency), or the rectangle g,J{0)-Q,-Q)ull in Figure 3.1(b)
(when using a>ulj as the critical frequency). Then, the energy matrix is given by
e e e
i\ \2 ••• \n
e2] e2Z ... e2n

E= = G(0)®Q (3.3)
_en\ e
nl ••• e
m _
where
"g„(0) g„(0) •• ft.(0)"

g 2 .(0) g 22 (0) - g 2 „(0)
G(0) =
&,(0) g.2(0) ••• g„(0)
and
(0 c.U a
cM - Vein
'c,2I ®c.22 - ^c.2.

Q =
fl'c-l <»c.«2 - ^c.™
are the steady-state gain and the critical frequency matrices, respectively. Even
though Equation (3.2) is more accurate, Equation (3.3) is easier to use. As the
energy matrix is used to derive relative interaction measures, the same
approximation using Equation (3.3) for both open-loop and closed-loop cases will
not affect the outcomes.

56
Interaction Analysis and Loop Pairing for M u l t i v a r i a t e Processes
Similar to relative gain, the relative energy, tj>n, of g,-(y'<y) is defined as the
energy ratio of the loop yt - Uj when other loops are open or closed, i.e.,
all loops open

l=% =
^(dy./dujde,
. all other loops closed except for loop y, - u ,
where d) • and e are the critical frequency and the energy of gn(s) when all
other loops are closed, respectively.

When the relative energies are calculated for all the input/output combinations
of a multivariable process, it results in relative energy array (REA):
£l <t>ll " <t>2n

0 = (3.4)
4
Following the same arguments as in DRGA and RGA, if all yk, k 5*/', are
perfectly controlled by the use of u,, l*i, then the energy from «, to v is equal
to the energy provided by l/| E~] 1 and Equation (3.4) can be conveniently
computed by
<& = E®E-T. (3.5)
Since REA is also a relative interaction measure, it has similar properties as
RGA and DRGA:
(1) The value of </> is a measure of the relative energy interaction expected in
the 1* loop if its output v, is paired with wy.
(2) The elements of the REA in any row or column sum up to 1, i.e.
(-1 7=1
(3) Let e represent the loop / area when all the other loops are closed,
whereas etJ represents the normal, open-loop area, then:
1
e = (3.6)
V

Interaction Analysis and Loop Pairing for Multivariate Processes —
(4) When ^ is negative, it produces an energy change in yl in response to a
change in u) when the other loops are open totally opposite in direction to
that when the other loops are closed.

Furthermore, when all n loops are closed, the multi-loop system will be
unstable for all possible values of controller parameters (i.e., it will be
"structurally monotonically unstable") if
NI=
l' G (°)l
V ;|
<0,
n&(°)
1=1
where G(0) denotes the determinant of matrix G(0) (Niederlinski, 1971).
Therefore, NI > 0 is a necessary stability condition, and consequently constitutes a

complementary tool to the REA in variable pairing selection.
Parallel to RGA and NI based pairing rules, the REA and NI based loop pairing
rules are given below:
(1) Corresponding REA elements are closest to 1.0.
(2) The NI is positive.
(3) All paired REA elements are positive.
(4) Large REA elements should be avoided.
Remark 3.1. For control system configuration, both bandwidth and ultimate
frequency can be used to find the REA. However, ultimate frequency (a>„„) is
more convenient to use since it is linked to control system dynamic performance

and familiar to control engineers. For the transfer function matrix with some
elements without phase crossover frequencies, such as transfer functions of first
order or second order without time delay, a finite time delay factor, e~A,e', can be
used to multiply all elements in G(s), such that ultimate frequency is available to
calculate relative energy for each element. As it is a relative measure, adding a
time delay to each element will not affect the relative measures. Without loss of
generality, we will use ultimate frequency in our following development.
Remark 3.2. Both REA and NI play important roles for control structure selection.
REA is used to measure interactions, while NI is used as a sufficient condition to
rule out the closed-loop unstable pairings. Since we are using energies instead of
simply steady-state gains, dynamic interactions up to the critical frequency can be

effectively reflected. Therefore, comparing REA method with RGA and DRGA
methods, it has the following advantages:
(1) In additional to steady-state gains, only critical frequency needs to be
calculated in REA method, it is far easier to calculate than those DRGA
methods.
(2) REA combines both steady-state gains and critical frequencies in
measuring the loop interactions, it provides better pairing results than that
of RGA based pairing and comparable with DRGA ones.
(3) REA only uses information of open-loop process transfer functions, it is
controller type independent.
3.4. Case studies
In this section, we use various examples to show the effectiveness of the above
REA based method in both simplicity and correctness.
Example 3.2 (Continue with Example 3.1),

5e-40s e-As
1005 + 1 105 + 1
G(3) =
5e^0s
105 + 1 1005 + 1
The RGA is
0.8333 0.1667
A(G(0)) =
0.1667 0.8333
This result implies the diagonal pairing 1-1/2-2 is a good choice. Hence, according
to the RGA based loop-pairing criterion, the pairing of 1-1/2-2 should be preferred
for the smaller interaction to any one loop from other closed-loop. However, in
REA approach, the energy matrix is
"0.0500 0.1000
E=
-0.5000 0.0500
and <j\x =0.0476 strongly suggests off-diagonal pairing.

Mc Avoy et al. used DRGA and optimal decentralized PI controllers for
various configurations, and found that the diagonal pairing resulted in a poor
closed-loop performance (Mc Avoy et al., 2003). In order to test the pairing

Interaction Analysis and Loop Pairing for Multivariable Processes —
results for this process, two sets of optimal PI controllers were designed for
diagonal and off-diagonal pairings, respectively. For each configuration, two
controller gains and two integral gains are optimized for a step change in _>>,
followed by a step change in y2 using the same objective function that was used
for the optimal control calculation. Equal weighting is given to the measurements
and the manipulated variables. In calculating the errors for y] and y2 the
difference between these measurements and their set-points was used. Since the
minimization involved in tuning the PI controllers is non-convex, several different
starting points are used to determine the best values of the tuning parameters.
After all, it is concluded that the DRGA correctly indicates that an off-diagonal
pairing produces a better overall control system response than that of the diagonal
pairing.
Example 3.3 Consider a 3x3 process (Huang et al, 1994, given by
-2e's \.5e-s es
105 + 1 5 + 1 5+1
1.5g" J
e" -2e's
G(s)
5+1 5 + 1 105 + 1
e" -2e" \.5e-'
5+1 105 + 1 5 + 1
The RGA of the system is

-0.9302 1.1860 0.7442
A = 1.1860 0.7442 -0.9302
0.7442 -0.9302 1.1860
Obviously, two possible parings 1-2/2-1/3-3 and 1-3/2-2/3-1 are comparable

because all related RGA elements close to 1. Therefore, RGA pairing approach
cannot determine which pairing is better.
Using REA method, the critical frequency matrix
"0.1 1 1 "
a= I I o.i
1 0.1 1
and REA of the system is obtained as

"0.0554 0.6977 0.2468
O 0.6977 0.2468 0.0554
0.2468 0.0554 0.6977

This indicates the best paring is 1-2/2-1/3-3 because the corresponding REA
elements are much closer to 1 than the other option. This was confirmed by the
generalized dynamic relative gain (GDRG) approach (Huang et al, 1994).
However, the proposed method is much simpler.
Example 3.4 Consider a 4x4 industrial reactor/recycle system (Robinson et al,
2001) given by
3.96(1975 + 1) 0.536(2585 + 1) 9.7
0
49.4s2 +14. Is+ 1 83.3s2 +18.35 + 1 24.3s + l
0.00111 0.044 -0.0152<T32s 0.039g-2Oj
s 46.9s2 + J s
G(5) =
-232.2 -15.96(529s T!) 139.2
0
32.25 + 1 10417s 2 +204s+1 7.27s + l
-0.582 -2.54(8.115 + 1) 0.0462(45.65 + 1K 35 ' -0.0358e~30s
s 6.25s3 + 5s 2 +s 306s3 + 35s 2 +s s
As proposed by Arkun and Downs (1990), for processes that contain pure
integrator elements (that has no steady-state) the steady-state gain and RGA of
integrating processes can be calculated by letting / = 1 / s, substituting / into the
original transfer function matrix, and taking the limit /C = limG(s) in calculating
j->0
the RGA
0 0.107 0.148 0.745
-0.072 0.196 0.829 0.046
A=
0 0.723 0.067 0.210
1.072 -0.027 -0.044 0
It indicates the parings: 1-4/2-3/3-2/4-1.
To use REA method, the bandwidth matrix of G(s) and REA are obtained,
respectively, as
0 5.6400 4.3800 0.0410
1.4500 0.1750 1.4500 1.4500
fi =
0 0.0310 0.0710 0.1380
1.4500 1.2980 0.4550 1.4500
and
0 0.7856 0.1387 -0.0047
-0.0806 -0.0285 1.2895 -0.0244
<D =
0 -0.0026 -0.0007 0.8426
0.9438 -0.0291 0.0215 0

This indicates parings: 1-2/2-3/3-4/4-1. This is the same as the pairing option
given by the DRGA approaches (Mc Avoy et al, 2003; Robinson et al, 2001).
Robinson et al (2001) confirmed this conclusion and they showed that the
transients produced by this pairing option are significant superior to those
produced by pairing option suggested by RGA. Nevertheless, the proposed
method is much more direct and the computation is much easier.
3.5. Conclusions
In this chapter, both the steady-state gain and the critical frequency of the
process are used to provide a simple yet comprehensive description of loop
interactions for MIMO processes. The REA can be conveniently calculated by
control engineers since the bandwidth or ultimate frequency can be easily
obtained from the given transfer function matrix. The effectiveness of the method
is demonstrated by several examples, for which the RGA based loop pairing
criterion gives an inaccurate interaction assessment, while the proposed
interaction measure and loop pairing criterion provides accurate results and they
are very easy to be calculated. Furthermore, the REA method is not only an
effective tool for loop pairing, but also very useful in the design of the
decentralized and decoupling control systems, especially, for high dimensional
processes. These issues will be discussed later.
References
Arkun, Y.; Downs, J. A general method to calculate input-output gains and the
RGA for integrating processes. Comp. Chem. Eng. 1990. 14, 1101-1110.
Bristol, E. H. On a new measure of interactions for multivariable process control.
Bristol, E. H. Recent results on interactions in multivariable process control. In

Proceedings of the 71" Annual AIChE Meeting, Houston, TX, USA; AIChE, 1979.
Chiu, M. S.; Arkun, Y. Decentralized control structure selection based on integrity

considerations. Ind. Eng. Chem. Res. 1990, 29. 369-373.


Grosdidier, P.: Morari, M. Interaction measures under decentralized control.

Automatica, 1986, 22, 309-319.
Grosdidier, P.; Morari, M. A computer aided methodology for the design of
decentralized controllers. Comp. Chem. Eng. 1987, 11, 423-433.
Hansen, J. E.; Jorgensen, S. B.; Heath, J.; Perkins, J. D. Control structure selection
for energy integrated distillation column. J. Proc. Cont. 1998, 8, 185-195.
He, M.-J.; Cai, W.-J. New criterion for control loop configuration of multivariable
processes. Ind. Eng. Chem. Res. 2004, 43, 7057-7064.
Huang, H.-P.; Ohshima, M.; Hashimoto, I. Dynamic interaction and multiloop

control system design. J. Proc. Cont.1994, 4. 15-27.
Mc Avoy, T.; Arkun, Y.; Chen, R.; Robinson, D.; Schnelle, P. D. A new approach
to defining a dynamic relative gain. Control Eng. Practice 2003, 11, 907-914.
Niederlinski, A. A heuristic approach to the design of linear multivariable

interacting subsystems. Automatica 1971, 7, 691-701.
Robinson, D.; Chen, R.; Mc Avoy, T.; Schnelle, D. An optimal control based
approach to designing plantwide control system architectures. J. Proc. Cont. 2001,
11,223-236.
Seborg, D. E.; Edgar, T. F.; Mellichamp, D. A. Process Dynamics and Control.

John Wiley and Sons: New York, 1989.
Shinskey, F. G. Process Control Systems; McGraw-Hill: New York, 1988.
Tung, L.; Edgar, T. Analysis of control-output interactions in dynamic systems.

AIChEJ. 1981, 27, 690-693.
Witcher, M.; Mc Avoy, T. J. Interacting control systems: steady state and dynamic
measurement of interaction. ISA Trans. 1971, 16. 83-90.
Wolff, E. A.; Skogestad, S. Operation of integrated Three-Product (Petlyuk)

distillation columns. Ind. Eng. Chem. Res. 1995, 34, 2094-2103.

63
Decentralized Control System Design
Chapter 4
4.1. Introduction
In process control industry, more than 95% of the control loops are of PI/PID
type (Astrom and Hagglund, 1995). This is mainly attributed to its effectiveness
and relatively simple structure, which can be easily understood and implemented
in practice. Consequently, the research on PID control algorithm development and
their applications is still a very active area; many formulas have been derived to
tune the PID controllers over the years.
Due to the high product quality and energy integration requirements, most of
modern industry processes, however, are MIMO processes. For easier field
implementation, it is desirable to apply well established single loop PID tuning
principles to these MIMO processes (Grosdidier and Morari, 1987). However,
compared with SI SO counterparts, MIMO systems are more difficult to control
due to the existence of interactions between different input and output variables.
Adjusting controller parameters of one loop affects the performance of the others,
sometimes to the extent of destabilizing the entire system. To ensure stability,
many decentralized controllers in industry are tuned loosely, which causes
inefficient operation and higher energy costs.
The main parts of this chapter were published in Industrial and Engineering Chemistry Research
45 (2006) 2769-2776 and Journal of Process Control 16 (2006) 773-784.

64
Although considerable effort has been dedicated to this problem and many
design techniques, such as sequential loop closing methods (Lee et al, 1998; Loh
et al., 1993; Shen and Yu, 1994), detuning factor methods (Luyben, 1986; Chen
and Seborg, 2003), and independent design methods (Grosdidier and Morari, 1986;
Skogestad and Morari, 1989; Pomerleau and Pomerleau, 2001; Huang and Jeng,
2002; Huang et al, 2003) have been proposed, decentralized control system
design and implementation is still difficult to deal with for control engineers due
to the lack of simple and practical approach. Since controllers interact with each
other in a decentralized control system, the performance of one loop cannot be
evaluated without the information of the controllers of other loops.
In this chapter, a novel equivalent transfer function method in terms of REA

for multivariate processes is proposed. By considering four combination modes
of gain and phase changes for a particular loop when other loops are closed, this
equivalent transfer function can effectively approximate the dynamic interactions
among loops. Consequently, the design of decentralized controller for MIMO
processes can be converted to the design of single loop controllers. The method is
simple, straightforward, easy to understand and implement. Several multivariable
industrial processes with different interaction characteristics are employed to
demonstrate the effectiveness and simplicity of the design method compared with
the existing methods.
4.2. Equivalent transfer function
Consider an open-loop stable multivariable system as shown in Figure 2.2

with decentralized control structure, Equations (2.1) and (2.2), respectively. From
Chapter 3, we have the relative energy, ^, between output variable y, and input
variable ul:
*,-. (4.1)
'J
where e,y is the effective energy transmission ratio between output variable y, and
input variable ul when all other loops are closed. When the relative energy are

65
calculated for all the input/output combinations of a multivariable process, it

results in an array of the form, i.e., REA, which can be calculated by
<D = E ® E " ' = * ' ** - ^n
In this chapter, we will employ this interaction measure to develop equivalent

transfer functions (ETFs) for controller design under decentralized control
structure.
Suppose that the best loop configuration has been determined and the best
pair is diagonally placed in the transfer function matrix as shown in Figure 2.2. As
mentioned in Chapter 3, let the energy of g,j(s), etJ, when all other loops are
closed be el} =gij(0)a>cjj i,j = l,2,---«, where g,7(0) and cbCIJ are the steady-state
gain and ultimate frequency between output variable v, and input variable
Uj when all other loops are closed, respectively. Then, from Equation (4.1)
(4.2)
By the definition of RGA, we have
*»•*? (4.3)
where Atl is the relative gain.
Substitute Equation (4.3) into (4.2) and rearrange to result

CO
(4.4)
where yIJ represents the critical frequency change of loop i-j when other loops are
closed, defined as relative critical frequency. When the relative critical

frequencies are calculated for all the input/output combinations of a multivariable
process, it results in an array of the form, i.e., relative critical frequency array
(RCFA) defined as

YlH
Y%\ Y22 Y2„
r= (4.5)
Y „\ Yn2 "• Yn,

The elements of Equation (4.5) can be calculated by
r = o<g>A
0.1 0i2 - K
<t>2X A l - <t>2n (4.6)
<t>M
l_~«l <t>n2 4> _1/X, l/i, 2 ... 1/A,
In Chapter 3, we have derived the energy for a control loop when other loops
are closed. In the following, we will derive ETFs, g,;(s), using REA for loop
v, - uj when all other loops are closed.
Since control loop transfer functions when other loops closed will have similar
frequency properties with when other loops open if it is well paired (Chien et ai,
1999), we can let the ETFs have the same structures as the corresponding open-
loop transfer functions but with different parameters
i„(s) = i,(0)g,r,(s)e-.-0„t'"s (4.7)
where g'„ (s) is defined by
g;,(s)=g:(s)^
and §ti is the time delay of the ETF.
As the change in ultimate frequency of a control loop is generally affected by

changes in both time constant and time delay when other loops are closed, and
they may be reasonably represented by a linear approximation, it is reasonable to
change only time delay to reflect the phase changes.
In Equation (4.7), g,y(0) can be determined by using Equation (4.3), while by the
definition of the ultimate frequency.
-3d) +Zg°(jd) ) = -Gco +Zg°(a) ) = -/r,

IJ C.IJ blJ^J C,t] ' Ij c,tj o;y\ c,/// '
0n can be easily determined by

h7
0„ = - Y„ • t4-8)
Notice that
(1) g,'(s) is usually low order transfer functions, their contribution to the
phase change at low frequency range are small and can be equivalently
represented by the additional time delay term.
(2) In many decentralized control system designs, such as gain and phase
margin method, individual loop is tuned around the critical frequency
region of each control loop. Accurate estimation is required around critical
frequency point, not who contribute to the change.
Assume the contribution of g°(s) in the phase change at critical frequency is
small, we can let Zg'n(jcoc „)« Zg'n(coc „), and make further simplification to
Equation (4.8) as
-6 6) * -6> to
IICJI II C.ll
which results by considering Equation (4.4)
4*^r=H=^ (4.9)
This is the practical formula which will be used to create ETFs. Even though
Equation (4.9) is less accurate than Equation (4.8), several simulation results have
showed that the control system performances are comparable by the two
approximations, but Equation (4.9) is much more straightforward and easier to
understand than Equation (4.8).
Since it is desirable that the controlled system possesses integrity property;
that is, the overall control system remained stable regardless of input and output
of other control loops, g„(0) and 6lt in ETF must take different values for
different combination of kn and yn. For the four different combinations of \ and
Yu, S,,(s) ma
y take different modes shown in Figures 4.1-4.4, and are discussed
below:
Case 1: Au < 1, yu < 1, Figure 4.1.

In this case, 1 >0and (yn - l ) < 0 . According to Equations (4.3) and (4.9),
v4< )
we haveg„(0) > g„(0), and 0„ < 0„.
• Sn (0) > g„ (0). This means that the magnitude of the frequency response
when the other loops closed is not less than that of when the other loops
open. Since the retaliatory effect from the other loops magnifies the main
effect of ut on yn we need to reduce the controller gain to assure system
stability. In this case, the gain is by Equation (4.3)
2„(0) = g„(0)M,.
• 0n < 0U. This means that the time delay when the other loops closed is not
bigger than that of when other loops open. The reduced time delay will
increase the phase margin. However, by considering the control system
integrity, the time delay needs to be kept as before, i.e..
0. = 0..
11 11
T T r
' • • 1 • 1 '• 1 •
g*(0)
1 1 1 1 1111 1 1 1 i T S
Frequency
Figure 4.1. Interaction mode with kn < 1, ytt < 1
Case 2: ku < 1, yu > 1, Figure 4.2.
In this case, ' 1 - ; >oand (yn - l ) > 0 . According to Equations (4.3) and (4.9),
>*
we have g„ (0) > g„ (0), and 0„ > 0„.

M
• gu (0) > g„ (0). Same as in Case 1,
g„(0) = g„(0)/A„.
• 6n > 0„. This means that the time delay when the other loops closed is
bigger than that of when the other loops open. The enlarged time delay
will reduce the phase margin. In this case, the time delay is determined by
Equation (4.9)
0 =y 0 .
II / I! II
1,(0)
!g,(0)
en
ID
2
• * * • " ' • xJ 1—i—i i m i l 1—I—I I i n i l ' • *' • « ' • • •
Frequency
Figure 4.2. Interaction mode with ku < 1, yu > 1
Case 3: An > 1, yh < 1, Figure 4.3.
1
In this case, •1 <0and (y„ -1) < 0 . According to Equations (4.3) and (4.9),
v4. J
wehaveg„(0)<g„(0), and0 (1 <0„.
• g„(0) < g„(0) • This means that the magnitude of the frequency response
when the other loops closed is smaller than that of when the other loops
open. Even if the retaliatory effect from other loops acts in opposition to
the main effect of u, on yt, we cannot enlarge the controller gain for
better performance due to the system integrity consideration. Hence, the

gain should be unchanged, i.e.,
&(0) = &(0).
Ncmyang Technological University

70
0..
a
< /(0„. Same as in Case 1
0=0.
n n
Frequency
Figure 4.3. Interaction mode with An > 1, yn < 1
Case 4: An > 1, yu > 1, Figure 4.4.
In this case. <0and (yHr-l)>0. According to Equations (4.3) and (4.9).
wehaveg„(0)<g„(0), and0„>0„.
• g„ (0) < gn (0). Same as in Case 3,
£„(0) = g„(0).
• 0n > 0n. Same as in Case 2,
0 =y0 .
^11 / a II

71
g,(0)
g,(0)
• ...........i. ..!.... 1...1...l.l..ixi................1. ....i.....l....t...l.i,.*.tl... .. i. ,M.l A....M.4 jt...il.iiil>j,i*a
1 -
://
^ S * , "-». "^
-
! i l l ^ \
CO <Z>. jj
Frequency
Figure 4.4. Interaction mode with AH > 1, yfi > 1
4.3. SISO controller design
To illustrate the control system design procedure based on ETFs, the gain and
phase margins design approach (Wang and Cai, 2002) is presented here for the
sake of completeness.
Without loss of generality, we assume that each main loop, i.e., diagonal
element in the transfer function matrix is represented by a SOPDT model, which
can be used to describe most of the industrial processes:
•'O.u
&,(') =
a
2MS +
°UiS + l
Similarly, ETF is represented as
&(0) --#/!
& ( * ) = •
a S a S+]
2,n '+ u,
The decentralized controllers can then be independently designed by single loop
approaches based on corresponding ETFs. Here we employ the gain and phase
margins approach. This is primarily because the frequency response method
provides good performance in the face of uncertainty in both plant model and
disturbances. The PID controller of each loop is supposed of the following
standard form:

As2+Bs + C
gcA*) = k„+ — + kIIJs = k
s s
where A- kdl/k , B = kpJk and C = klt/k . By letting A = a2, B = a, and C = 1,
the open-loop transfer function becomes
Denoting the gain and phase margin specifications as Ami and *¥mJ, and their
crossover frequencies as cog, and at ,, respectively, we have
M$[gcAJ°>llMi(Jtog.J]=-7r
^./kc,(H,,)^,o?,)|=1
|^c,(H,)i,-0'^,,)| = 1
*,., = «" + arg [ & , O p , , )£„ 0 » „ ) ]

By substitution and simplification to above equations, we obtain
it
a 9 =• JL.=.°»
kg,m
kgM = o>PJ ¥ =--<a 9

m,i « />,/ i/
which result
n
2 4./ 2,4 ftx..(0)
By this formulation, the gain and phase margin are interrelated to each other,
some possible gain and phase margin selections are given in Table 4.1.
Table 4.1. Typical gain and phase margin values
V.J njA T/3 3TT/8 2^/5
4., 2 3 4 5
The PID parameters are given by
~kp.i
7T *!,«
(4.10)
J.I
24A&«» <72.H

73
Applying Equation (4.10) for each case discussed in Section 3, we can easily
obtain both ETFs and the P1D parameters which are summarized in Table 4.2.
Table 4.2. ETFs and PID controllers for gain and phase margin method
Mode aW K, KM K,
&(0)/4 c -,„ J *Kau, *K XA„a2,„
a
2.ns2+ai„S +
^ iÂsM T-KAzM 2AmJdHgH(0)
n
g„(0)/A„ cê„s Kau, *K n a
k i.u
2Amj/„0ug„(O) 2Amjy„0„gAO) 2AmjAg„(0)
&(0) c-e„s ™\.n n n<*i,n

a2y+aiJls + ^ 24.^„R,(0) 2Am,0llg„(.O) 2AmA,g„(0)
4*L S„(0) c-,„e„*

na
v„ n na
2.»
a
2 , „ s 2 +
^ M S + X
" ^mjYA,sM 2Amj„0ugll(O) 2Amj,A,g„(0)
Remark 4.1. A unique problem for decentralized control of MIMO processes is

the zero crossing (Cui and Jacobsen, 2002): stable or unstable zeros might be
introduced into a particular control loop when other loops are closed. If an
unstable zero is introduced, it will result in a phase shift to the left in the
frequency domain. In order to guarantee the entire system stability, the controllers
are normally conservatively designed by conventional detuning approaches (Cui
and Jacobsen, 2002). By introducing the relative critical frequency, yu, to indicate
phase changes after the other loops closed, the effects of unstable zeros can be
accurately estimated in each control loop. Consequently, the resultant control
systems will be much less conservative.
Remark 4.2. Mathematically, the equivalent transfer function should incorporate

the controllers of all other loops. To solve such a complex problem, recursive
solution is required by first assigning initial controllers, then finding the
equivalent loop transfer functions and designing controllers again. This process is
continuous until a stable solution is obtained. To simplify the problem, both
detuning and independent methods proposed so far assumes that all other closed-
loops are under perfect control when design controller for a particular loop and
consider only the gain change. In the proposed method, the changes are
considered for both gain and frequency. As will be shown later, it is far more
accurate than the current existing methods.

As can be seen, the design procedure is simple, straightforward and can be

fully automated. The procedure for designing decentralized PI/PID controller for
MIMO processes can be summarized as follows.
Algorithm 4.1:
Step 1. For a given transfer function matrix, calculate E, <S>, A and V by
Equations (3.3), (3.5), (3.1) and (4.6), respectively.
Step 2. Select the best pairing and put them on the diagonal position
according to REA, determine the equivalent transfer functions for
diagonal elements by calculating the two correcting factors using
Equations (4.3) and (4.9), respectively.
Step 3. Determine the parameters of the diagonal controllers by looking up
Table 4.2.
4.4. Case studies
In this section, we apply the proposed design method to a number of industrial

processes. The gain and phase margins for all examples except Example 4.2 are
specified to be 3 db and ;r/3 rad , respectively. Those margins are generally
regarded as making the best compromise between performance and robustness in
process control. The proposed design method is compared with the following
popular decentralized control design approaches to show the effectiveness of the
proposed design method:
(1) The Ziegler-Nichols with detuning factor approach proposed by Mc Avoy
(1983):
(2) The biggest log modulus (BLT) tuning approach proposed by Luyben
(1986):
(3) The RGA based tuning approach proposed by Chien et al. (1999);
(4) The sequential loop tuning approach proposed by Shen and Yu (1994);
(5) The relay based auto-tuning approach proposed by Loh et al. (1993); and
(6) The describing function matrix approach proposed by Loh and Vasnani
(1994).
Example 4.1. Consider a process given by

75
5>
5e" 3 s 2.5<?
45 + 1 15^ + 1
G(5) =
-4e-*s
^205 + 1 55 + 1
This process seems easy to control. However, as can be seen below, both variable
pairing and control design are difficult with conventional approaches.
The RGA, critical frequency array, REA and RFA are given respectively by
0.3333 0.6667 0.6308 0.2732
A = Q:
0.6667 0.3333 0.2856 0.4909
0.6649 0.3351 1.9948 0.5026

o 0.3351 0.6649 0.5026 1.9948
It can be seen that the best pairing according to RGA (0.6667) and NI (=1.5000>0)
is off-diagonal, i.e., 1-2/2-1. Using the method (1), we obtain the detuning factor
0.6667+ V0.66672-0.6667 =0.8165 , and corresponding PI controller
parameters which results in an unstable system. By keeping detuning the original

PI controller using the same detuning factor four times, we obtain the PI controller
parameters as listed in Table 4.3. Both loops are stabilized but with very
oscillatory response as shown in Figure 4.5.
Following REA (0.6649) and NI (=0.6152>0), the best pairing is diagonal, i.e., 1-
1/2-2. As An < 1 and y„>\, the equivalent process for two loop are calculated by
proposed method as
-59844.! -19192s
15.0000? 3.0000?
and
45 + 1 55 + 1
respectively. The resulting PI controllers and the closed-loop responses are also
listed in Table 4.3 and shown in Figure 4.5. respectively, where the unit set-points
change in r, at t=0 and /-, at t=1250. It can be seen that the proposed method
results in a better loop pairing and control system performance while the
conventional detuning method (1) results in a unsatisfactory performance even
detuning by four times.


Controller Proposed Mc Avoy
K, T
u. K, T
,.»
Loop 1 0.0233 4.0000 -0.3333 18.0000

Loop 2 0.1094 5.0000 0.4800 15.0000
1000 1500 25CC

Time(minute)
1.E
05
|5pp*
-05
51 1000 1500 210 2500
Time(minute)
Figure 4.5 Closed-loop responses for Example 4.1

Example 4.2. Consider an industrial-scale polymerization reactor given by (Chien
etai, 1999)
22.89 ^ , -11.64 _*,
4.5725 + 1 1.807^ + 1
G(s) =
4.689 ^ 5.80 ,-0 4j
-e
.2.1745 + 1 1.8015 + 1
The time scales are in hours, so it is a quite slow process. In addition, it is easy to
verify that it is not diagonally dominate. The two controlled variables are two
measurements representing the reactor condition, and two manipulated variables
are the set-points of two reactor feed flow loops with load disturbance as the purge
flow of the reactor.

-7
0.7087 0.2913 8.0554 4.1888
A= fi
0.2913 0.7087 7.8540 4.3036
0.7193 0.2807 1.0151 0.9633

0 = r = 0.9633
0.2807 0.7193 1.0151
Both REA and RGA indicate diagonal pairing (NI =1.4111>0). In this example,
we have \ < 1 and yn > 1 for ;' = 1,2 . According to the proposed method the
equivalent process for the two loop are calculated as
32.3003 _02030s 81844 - 0 4060s

-e and
4.572^ + 1 """ 1.801J + 1
respectively. For the gain and phase margins of 5 and 2^/5. respectively, the
resulting PI controllers determined by the proposed method together with those
determined by methods (2), (3) and (5) are listed in Table 4.4.
Controller Proposed Luyben Chien et al. Loh et al.
K, T
,.„
K, T
,,n K, T
,.n K, T
r,„
Loop 1 0.2190 4.5720 0.210 2.26 0.263 1.42 0.620 0.60

Loop 2 0.1703 1.8010 0.175 4.25 0.163 1.77 0.247 1.78
Figure 4.6 shows the closed-loop responses for different controllers, where the
unit set-points change in /-, at t=0 and r2 at t=20. It can be seen that for this
process which is not diagonally dominant, the proposed design method gives the
best performance in terms of overshoot and oscillation.

7S
1 8
Proposed
1.6 Luyben
' 4
* — Chien et al.
Loh et al
1 2 ! >'V, -
T
•
GB
0.6 3 -
04
0.2
5 10 15 20 25 30 35 40
Time(hour)
15
Proposed
— Luyben !*;
— Chien et al
Loh et al
ip
<N 05 •
[A
!; '
-0 5 i i i i i i
15 20 25 tl 40
Timefhour)

To investigate the robustness of the system, we mismatch the process model by
increasing all four time delays by 25%. The closed-loop responses under same
controllers are shown in Figure 4.7. As can be seen, the proposed method still
shows the best performance.

T ™ T '•" t I 1
Proposed
— Luyben
— Chien et al
15 • Loh et a I
35 I
_i i—
10 15 2D 25 30 35 40
Time(hour)
Proposed
Luyben
1 5 Chien el al.
Loh et al
35
A
w*
0 5
15 20 25 30 35 43
Time(hour)

Example 4.3. Consider the 3x3 process given by (Loh and Vasnani, 1994)
119<?" I53e" -l.\eis

21.75 + 1 3375 + 1 lO.v + 1
5
37.0e~ ' -5.0e" 5 s
G(5) =
500s +1 285 + 1 105 + 1
93.0e 55 -66.7e"5* -103.3e" 5 j
5005 + 1 1665 + 1 235 + 1
Loh and Vasnani used this process to verify their design for high dimensional
systems. It is also not diagonally dominate process.
2.1896 -1.1446 -0.0449 0.3290 0.3222 0.3611

-1.3147 2.0677 0.2470 Q = 0.3142 0.3307 0.3696
0.1252 0.0769 0.7979 0.3142 0.3142 0.3491
2.0095 -0.9693 -0.0402 0.9178 0.8468 0.8948

d> = -1.1226 1.8976 0.2250 r = 0.8539 0.9177 0.9108
0.1131 0.0717 0.8153 0.9034 0.9322 1.0217

The best pairing according to both RGA, RE A and NI (=0.4759 >0) is 1-1/2-2/3-3.
For i = 1,2, \ > 1 and yu < 1. Hence, the original transfer functions are selected
as corresponding ETFs. For ;' = 3, X,t < 1 and yn > 1. The equivalent process is
calculated as
-129.4574e'-'084'
235 + 1
The resulting PI controllers by proposed method are listed in Table 4.5 together
with those of method (6).
Controller Proposed Loh and Vasnani
K, T
u, K, T
,.n
Loop 1 0.0191 21.7000 0.0181 14.4936

Loop 2 0.0382 28.0000 0.0335 14.4936
Loop 3 -0.0182 23.0000 -0.0260 14.4936
Figure 4.8 shows the closed-loop responses, where the unit set-points change in r,
at t=0, r2 at t=100 and r3 t=200. As can be seen, the proposed method results in a
better performance with the integral times much different from those of Loh and
Vasnani.

81
1 4
Proposed
— Loh and Vasnam
1 2
1
A. „.
08
>
26
04
•
32 ' •••
50 100 150 200 250 30C

Time(second)
• 4
Proposed
1.2 / N. — Loh and Vasnani -
_^^_
"7
1
08
0.6 -
04
02
-0 2 ' •
50 100 150 200 250 300

Time(second)
- Proposed
12 . . Loh and Vasnani -
1 V
1
0.8
J
1
06
1
04 •
1
0.2
1
0
1
•02
V^"
50 100 150 200 250 30C
Time(second)

Example 4.4. Consider the 4x4 process given by (Doukas and Luyben, 1978)

82
-1.15s
-9.81 Wf 0.374e -2.368e -27.33s •11.3e"
11.365 + 1 22.225 + 1 33.3-v + l (21.745 + 1)2
5.984<T224j -1.986e-°71i 0.422e~872s 5.24<?-*0y
14.295 + 1 66.675 + 1 (2505+ 1)2 4005 + 1
G(s) =
2.3&?-0-42' 0.0204^ 5Qs 0.513<TS -0.33e^ 58i
(1.435 + 1)2 (7.14.v~l) : 5+ 1 (2.385+ 1)2
-11.3e"379s -0.176e-° 48j 15.54e-J 4.48e"°52s
(21.745+ 1)2 (6.95+ 1)2 5+1 11.115 + 1
This is a real process with complicated delay and interaction modes. The problems
of some of the unrealistic pathological cases would be avoided using such kind of
processes to verify design.
The RGA, critical frequency array, REA and RFA are given by
"0.1264 -0.1013 -0.0314 1.0063"

0.0107 1.0935 0.0003 -0.1045
0.7264 0.0025 0.1630 0.1081
0.1366 0.0054 0.8680 -0.0099
1.0134 0.2285 0.0757 0.1571

0.7306 2.2440 0.0306 0.0279
Q=
1.7952 0.6830 1.9635 1.0833
0.1551 0.7854 2.0268 3.1416
"0.5237 -0.0052 -0.0222 0.5036"

-0.0024 1.0035 0.0000 -0.0012
o= 0.4719 0.0012 0.2965 0.2303
0.0067 0.0004 0.7256 0.2673
and
4.1448 0.0511 0.7066 0.5004
-0.2239 0.9177 0.0978 0.0113
r= 0.6497 0.4888 1.8192 2.1309
0.0492 0.0793 0.8359 -26.9057
respectively. The best pairing according to REA and NI (=1.1814>0, ideal pairing)
is 1-4/2-2/3-1/4-3. This example has complicated interaction modes. For / = 1,2,
\ > 1 and y„ <\ , the main process is selected as corresponding equivalent
process. For # = 3,4, \ < 1 and y„<\, the equivalent process is calculated as

Decentralized Control System Design x;,
17.903 l e " -11.2289<T379'

and
5+1 (21.745 + 1)2
respectively. The resulting PI/PID controllers by the proposed method are listed in
Table 4.6 together with method (2) based on the diagonal pairing. It should be
noticed that different designs have much different controller parameters due to
different pairing schemes.
Controller Proposed Luyben
*„, T
,M
T
d,„ *,, T
,,u
Loop 1 1.4981 2.8600 0.7150 -0.084 33

Loop 2 -24.7566 66.6700 0 -5.16 15.5
Loop 3 0.0292 1 0 0.305 17.0
Loop 4 -0.5349 43.4800 10.8700 0.529 11.2
at t=0, r2 at t=400, r3 t=800 and r4 t=1200. Although there are larger overshoots
in some loops, the system has better overall performance. This example illustrates
that the proposed method can be easily applied to high dimensional processes with
complicated interaction modes. In fact, the dimension does not increase the design
difficulty.
Ncmyang Technological University

200 400 600 800 1000 1200 1400 1600

Time(minule)
1 6
Proposed
1 4 — Luyben
1.2
1
-t—— Hr\A- - / -
DE
3E
14
31
-0 2
0 200 400 600 800 1000 1200 1400 1600
Time(minute)
02
0 200 400 600 800 1000 1200 1400 1600
Time(minute)
0 200 400 600 800 1000 1200 1400 1600

Time(minute)

To investigate the robustness of this higher dimensional system, we mismatch the

process model by increasing all four main process steady-state gains by 70%. The
closed-loop responses under same controllers are shown in Figure 4.10. It is noted
that the proposed method still shows satisfactory performance (almost no change)
under such a severe model variation while the other one is more oscillatory.

Sh
0 200 400 600 800 1000 1200 1400 1600

Time(mtnute)
16 1 1
Proposed
1 i - Luyben
12 r h .
I M A . . . J.
1 - 1'• /
)!\\,!J' •
06
V
06
0.4 • -
02
0
' ' •
-02 • •
0 200 400 600 800 1000 1200 1400 1600

Time(minute)
0 200 400 600 800 1000 1200 1400 1600

Time(mmute)
0 200 400 600 800 1000 1200 1400 1600

~in°;minutej

X7
4.5. Conclusions
This chapter presents a novel method to construct equivalent transfer function

models for decentralized control system design of multivariable interactive
processes. The simplicity and effectiveness of the model is based on the energy
transfer ratio of each individual transfer function which provides necessary
information of gain and frequency changes when all other loops are closed.
Consequently, the decentralized controllers can be obtained by simply using
single loop design approaches. Simulation results for a number of industrial 2 x 2,
3*3 and 4*4 processes show that the proposed method provides a better overall
performance compared to other reported design approaches. The advantage of the
proposed method is even more significant when applied to higher dimensional
processes with complicated interaction modes. The method can be easily
integrated into an auto-tuning control structure combined with some on line
parameter identification module and implemented for industrial control systems.
References
controllers with no proportional kick. Ind. Eng. Chem. Res. 1999, 38, 1456-1468.
Cui, H.; Jacobsen, E. W. Performance limitations in decentralized control. J. Proc.
Cont. 2002, 12, 485-494.
Doukas, N.; Luyben, W. L. Control of sidestream columns separating ternary

mixtures. Instrum. Techno!. 1978, 25, 43.
Grosdidier, P.; Morari, M. Interaction measures under decentralized control.

Automatica, 1986, 22, 309-319.
Grosdidier, P.; Morari, M. A computer aided methodology for the design of
decentralized controllers. Comp. Chem. Eng. 1987, 11, 423-433.

Decentralized Control System Design xs
Huang, H.-P.; Jeng, J.-C. Monitoring and assessment of control performance for
single loop systems. Ind. Eng. Chem. Res. 2002, 41, 1297-1309.
Huang, H.-P.; Jeng, J.-C; Chiang, C.-H.; Pan, W. A direct method for multi-loop
PI/PID controller design. J. Proc. Cont. 2003, 13, 769-786.
Lee, J.; Cho, W.; Edgar, T. F. Multiloop PI controller tuning for interacting
multivariable processes. Comp. Chem. Eng. 1998. 22, 1711-1723.
Loh, A. P.; Hang, C.C.; Quek, C. X.; Vasnani, V. U. Autotuning of multiloop
proportional-integral controllers using relay feedback. Ind. Eng. Chem. Res. 1993,
32,1102-1107.
Loh, A. P.; Vasnani, V. U. Describing function matrix for multivariable systems
and its use in multiloop PI design. J. Proc. Cont. 1994, 4, 115-120.
Luyben, W. L. Simple method for tuning SISO controllers in multivariable

systems. Ind. Eng. Chem. Proc. Des. Dev. 1986, 25, 654-660.
Mc Avoy, T. Interaction Analysis: Principles and Applications; Inst. Soc. of

America: North Carolina. 1983.
Pomerleau, D.; Pomerleau, A. Guide lines for the tuning and evaluation of
decentralized and decoupling controllers for processes with recirculation. ISA
Trans. 2001,40, 341-351.
Shen, S. H.; Yu, C. C. Use of relay-feedback test for automatic tuning of
multivariable systems. AIChE Journal 1994, 40, 627-646.
Skogestad, S.; Morari, M. Robust performance of decentralized control systems
by independent designs. Automatica 1989, 25, 119-125.
Wang, Y.-G.; Cai, W.-J. Advanced Proportional-Integral-Derivative tuning for
integrating and unstable Processes with gain and phase margin specifications. Ind.
Eng. Chem. Res. 2002, 41, 2910-2914.

84
Decoupling Control System Design
*
Chapter 5
5.1. Introduction
For MIMO processes with severe loop interactions, the decoupling control
schemes are often preferred. The decoupling control usually requires two steps: 1)
design of decoupler to minimize the interactions among loops; and 2) design the
main loop controllers for overall system performance. The key advantages of the
decoupling control approach are that: 1) it allows the use of SISO controller
design methods; and 2) in case of actuator or sensor failure, it is relatively easy to
stabilize the loop manually, since only one loop is directly affected by the failure.
However, such a design may result in very complicated control structures,
especially, when the system dimension is high. Consequently, the research
activities are primarily focused on TITO processes (Shinskey, 1996; Gawthrop.
1987; Palmor et al, 1995; Ho et al, 2000; Wang et al, 2000; Gilbert et al, 2003;
Chen and Seborg, 2002; Tavakoli et al, 2006). For high dimensional processes,
Wang et al. developed some methods for full dimensional PID or non-PID
controllers with systematic design procedures and relatively simple system
structures (Wang et ai, 1997; Wang and Yang, 2002; Wang et al, 2003).
However, due to the mathematic approximation of the decouplers for high
dimensional process, the robustness of such a control system can not be
The main part of this chapter was published in Journal of Process Control 17 (2007) 665-673.

40
guaranteed. Another approach is the static decoupling which can significantly

improve control performances with their robust stability comparable to that of
multi-loop control systems are adopted by the industry to deal with severe loop
interactions for high dimensional MIMO processes (Lee et al, 2005). Although
static decoupling is simple to design and implement, they may not always provide
satisfactory closed-loop performances if there exists large difference in dynamic
characteristics among the transfer function elements.
In this chapter, a novel decoupling design method for multivariable processes
is proposed. The method include three steps: 1) using the concepts of energy
transmission ratio to obtain the effective relative gain, relative gain and relative
frequency of a given transfer function matrix; 2) using the information obtained in
the first step to obtain an equivalent transfer function matrix for closed-loop
system; and 3) designing the off-diagonal controllers based on interaction analysis
and the diagonal controllers for original transfer functions of main loops.
Although this method is basically a full dimensional controller design, it is simple,
straightforward and easy to be applied by field control engineers. The resulting
controller can have satisfactory performance even when the primary process
parameters change significantly. Several multivariable industrial processes with
different interaction characteristics are employed to demonstrate the simplicity
and robustness of the design.
5.2. General formulation of MIMO control
Consider an open-loop stable multivariable system as shown in Figure 2.2.

Again, a SOPDT model is adopted to approximate each element of the process
transfer function matrix, i.e.,
a
2.,jS +aVJS + i
For a closed-loop controlled MIMO system, it is desirable that the forward

transfer function to be of the form
G(s)Gc(s)*G(s)G-](s) = I/s (5.1)
It is, however, not an easy task to find a suitable solution for G~\s) under the
fixed control structures, as it may end up with very complicated expressions,

41
especially, when the system dimension is high. In the following, technique for
computing loop dynamic interactions are employed to find a simple description
for G\s), and establish the relationship between G \s) and the closed-loop
transfer function matrix.
When a MIMO control system is closed, there exist interactions among loops
as a result of the existence of non-zero off-diagonal elements in the transfer
function matrix. The interactions can be dynamically measured by the dynamic
relative gain defined by (Bristol, 1966):
W = gu{s)/giJ(s) (5.2)
where g,;(s) is the equivalent transfer function of g,;(s) when all other loops are
closed. For overall system, Equation (5.2) can be written in a matrix form which
results in the dynamic relative gain array (DRGA),
4,(5) \2(s) ... 4.(5)"
4,(5) 4,(5) ... 4„(5)
A( 5 ) =
4,(5) 4,(5) ... 4,(5)
Substituting Equation (5.2) into (5.3) results in

gn(s)/gn(s) gn(s)/gl2(s) ... glm(s)/gln(s)
A{ v g^)/g2l(s) g22(s)/g22(s) ... g2„(s)/g2n(s)
_gm(s)/g„i(s) gn2(s)/gn2(s) ... gnn(s)/gnn(s)
g\\(s) gn(s) - gu(s) !/£„(*) i/ftaw ... y&M

gll(S) #32 0 ) - &„(*) l/g2l(S) l/g 22 (5) ... l/g 2 „(5)
_gnl(S) gnl(S) - gnn(S).

y^(S) ygnl^) - ygnn(s)
(5.4)
where the operator ® is the hadamard product. Since g (s) is controller
dependant, it is impossible to compute A (5) without first knowing the controller
parameters. By assuming the process is under perfect control, however, a simple
computational algorithm for A (5) can be obtained to calculate the relative gains
at each frequency point (Grosdidier and Morari, 1985):

A(s) = G(s)®G-r(s)
gixis) gn(s) ... glH(s)
g2l(s) g21(s) ... g2n(s) (5.5)
S21OO £22^) - &„(*)
_gM gAS) •• g„n(S) gJS) grl(S) gnnW
Comparing Equations (5.4) and (5.5), we have
gn(s) gn(s) gin(s)

gn(s) g22(s)
G" r (s) =
gln(s)
_£„i(*) gM
gmW
which implies
Vine) Vs2i<*) - V£„i<*>'

l/il2(*) 1/^22^) - Vin2^> (5.6)
As a stable and well designed decoupling control system asymptotically

approaches the perfect control, Equation (5.6) provides the fundamental
relationship between open and closed-loop transfer functions. Consequently, the
problem of designing PI/PID controller for MIMO processes is converted to:
(1) Find the best loop pairings, place them on the diagonal position, and obtain
the equivalent transfer function g (s) by taking loop interactions into
account. This topic has been discussed in Chapter 3 and Chapter 4.

(2) Determine the structure and parameters of gCIJ(s) such that the
interactions among loops are minimized and the desired control system
performance is obtained. This will be discussed in the following sections.
5.3. Full decoupling control
By full dimensional controller structure, Equations (2.1), (2.3), (5.1) and (5.6),
we can easily establish the relation between the controller elements and the ETFs
as

Decoupling Control System Design <r>
&.ll(*) gc,dS) •• gc,JS)

&.a(*) gc,AS) • • & , 2 » 1
GM- (5.7)
5
&,,» gcnlW • • gc.Js)

Consequently, the problem of design the PI/PID controller for M1MO processes is
converted to find Gc(s) such that Equation (5.7) hold. This can be divided into
two sub-tasks:
(1) Off-diagonal controllers: The main task of the off-diagonal controllers is
to minimize the interactions among loops. According to Equation (5.7),
we have
g,j(s)
Image that gcy-,Cs)g (s) is the forward transfer function of an fictitious
closed-loop control system, the control objective, gc. j,(s)g,j(s) - Vs > ' s to
obtain an ideal control for this loop. As the ideal control cannot be realized
in practice, the PI/PID controller structure, such as, the gain and phase
margin approach described in Chapter 4 can be used to design the
controllers.
(2) Diagonal controllers: The diagonal controllers are to provide the desired
performance of the closed-loop control system. As the interactions among
the loops have already been dealt with by the off-diagonal controllers, the
original transfer functions of diagonal elements after loop pairing are used
for the controller design. Again, the gain and phase margin approach in
Chapter 4 is one of the options to be used for designing the controllers.
Remark 5.1. Theoretically, any SISO controller design approach can be
employed. This is because the interaction is already approximately considered into
the equivalent transfer functions. The gain and phase margins approach is selected
because it provides good performance in terms of robustness with respect to the
uncertainties in both process model and disturbance, and might be more
acceptable by process control engineers.
It is worth noting that the only difference between off-diagonal and diagonal
controllers is the proportional gain. For off-diagonal proportional gains, they are
corrected by both relative gain and relative frequency factors, while the original

94
parameters are used for diagonal ones. In addition, it should be noticed that the
control performance may not be met well for some high dimensional systems
since a fixed PI/PID structure is adopted.
In summary, the whole design procedure is very simple, straightforward and
can be fully automated. The general procedure for designing full dimensional
PI/PID controller for MIMO processes can be summarized as follows.
Algorithm 5.1:
Step 1. For a given transfer function matrix, calculate E, O, A and T by
Equations (3.3), (3.5), (3.1) and (4.6), respectively.
Step 2. Select the best pairing and put them on the diagonal position
according to REA, determine the equivalent transfer functions for
off-diagonal elements by calculating the two correcting factors using
Equations (4.3) and (4.9), respectively.
Step 3. Determine the parameters of the off-diagonal decoupling controllers
and main loop diagonal controllers by Equation (4.10) with modified
and original process parameters, respectively.
5.4. Block decoupling control
The block decoupling controllers balance the high performance given by fully
multivariable controllers and the easy implementation and maintenance associated
with fully decentralized controllers. The use of block decoupling control is also
because of the fact that in most industrial processes, the interactions are limited in
scope and do not include the full scope of the plant.
For block decoupling controller structure, Equations (2.1) and (2.4), to identify
the structure of Gc(s), a threshold, p, is to be specified by the designer. If
|&,|> A k,l = 1,2,...n, k*l,
the k'h and /* loops are considered to be closely coupled under the p criterion,
even if | ^ f t | < p . Once the structure of Gc(s) is determined, rearrange the
columns and rows of the transfer function matrix such that G(s) can be
partitioned as

G,(5)
G2(s) ...
G(s) =
- * Gm(s)
where G,(s), I = l...m are sub-matrices of closed-loop ETFs with the same
dimensions as Gcl(s). The block decoupling control starts from the 2*2 controller
sub-blocks, let
Gck (s) -
_&c,p~\p{s> Sc.p+\p+\ys) _
where p +1 < n. Correspondingly
Gk(s) =
gP+iP(s) (-0
and
Ypp Tpp+\
®k =
rp+\p rp+\p+\
Then, for perfect control of the sub-block Gk(s), it requires
Gck(s)*Gk(s).
The overall design of block decoupling control is summarized as follows.
Algorithm 5.2:
Step 1. Design decouplers for those 2*2 sub-blocks according to design
procedures for off-diagonal decoupling controllers using Equation
(5.8).
Step 2. Recalculate <f> and $P+iP+\. As the off-diagonal elements have been
dealt with by decoupling controllers, (f>pp and <f>p^p+l are to be
updated by
Ypp • Ypp Ypp+\

(5.9)
Tp+\p+\ • Yp+\p+\ 'P+lp
to result in a less conservative decentralized control design.

Step 3. Approximate the ETFs of the diagonal elements using Equation (4.7)
in the sub-block Gk(s) by the new <f)pp and flp+\p+\ obtained in
Equation (5.9).
Step 4. Design the decentralized controller for the diagonal elements
following the principles given in Chapter 4.
Remark 5.2. In decoupling control system design. Equation (4.7) is directly used
to find the ETFs for off-diagonal elements, gtJ(s), i * j , using XtJ and y:j, since
the purpose of the off-diagonal ETFs is to design decoupling controllers to

improve the system performance while control system integrity is not our main
concern here. The no-causal elements, e ,J , which may appear in ETFs for some
i, j , can be approximated by transfer function zeros using first order Taylor series
in the decoupling controller design. Furthermore, since the off-diagonal
decoupling controllers are PID type instead of simply the ratio of transfer function
elements as in conventional decoupling controller design, the robustness
improvement of the control system can be expected.
5.5. Case studies
In this section, we apply the proposed decoupling design methodology to a

number of industrial processes. All controllers are designed by gain and phase
margin method given in Chapter 4 and gain and phase margins are assigned to be
5dB and In/Srad , respectively. They are compared with the following
decoupling control approaches to show the effectiveness of the proposed designs:
(1) The Gershgorin bands based decoupling approach proposed by Chen and
Soberg (2002);
(2) The integral decoupling approach proposed by Lee et al. (2005).
The effectiveness of the designed full decoupling and block decoupling
control schemes are measured by:
(1) Column dominance index (Chen and Seborg, 2003):
dciw) = ,~ -lJ'J. ,y = l,2,..77, (5.10)

\GAjo))\

<;7
where G(ja>) is the forward transfer function matrix of the MIMO process
(original or controlled), dCJ{co) = 1 implies ideal decoupling (diagonal).
(2) Robustness index: let the process input uncertainty be A, which is used to
describe modeling errors, then, the closed-loop system is stable if and only
if (Morari and Zafiriou, 1989)
1
a[H(jo))] r,Vfl>,
a[A(jco)]'
where <r[»] is the maximum singular value of • and H is the
complementary sensitivity function of the closed-loop system. A design

method is more robust if the robustness index is smaller within interested
frequency range.
Example 5.1. Consider an industrial-scale polymerization reactor (Chien et ai,
1999)
22.89 _o -11.64
-e 2v
4.5725 + 1 1.8075 + 1*
G(s) =
4.689 5.80
.2.174s + l 1.8015 + 1
For this process, the calculation of open-loop transfer function matrix gives
184.3873 -48.7575 0.7087 0.2913 8.0554 4.1888

E= A = Q
36.8273 24.9606 0.2913 0.7087 7.8540 4.3036
0.7193 0.2807 1.0151 0.9633

O and r -
0.2807 0.7193 0.9633 1.0151
Both REA and RGA pairing rules indicate diagonal pairing (NI =1.4111>0).
Full Decoupling Control: By the proposed method, the full dimensional
decoupling controller is
1 1
0.3137(1 + ) 0.1336(1 +
Gc(s) = 4.57205 2.17405
-0.0224(1 +
1
) 0.2439(1 + _J
1.80705 1.80105
The designed decoupling controllers are evaluated by
(1) The column dominance index of the decoupled and the original processes
are calculated from Equation (5.10) and shown in Figure 5.1, which
indicates that the original process is indeed not diagonally dominant; it is

98
actually off-diagonally dominant since dc2(a>) equates to -1 for entire

frequency range, while the degree of diagonal dominance has been
substantially increased by the proposed method.
frequency (racVsecond)
Figure 5.1. Column dominance for Example 5.1

(2) The closed-loop system responses for the proposed REA based full
decoupling controller (REA-FDC) are shown in Figure 5.2, where the unit
set-points change in r, at r=0 and r2 at f=40, load disturbance injected at

/=80. respectively. For comparison, the results of the decoupling control
given by Chen and Soberg (2002) are also shown in the same figure. The
proposed design method gives better overall performance.

I 5
- REAFDC
— Chen and Soberg
r
C5
10 2 0 3 0 4 0 5 0 6 0 7 0 8 0
Time(hour)
12
1 •
l*>
C8
06
0.4
C2
Q X RFJVFDC
J~
Chen and Soberg
•C2
XI 30 40 50 B0 70 EC
Time(hour)
(3) The value cr[//(yVij)] of different design approaches are shown in Figure
5.3, the proposed method gives better robustness, and the maximum
singular value increases slowly with the frequency.
12
REArFDC
/ Chen and Soberg
2 - — — '•
01 02 03 04 05 06 0.7 09 1
frequency (rad/second)
Figure 5.3. Singular value based robustness index for Example 5.1
(4) The robustness of the different methods is further investigated by
increasing all four time delays by 70%. The closed-loop responses for the
two control schemes are shown in Figure 5.4. It is obvious that the
performance of the proposed control method is still satisfactory while the

10(1
other one becomes oscillatory.
RE/VFDC
Chen and Soberg
I'hWr
iW'JWvV' .'
_l l_
10 20 30 40 50 60 70 80
Time(hour)

Example 5.2. Consider a 3x3 process given by (Ogunnaike and Ray, 1979)
0.66?" -0.6 \e"5s -0.0049e"

6.75 + 1 8.645 +1 9.065 +1
1.1 le^ 5 ' -2.36g- 3 ' -0.0 lg-' 2 '
G(5) =
3.255 + 1 55 + 1 7.095 +1
-34.6fc>-921 46.2e"94s 0.87(11.6I5 + I K
8.155 + 1 10.95 + 1 (3.895 + 1X18.85 + 1)
The calculation of open-loop transfer function matrix results

0.0985 -0.0706 -0.0005 2.0084 -0.7220 -0.2864
E = 0.3415 -0.4720 -0.0014 A = -0.6460 1.8246 -0.1786
-4.2552 4.2385 0.0701 -0.3624 -0.1026 1.4650
0.1493 0.1157 0.1104 2.4267 -1.1510 -0.2758

Q = 0.3077 0.2000 0.1410 <D -0.8244 1.9746 -0.1502
0.1227 0.0917 0.0805 -0.6023 0.1764 1.4259

1.2083 1.5942 0.9628

r 1.2762 1.0822 0.8409
and
1.6621 -1.7197 0.9733
The best pairing according to REA and NI (=0.3859>0) is 1-1/2-2/3-3.

Block Decoupling Control: Since ^l2 = -1.1510 and <f>lx = -0.8244 are large in
amplitudes, block decoupling control scheme is preferred. The proposed REA

based block decoupling controller (REA-BDC) can be easily obtained as
1
1.5911(1 + ) -0.0716(1 + ) 0
6.70005 3.25005
1
Gc(s) = 0.5758(1 + ) -0.2679(1 + -) 0
8.64005 5.00005
0 0 4.4831(1 )
12.41505
The closed-loop responses are shown in Figure 5.5, where the unit set-points
change in r, at t=0, r, at t=500 and r3 t=1000. The responses of full decoupling
control given by Lee et al. (2005) are also shown in the same figure. It is clear that
the proposed method has a better performance.

1(12
1.2
1 K
ca f
C6 -
ci
02
0 — REA-BDC '
Lee et al
-0 2
500 1000 I SOD
I
Time(minute)
1.5
REA-BDC
Lee et al
0J
L
Q5
500 10O0 '500

Time(minute)
REA-BDC
Lee et al
500 1000 1500

Time(minute)

To investigate the robustness of the system, we mismatch the process model
by increasing steady-state gain for loop 1, time delay for loop 2, and time
constants for loop 3 all by 40%. The closed-loop responses under same controllers
are shown in Figure 5.6. In this case, the proposed method still gives satisfactory
performance while the other method is oscillatory even it has satisfactory nominal
performance under exact model. In fact, further investigation shows that the
proposed controller can even work for 100% model parameter mismatches while
the other one is unstable.

REArBDC
Lee et ai
500 1000 •500

Time(minute)
1 2
1
0.8
0.6
C J
02
-C2
.4 RBVBDC
Lee et at.
0.6
500 1000 1£OC
Time(minute)
500 1000
Time(minute)

Example 5.3. Consider a 4x4 process given by (Doukas and Luyben, 1978)
-9.81 U?-"9' 0.374e~ -2.368e> -11.3e"

11.36^ + 1 22.225 + 1 33.35 + 1 (21.745+ 1)2
72j
5.984g-224' -1.986e-°71j 0.422g-* 5.24e^"
14.29^ + 1 66.675 + 1 (2505+ 1)2 4005 + 1
G(s) =
0.0204e^ 9j 0.513e" -0.33e-°68j
(1.435+ 1): (7.145 + 1)2 5+1 (2.385+ 1)2
37
-11.3e- " -0.176e^ 48 ' I5.54e~' 4.48e^ 52 '
(21.745 + 1)" (6.95 + 1)- 5+ 1 11.115 + 1
The calculation of open-loop transfer function matrix results

Hi4
-9.9426 0.0855 -0.1793 -1.7750

4.3719 -4.4566 0.0129 0.1463
E=
4.2726 0.0139 1.0073 -0.3575
-1.7531 -0.1382 31.4970 14.0743
0.1264 -0.1013 -0.0314 1.0063^

0.0107 1.0935 0.0003 -0.1045
A =
0.7264 0.0025 0.1630 0.1081
0.1366 0.0054 0.8680 -0.0099
"1.0134 0.2285 0.0757 0.1571

0.7306 2.2440 0.0306 0.0279
Q =
1.7952 0.6830 1.9635 1.0833
0.1551 0.7854 2.0268 3.1416
0.5237 -0.0052 -0.0222 0.5036

-0.0024 1.0035 0.0000 -0.0012
<D =
0.4719 0.0012 0.2965 0.2303
0.0067 0.0004 0.7256 0.2673
4.1448 0.0511 0.7066 0.5004

-0.2239 0.9177 0.0978 0.0113
and r= 0.6497 0.4888 1.8192 2.1309
0.0492 0.0793 0.8359 -26.9057
The best pairing according to REA and NI (=1.1814>0, ideal pairing) is 1-4/2-2/3-
1/4-3.
Block Decoupling Control: Since the closely coupled loops are 1 and 3,
4 3 =0.5237 and ^ =0.2673, block decoupling control scheme is selected. By
proposed method, the REA-BDC is designed as
1 _ 1
-03983(1- +1Q87M?) 0 -Q338Q(l4 +1.190Os) 0
43.4800? 47600s
1
0 -148540(1+ 0 0
65.6700*
m- 1 1
-00070(1+ ) 0 08913(1+ +07150?) 0
113600s 28600?
0 0 Q02O2(l-

105
at t=0, r2 at t=200, r3 t=400 and r4 t=600. Since we cannot find any existing
decoupling control scheme in literatures for this process, we compare the
responses with the REA based decentralized controller (REA-DCC) proposed in
Chapter 4, which are shown in the same figure. As can be seen, the REA-BDC
method has better overall performances.

Decoupling Control System Design 1(16
1 2
OB
06
04
02
0.2
-04 REA-BDC
REA-DCC
-06
100 200 300 400 500 600 700 500
Time(minute)
REA-BDC
1.4 REA-DCC "
i:
i
•f '
^v~
r
0.8 -
06 • -
0.4 •
02 •
0
' L • ' '
0 100 2 0 0 3 0 0 4 0 0 5 0 0 6 0 0 700 800
Time(minute)
1 4
REA-BDC
1.2 - REA-DCC -
1 It-
y-
06 -
06
0.4
02 •
-0.2 1 1 1
0 100 2 0 0 3 0 0 4 0 0 5 0 0 6 X 7 0 0 800
Time(minute)
y--
•f- -f-*-
REA-BDC
REA-DCC
100 200 300 400 500 532 800

Time(minute)

11)7
To investigate the robustness of the system, we mismatch the process model

by increasing all steady-state gains by 50% and all time delays by 20%. The
closed-loop responses under same controllers are shown in Figure 5.8. In this case,
the REA-BDC still gives satisfactory performance while the REA-DCC is more
oscillatory.

0 100 200 300 400 500 600 700 800

Time(mmute)
REA-BDC
REA-DCC
1 5
^*W*«- ^fr^*r>»
05
-05
1 DO 200 300 400 500 600 700 BOO
Time(minute)
100 2 0 0 3 0 0 4 0 0 5 0 0 6 0 0 7 0 0 8 0 0
Time(minute)
1 X 2 0 0 3 0 0 4 0 0 5 0 0 6 0 0 7 0 0 8 0 0
Time(minute)

MW
5.6. Conclusions
In this chapter, a novel full dimensional PI/PID controller for MIMO

processes is proposed which consists of diagonal main controllers and off-
diagonal decoupling controllers. Based on the equivalent transfer function matrix,
both off-diagonal decoupling controllers and main loop diagonal controllers can
be easily designed using the existing PI/PID tuning rules. The design procedure is
very simple and straightforward, which can be easily understood and implemented
by field control engineers. The simulation results for a number of industrial
processes show that the proposed method gives better or comparable nominal
system performance to other reported techniques. Furthermore, due to the
implementation of PI/PID controllers for decoupling loops, the proposed method
has good robustness and still works with satisfactory performance even if the
primary process parameters have significant variations while other designs
become oscillatory or unstable.
References
Bristol, E. H. On a new measure of interactions for multivariable process control.

Chen, D.; Seborg, D. E. PI/PID Controller Design Based on Direct Synthesis and
Disturbance Rejection. Ind. Eng. Chem. Res. 2002, 41, 4807-4822.
controllers with no proportional kick. Ind. Eng. Chem. Res. 1999. 38, 1456-1468.
Doukas, N.; Luyben, W. L. Control of sidestream columns separating ternary
mixtures. Instrum. Technol. 1978, 25, 43.
Gawthrop, P. J. Continuous-time self-tuning control, 1, design. Research Studies

Press: UK, 1987.

Decoupling Control System Design no
Gilbert, A. F.; Yousef, A.; Natarajan, K.; Deighton, S. Tuning of PI controllers

with one-way decoupling in 2><2 MIMO systems based on finite frequency
response data./ Proc. Cont. 2003, 13, 553-567.

Ho, W. K.; Lee, T. H.; Xu, W.; Zhou, J. R; Tay, E. B. The direct Nyquist array
design of PID controllers. IEEE Trans. /£2000, 47, 175-185.
Lee, J.; Kim, D. H.; Edgar, T. F. Static decouplers for control of multivariable
processes. AIChE J. 2005. 51, 2712-2720.
Morari, M.; Zafiriou, E. Robust Process Control; Prentice-Hall: Englewood Cliffs,
NJ, 1989.
Ogunnaike, B. A.; Ray, W. H. Multivariable controller design for linear systems
having multiple time delays. AIChE J. 1979,25, 1043.
Palmor, Z. J.; Halevi, Y.; Krasney, N. Automatic tuning of decentralized PID

controllers for TITO processes. Automatica 1995, 31, 1001-1010.
Shinskey, F. G. Process control systems: Application, design and tuning;
McGraw-Hill: New York, 1996.
Tavakoli, S.; Griffin, I.; Fleming, P. J. Tuning of decentralised PI (PID)

controllers for TITO processes. Control Eng. Practice 2006, 14, 1069-1080.
Wang, Q. -G.; Huang, B.; Guo, X. Auto-tuning of TITO decoupling controllers

from step tests. ISA Trans. 2000, 39, 407^118.
Wang, Q. -G.; Yang, Y. -S. Transfer function matrix approach to decoupling

problem with stability. Systems & Control Letters 2002, 47, 103-110.
Wang, Q.-G.; Zhang, Y; Chiu, M. S. Non-interacting control design for
multivariable industrial processes. J. Proc. Cont. 2003, 13. 253-265.
Wang, Q. -G.; Zou, B.; Lee, T. H.; Bi, Q. Auto-tuning of multi-variable PID
controllers from decentralized relay feedback. Automatica 1997, 33. 319-330.


Chapter 6
Application to an Experimental HVAC

System
6.1. Introduction
The HVAC has now become an indispensable part of our life. Different kinds
of HVAC systems have been installed in housing estates, commercial complexes,
office buildings, and industrial estates, etc. These systems are playing an
important role in providing thermal conditions for people and appliances inside
the air-conditioned space. With the wide usage of air-conditioning system, the
electricity usage for space air-conditioning amounts up to 30% of the total
electrical energy consumption for countries with mild climate (Yaqub and Zubair,
2001; Shoureshi, 1993). As Singapore is a tropical city country with an annual
average ambient temperature of 29.4 °C and an annual average ambient relative
humidity of 85%, the indoor thermal condition of buildings is very important
since people live and work in buildings with air-conditioning systems running
almost all the time. Therefore, it is very important to increase the system
efficiency and improve interior environmental control in a more cost-effective,
high quality manner. The research in the area of control of air-conditioning system
has, therefore, a great impact to the economy and environment.
The main part of this chapter is submitted to Control Engineering Practice.

113
Application to an Experimental HVAC System
As shown in Figure 6.1, an HVAC system consists of indoor air loop, chilled
water loop, refrigerant loop, condenser water loop, and outdoor air loop. Each
component involves several variables, the working principle and system analysis
can be very complicated. To minimize the energy consumption while meeting the
human comfort requirement, an accurate mathematical model for each component
is needed to describe its working conditions. Based on the models, optimization
problem for buildings can be formulated as minimization of the energy cost with
the constraints that all the requirements on air quality in the building such as room
temperature, humidity, air movement and oxygen concentration be met. While for
every set-point change determined by the optimization algorithm, the control
system is to make sure that each device can reach its new set-point smoothly and
promptly.
water «- tower «-
X pumps fans
W
V
< i
•
nvironm
outdoo
a
1a cooling
B
u towers
,1 /DUUJOOf 0.'
Watey
—•
chillers
Figure 6.1. Scheme of a typical building HVAC system

HVAC systems require control of environmental variables such as pressure,
temperature, humidity, etc. Over the years, different control strategies have been
developed to control fully or partially the HVAC process operation (Teeter and
Chow, 1998; Liu and Dexter, 2001; Sousa et al., 1997; Ghiaus, 2001). However,
as in other process control applications, classical control techniques, such as on-
off switching controller (thermostats) and PID controllers are still widely used in
practice due to easier implementation, low cost and reliable performance etc
(Seem, 1998). However, the PID tuning procedure can be a time-consuming,
expensive and difficult task (Pinnella et al., 1986; Krakow and Lin, 1995). Some
control engineers may take 1 to 3 days to search for a workable PID controller
setting for an air pressure loop in a building. For a temperature loop, the time
spent on PID parameter tuning may be even longer. For many processes, their

Application to an Experimental HVAC System nt
inputs and outputs are cross-coupled, i.e., one process input can influence other
outputs of the process. For example, if the chilled water supply is used to regulate
the air humidity and the room damper controls the room temperature, then
changing the valve position of the chilled water supply will affect both humidity
and room temperature. Also, changing one room damper to regulate the room
temperature will also affect the temperature of another room. For these kinds of
multivariable processes, if independent loop control is adopted without taking into
account the coupling, the resulting performance is usually poor. To enhance
performance, the control design should address the coupling issue, which makes
loop tuning more difficult. As a result, many HVAC loops are poorly tuned
(Pinnella et al, 1986). A poorly tuned control system may waste energy, reduce
occupant comfort, and cause excessive and unnecessary wear of actuators.
The Ziegler and Nichols methods (Ziegler and Nichols, 1942) are classical
PID tuning approaches which are widely accepted in industries. But those
methods have shortcomings such as long testing time and limited control
performance. It is more suitable for the initial setting of a PID controller. Recently,
some complex control strategies based on the classical control concepts have been
proposed in attempting to improve the system performance. Among those works,
Pinnella et al. (1986) proposed a self-tuning method of an integral only controller.
Brandt (1986) applied the self-tuning control in an HVAC system. Nesler (1986)
developed a computer-based auto- and self-tuning controller, where the open-loop
step test of the Ziegler and Nichols method was used to derive a process model
and a PID controller. Krakow and Lin (1995) designed their PI controller based on
the analytical models of HVAC components. Salsbury (1998) described a
feedforward controller as a supplement of the conventional PI feedback controller
based on simplified physical models. Kasahara et al. (1999) proposed a robust
PID control scheme to deal with the model uncertainty caused by changing
characteristics of the plant. Bi et al. (2000) developed an advanced auto-tuning
PID controller for both temperature and pressure control loops. However, the
above literatures are mainly focused on the single loop PID design, due to the
interactions among individual loops many controllers are tuned conservatively in
order to make the overall system stable. This may result in a unsatisfactory
performance and waste of energy. In this chapter, the REA based MIMO
techniques are applied to analysis and control design of an experimental HVAC

Application to an Experimental HVAC System ill
system, which attempts to replace independent PID controller tuning/re-tuning

with systematic MIMO system design.
6.2. Four-room air temperature control system
The four-room air temperature control application for HVAC system is tested
on an experimental centralized HVAC system. The system has three chillers, three
sections with three air handling units (AHUs), three cooling towers and flexible
partitions up to twelve rooms. All motors (fans, pumps and compressors) are
equipped with variable speed drives (VSD). The system is made very flexible to
configure these three units to form different schemes. The portrait of the
experimental system is shown in Figure 6.2, where number 1, 2, 3 and 4 indicate
the components of computer controller, VAV air-conditioning system, signal
process board (AI/AO) and signal transmission cables, respectively. The software
environment is Matlab/Simulink V6.5. In the system, the outside air is mixed with
the building return air. The mixed air, i.e., supply air is sucked through the cooling
coil by a supply fan. The cooled air is then supplied to different zones. The
temperature of the off-coil air can be maintained at a pre-specified value by the
regulating valve of chilled water supply or return. To ensure each control loop to
work properly, the supply air pressure needs to be kept at a constant value. This is
realized by regulating the supply air fan speed. The temperature of each room can
be fine tuned by the corresponding damper of VAV box. All motor-driven fans,
pumps and compressors in the system are controlled by VSDs.

* ft ft C ttes
1 , c
Q
I * pj
W . .. i» a . — J * "
i* 2 I B 11
|4 1 "**
Figure 6.2. Experimental centralized HVAC system
Here we consider the temperature control problem of four neighboring rooms

of one zone. The schematic diagram of the four-room multivariable system is
shown as in Figure 6.3.
In this system, four temperatures are controlled by regulating the position of

four variable air volume (VAV) dampers. When the VAV damper opens wider,
more cooling air enters the room and the room temperature will drop, and vice
versa.

117
Room 2 Room 4
VT„ Dit 2 VT.T DM

ufi' m OK pq
* t
Supply Air *• (, p *
* Vd |S
* *
ED- hd fT3^ DM r
Room 1 Room 3
Figure 6.3. Four-room HVAC system (Top View)
In the system, there are one air pressure sensor, PA, in the duct, and four
temperature sensors, Tm , TR2, TR3 and TR4, located in Rooms 1, 2, 3 and 4,
respectively. In addition, there are four supply air control dampers, DMX, DM2,
DM3 and DMA, and four exhausted air vents, VTX, VT2, VT3 and VT4 on the
room ceilings of Rooms 1, 2, 3 and 4, respectively. Due to the cross coupling,
change of any damper position will cause temperature offset for all four rooms.
Therefore, it forms an interactive 4x4 process. The main control objective is to
maintain the conditioned space temperature of each room at the set-point value.
The manipulating inputs for the system are the positions of the four dampers,
DMX, DM2, DM3 and DM,.
6.3. Modeling and simulations
The dynamics of the temperature control of the four-room process is modeled

by using closed-loop identification method (Mei et al, 2005), the parameters of
the transfer function matrix are obtained as:

IIS
-0.098 -0.036 -0.014 -0.017 _m

-e
1225 + t' 1495 + 1* 1585 + T 1555 + 1
r, -0.043 -0.092 -0.011 -0.012
T2 1475 + 1* 1305 + l' 1565 + 1' 1575 + 1*
T3 -0.012 -31J -0.016 -34s -0.102 -0.033
T4 1535 + 1* 15l5 + l' II85 + 1' 1465 + 1*
-0.013 -32s -0.015 -31 J -0.029 .-255 -0.108
1565 + l' 1595 + l' 1445 + 1' 1285 + 1*
where w,, «,, »3and « 4 , the positions of four dampers, are manipulating variables,
and T,, T2, T3 and T4, temperatures of four rooms, are controlled variables,
respectively.
The calculation of the above open-loop transfer function matrix results in

"-0.0092 -0.0023 -0.0007 -0.0009"
-0.0108 -0.0095 -0.0006 -0.0006
E=
-0.0007 -0.0008 -0.0103 -0.0021
-0.0007 -0.0008 -0.0091 -0.0100
1.2207 -0.2051 -0.0053 -0.0103

-0.1947 1.2198 -0.0136 -0.0116
A=
-0.0106 -0.0085 1.1095 -0.0904
-0.0154 -0.0062 -0.0907 1.1124
1.3874 -0.3899 0.0028 -0.0003

-0.3804 1.3859 -0.0049 -0.0006
0 =
0.0018 -0.0026 1.2251 -0.2242
-0.0088 0.0066 -0.2229 1.2251
1.1366 1.9008 -0.5321 0.0329

1.9541 1.1362 0.3641 0.0508
and
-0.1673 0.3111 1.1041 2.4787
0.5732 -1.0673 2.4577 1.1014
The best pairing according to both RGA and REA and NI (=0.7267>0) are
diagonal. In order to find the proper control strategy, several options are
considered below, where the gain and phase margins are specified as 5db and
2nl5rad, respectively.

ii')
Decentralized Control: For i = 1,2,3,4 , A„>1 and y„>\ , by the method

proposed in Chapter 4, the equivalent diagonal transfer functions are calculated as
-0.098 -0.092 -18 1786i -0.102 -0.108
and
1225 + 1 1305 + 1 1185 + 1 1285 + 1
respectively. The resulting REA-DCC is listed in Table 6.1 together with the
controllers designed by RGA base approach (Chien et ai, 1999) for comparison.
Table 6.1. Decentralized controllers for HVAC system
Controller REA-DCC R(iA based
K, T
,,» *,.
T
u<
Loop 1 -20.2407 122.0000 -32.2705 113.1520
Loop 2 -24.4200 130.0000 -38.9186 106.4960
Loop 3 -20.5729 118.0000 -31.8627 106.4960
Loop 4 -18.7813 128.0000 -29.0158 119.8080
Decoupling Control: Since the closely coupled loops are loop 1 and loop 2 with
^ 2 =-0.3899 and #,, =-0.3804, and loop 3 and loop 4 with ^ 4 =-0.2242 and
043 = -0.2229, respectively, the block decoupling control scheme is adopted. By

the method proposed in Chapter 5, the REA-BDC is designed as
1 _ 1 .
-23.7863(1+ ) 4.0161(1+ 0 0
122000O5 149.00005
1 1
5.5372(14 ) -28.9986(14 -) 0 0
147.00005 1300000s
Gt(s)=
1 1
0 0 -23.5699(1+ ) 22449(1+
118.00005 1460000s
1 1
0 0 20010(1+ ) -21.5432(14 )
144.000CK 128.0000s
at t=0, r2 at t=1000, r3 at t=2000 and r4 at t=3000. As can be seen, the REA-DCC

is better than the RGA based method which has some overshoot, the REA-BDC
method has the best overall performances.

Application to an Experimental HVAC System i:u
11
1.2 •
1 Jk_ _,\__
0.8 •
06
0.4
REA-DCC
0 2 REA-BDC
---- RGA based
0 500 1000 1500 2000 2500 3000 3500 4000

Time(second)
1.2
1 -A
S 04
o:
o
k REA-DCC
REA-BDC
RGA based
-02
0
I 500 1000 1500 2000 2500 3000 3500 4000
Time(second)
1.2
REA-DCC
1 . — REA-BDC
---- RGA based
.,, k
08 -
1
I:
02
\ h
0
0 500 1000 1500 2000 2500 3000 3500 4000
-0 2 Time(second)
1.2
REA-DCC
REA-BDC
RGA based
oe
o
I"
s 04
t-
02 K»>_
0
0 500 1000 1500 2000 2500 3000 3500 4000
•0 2 Time(second)
Figure 6.4. Closed-loop responses of four-room system

121
In order to compare robustness of different control schemes, we mismatch the

process model by increasing all 16 steady-state gains, 16 time constants and 16
time delays by 50%. With the controller parameters of both REA-DCC and REA-
BDC kept the same as before, the closed-loop responses are shown in Figure 6.5,
where the unit set-points change in rx at t=0, r2 at t=1000, r3 at t=2000 and r4 at
t=3000. It shows that even under such a severe model mismatch, both the REA-
DCC and the REA-BDC still give satisfactory dynamic performance.

Application to an Experimental HVAC System \n
14
1.2
08
i
0.6 •
-
C4
0.2 — RE/VDCC '

RE>VBDC
0 500 1000 1500 2000 2500 3000 3500 4000

Time(second)
1.2 I I
1 /V- A-
0.8
0.6
C4
C2
•0.2
L _l L_
READCC
REABDC
0 500 1000 1500 2000 2500 3000 3500 4000

Time(second)
1 2
1
REA-OCC
REABDC
JZ
CB
i"
C2
0
_1 L_
0 500 1000 1500 2000 2500 3000 3500 4000
-C2 Time(second)
1.2
READCC
REA-BDC
1—
08
i:
1
1
1
1-
02
0
ft
•02
0 50G 1000 1500 2000 2500 3000 3500 4000
Time(second)
Figure 6.5. Closed-loop responses of four-room system (model mismatched)

123
The above simulations confirm that the block decoupling control scheme is
superior in both nominal performance and robustness, but it is not very significant
compared with the decentralized control scheme. In addition, the uncertainty of
this pilot plant may not be so large. Therefore, considering the simplicity of
implementation and loop failure tolerance issues, it is decided that the REA-DCC
be implemented in the 4-room HVAC system.
6.4. Experiment results
The experiment is conducted under the following conditions: The chilled

water supply temperature is fixed and the cooling load variation is achieved
through the air and water flow rates. In the VAV system, the supply air
temperature is controlled at fixed set-point value and the speed of supply fan
varies according to the time varying cooling load. All those are realized by
separate controllers and a hierarchical two-level multi-model control strategy is
applied to the supply air temperature control (He et ai, 2005).
Figures 6.6 and 6.7 show the performance of the multi-loop control system, in
which the set-points values change from 22 C to 25 C at t=0, t=1000, t=2000
and t=3000 second, respectively. At time t=4000 second, the disturbance is
introduced by opening the door of ROOM 2 with some extent. As can be seen,
both the set-point tracking performance and the disturbance rejection performance
are satisfactory.

0 100 200 300 400 500 600 700 800 900 10C0
Time(x5 second)
0 100 2 0 0 3 0 0 1 0 0 5 0 0 6 0 0 7 0 0 8 0 0 9 0 0 1000
Time(x5 second)
0 100 200 300 400 500 600 700 800 900 1000
Time(x5 second)
0 100 200 300 400 500 600 700 800 900 1000
Time(x5 second)
Figure 6.6. Real-time closed-loop responses of four-room system

125
0 100 2 0 0 3 0 0 4 0 0 5 0 0 6 0 0 7 0 0 8 0 0 9 0 0 1000
Time(x5 second)
0 100 2 0 0 3 0 0 4 0 0 5 0 0 6 0 0 7 0 0 B 0 0 9 0 0 1000
Time(x5 second)
0 100 200 300 400 500 600 700 BOO 900 1000
Time(x5 second)
0 100 2 0 0 3 0 0 4 0 0 5 0 0 6 0 0 7 0 0 8 0 0 9 0 0 1000
Time(x5 second)
Figure 6.7. Real-time manipulated variables of four-room system

126
6.5. Conclusions
In this chapter, the proposed decentralized and decoupling design methods are
used to analyze and design an experimental HVAC system. From simulations and
real-time experiments, it is shown that the proposed control design works well in
the experimental HVAC system even if with significant model uncertainty. This
makes it possible at some extent to replace the tedious and poor manual tuning/re-
tuning of independent loop PID controllers in HVAC systems. Since many
processes in other industries have similar characteristics as those in HVAC
systems, therefore, it is conjectured that the proposed design approach can be
applied to the analysis and design of various industrial multivariable control
systems.
References
Bi, Q.; Cai, W.-J.; Wang, Q.-G.; Hang, C. C. Advanced controller auto-tuning and
its application in HVAC systems. Control Eng. Practice 2000, 8, 633-644.
Brandt, S. G. Adaptive control implementation issues. ASHRAE Trans. 1986, PO-

86-05,211-219.
controllers with no proportional kick. Ind. Eng. Chem. Res. 1999, 38, 1456-1468.
Ghiaus, C. Fuzzy model and control of a fan-coil. Energy and Buildings 2001, 33.
545-551.
He, M.; Cai, W.-J.; Li, S.-Y. Multiple fuzzy model-based temperature predictive
control for HVAC systems. Information Sciences 2005, 169, 155-174.
Kasahara, M.: Matsuba, T.: Kuzuu, Y.; Yamazaki, T.; Hashimoto, Y.; Kamimura,
K.; Kurosu, S. Design and tuning of robust PID controller for HVAC systems.
ASHRAE Trans. 1999, 105, 154-166.
Krakow, K. I.; Lin, S. PI control of fan speed to maintain constant discharge
pressure. ASHRAE Trans.: Research 1995, 101, 398-407.

127
Liu, X. F.; Dexter, A. Fault-tolerant supervisory control of VAV air-conditioning

systems. Energy and Buildings 2001, 33, 379-389.
Mei, H.; Li, S.-Y.; Cai, W.-J.; Xiong, Q. Decentralized closed-loop parameter
identification for multivariable processes from step responses. Mathematics and
Computers in Simulation 2005, 68, 171-192.
Nesler, C. G. Automated controller tuning for HAVC applications. ASHRAE
Trans. 1986, PO-86-05, 189-201.
Pinnella, M. J.; Wechselberger, E.; Hittle, D. C; Pederson, C. O. Self-tuning
digital integral control. ASHRAE Trans. 1986, PO-86-05, 202-210.
Salsbury, T. I. A temperature controller for VAV air-handing units based on

simplified physical models. HVAC&R Research 1998, 3, 264-279.
Seem, J. E. A new pattern recognition adaptive controller with application to

HVAC systems. Automatica 1998, 34, 969-982.
Shoureshi, R. Intelligent control systems: Are they are real. /. Dynamic Syst.,
Measur. Contr. 1993, 115, 15-19.
Sousa, J. M.; Babusla, R.; Verbruggen, H. B. Fuzzy predictive control applied to

an air-conditioning system. Control Eng. Practice 1997, 10, 1395-1406.
Teeter, J.; Chow, M. Y. Application of functional link neural network to HVAC

thermal dynamic system identification. IEEE Trans. Industrial Electronic 1998,
45, 170-175.
Yaqub, M.; Zubair, S. M. Capacity control for refrigeration and airconditioning

systems: A comparative study. Int. J. Energy Res. Tech. 2001, 123, 92-99.
Ziegler, J. G.; Nichols, N. B. Optimum settings for automatic controllers. Trans.
ASME 1942, 64, 759-768.

Conclusion and Recommendations [28
Chapter 7
Conclusion and Recommendations
7.1. Conclusion
In this thesis, in terms of relative interaction energy, a comprehensive

engineering oriented methodology for MIMO control system design is presented
which includes: interaction analysis, loop pairing, design of decentralized, full
decoupling and block decoupling controllers. The design procedure is very simple
and effective, which can be easily understood, and can guide control engineers in
applications. The main works in this thesis include:
(1) A novel interaction measure and loop pairing criterion is proposed which
can accurately reflect dynamic loop interactions under finite bandwidth
control. In this criterion, both steady-state gain and critical frequency of
the process are used to provide a simple yet comprehensive description of
loop interactions for MIMO processes.
(2) A novel design method for decentralized control system based on
equivalent transfer functions is proposed. The loop interactions are
quantified by two elements, i.e., relative gain and relative critical
frequency. The interaction effects for a particular loop from all other
closed-loops are then analyzed through both steady-state gain and critical
frequency variations, and directly incorporated into the equivalent transfer
functions in four ways in a simple and straightforward manner. Hence,

Conclusion and Recommendations 129
different equivalent transfer functions for different interaction modes are

derived such that the control system integrity can be preserved, and the
existing single loop tuning techniques can be employed.
(3) A novel engineering oriented decoupling control design method for
multivariable systems is proposed. An equivalent relationship between
open-loop and closed-loop transfer functions is established. For control
system design, individual loop controllers can then be directly designed
based on only one element in the closed-loop equivalent transfer function
matrix. Under this framework, the full decoupling or block decoupling
controller can be designed under the same formula as required.
The numerical simulations on several multivariable industrial processes with
different interaction characteristics and the experiments on a temperature control
system of an experimental HVAC system demonstrate the simplicity and
effectiveness of the design method.
The significances of the work of the thesis are: 1) the proposed loop pairing and
control system design technique results in a better control system performance
compared with the existing methods; 2) all control designs are based on single
loop control concept and all the existing single loop PI/PID tuning rules can be
directly used; 3) the control system is robust, it still works with satisfactory
performance even under severe model mismatches; and 4) the method is even
more significant when applied to higher dimensional processes with complicated
interaction modes.
7.2. Recommendations
In this thesis, an engineering oriented methodology for MIMO control system

design is proposed. To further develop the theoretic foundation of the method and
make it fully accepted, some possible issues for future research in this area are
briefly described below.
(1) As demonstrated, REA is an effective interaction measure. However, the

further investigation about REA is still attractive. The graphical
interpretation of REA may be one of the most significant aspects, which can

Conclusion and Recommendations no
provide a visual means of summarizing interaction information. Such a

graphical approach can be more convenient and straightforward in control
system analysis and design.
(2) Another interesting issue is the influence of process model uncertainty on
REA analysis. Analytical expressions for REA uncertainty bounds like
worst-case bounds, which concern model accuracy requirements and the
robustness of decentralized control systems, will be useful for control
system design.
(3) Closing the loop around one subsystem may introduce RHP transmission
zeros in other subsystems. If such zeros exist, and they are more severe than
the RHP zeros of the overall system, they will represent performance
limitations specific to the use of decentralized control. Hence, the
investigation on the performance limitation in terms of REA is an active
topic.
(4) Generally the performance for set-point change and load disturbance are
conflict in a simple PID loop. This can be improved by two stage or two
degree of freedom PID control. The inner P controller is designed first.
Then the outer controller is designed. However, the similar two stage
approach for multi-loop system is still challenging. Hence, the REA based
two stage decentralized controller design methodology would be attractive.
(5) Extensive comparison with other MIMO design approaches in terms of time
domain or frequency domain specification indices, and implementation of
decentralized/decoupling control scheme for the overall HVAC pilot plant
(3 sections and 12 rooms) are significant to further verifying the proposed
methods.

131
Author's Publications
(1) Qiang Xiong, Wen-Jian Cai and Mao-Jun He, Equivalent transfer function
method for PI/PID controller design of MIMO processes. Journal of
Process Control 17 (2007) 665-673.
(2) Qiang Xiong and Wen-Jian Cai, Effective transfer function method for
decentralized control system design of multi-input multi-output processes,
Journal of Process Control 16 (2006) 773-784.
(3) Qiang Xiong, Wen-Jian Cai. Mao-Jun He and Ming He, Decentralized
control system design for multivariable processes—A novel method based
on effective relative gain array, Industrial and Engineering Chemistry
Research 45 (2006) 2769-2776.
(4) Qiang Xiong, Wen-Jian Cai and Mao-Jun He, A practical loop pairing
criterion for multivariable processes, Journal of Process Control 15 (2005)
741-747.
(5) Qiang Xiong, Wen-Jian Cai and Ming He, A practical decentralized PID
auto-tuning method for TITO systems under closed-loop control,
International Journal of Innovative Computing, Information and Control 2
(2006) 1-18.
(6) Qiang Xiong and Guang-Yu Jin, Iterative decentralized PID tuning based
on gain and phase margins for TITO systems, International Journal of
Innovative Computing, Information and Control, Accepted in press.
(7) Hua Mei, Shao-Yuan Li, Wen-Jian Cai and Qiang Xiong, Decentralized
closed-loop parameter identification for multivariable processes from step
responses, Mathematics and Computers in Simulation 68 (2005) 171-192.
(8) Shao-Yuan Li, Wen-Jian Cai, Hua Mei and Qiang Xiong, Robust
decentralized parameter identification for two-input two-output process

132
from closed-loop step responses, Control Engineering Practice 13 (2005)

519-53.
(9) Qiang Xiong and Wen-Jian Cai, An engineering interaction measure with
application in vapor compression cycle control. Proceedings of the 9th
International Conference on Control, Automation, Robotics and Vision,
December 5 - 8 , 2006, Singapore.

Xiong Qiang 2007

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Xiong Qiang 2007

Uploaded by

Copyright:

Available Formats

This

Downloaded on 11 Nov 2022 04:04:24 SGT

Multivariate Control System Design

School of Electrical & Electronic Engineering

A thesis submitted to the Nanyang Technological University

First and foremost. I would like to express my sincere gratitude to my

Finally, my deep thanks go to all of my family members for their

Nanyang Technological University

Nanyang Technological University

Table of Contents ill

Nanyang Technological University

Nanyang Technological University

also been implemented in an experimental HVAC system to illustrate its

Nanyang Technological University

Figure 2.1. Hierarchy of control system functions in a typical processing plant

Figure 4.1. Interaction mode with An < 1, yn < 1

Figure 4.2. Interaction mode with An < 1, yu > 1

Figure 4.3. Interaction mode with An > 1. yn < 1

Figure 4.4. Interaction mode with kn > 1, yn > 1

Figure 4.5 Closed-loop responses for Example 4.1

Nanyang Technological University

Figure 6.5. Closed-loop responses of four-room system (model mismatched)

Nanyang Technological University

Table 4.1. Typical gain and phase margin values

Nanyang Technological University

Most industrial operating units are naturally characterized by MIMO

Consequently, multivariable control system design and implementation

Nanyang Technological University

Adjusting controller parameters of one loop affects the performance of others,

In summary, this thesis attempts to develop a practical and engineering

1.3. Major contributions of the thesis

In this thesis, a systematic design approach for decentralized/decoupling

Nanyang Technological University

quantified by relative gain and relative critical frequency. The interactions

(3) A novel full decoupling/block, decoupling design approach based on the

1.4. Organization of the thesis

The rest of thesis is organized as follows:

Chapter 2 gives a review in the area of multivariable control system,

Chapter 3 presents an interaction measure and loop pairing criterion based on

Chapter 4 presents a decentralized controller design approach based on the

Chapter 5 presents a framework for full decoupling and block decoupling

Chapter 6 describes the application of the proposed approach to the room

Nanyang Technological University

temperature control of an experimental HVAC system.

Nanyang Technological University

In modern processes, there are mainly two multi-level hierarchies of control

Nanyang Technological University

h g h ' t o * Select Logi: H g h ' L o * Select _og c

SUM SUM SUM S'JM

© ^) S / fc) $£) <xB)

(a) Conventional structure

Unit 1 Local Optimizer Unit n Local Cotimize* Optimization

Nanyang Technological University

Most industrial operating units are naturally characterized by MIMO

Indeed, decentralized PI/PID controllers either in the simple multi-loop form

(1) Hardware simplicity: The cost of implementing of a decentralized control

Nanyang Technological University

example, only 6 parameters need to be tuned instead of possible 12 for

(3) Robustness: The failure tolerance is usually easier to achieve in

Decentralized and decoupling control design consist of two stages: loop

A multivariate control system with n inputs and n outputs can be typically

function matrix expressed by

JAS) gn2(S) - gm(s)

Nanyang Technological University