Sie sind auf Seite 1von 198

Louisiana State University

LSU Digital Commons


LSU Historical Dissertations and Theses Graduate School

1971

Digital Control of Complex Systems Based on


Simple Models.
Kuo-cheng Chiu
Louisiana State University and Agricultural & Mechanical College

Follow this and additional works at: https://digitalcommons.lsu.edu/gradschool_disstheses

Recommended Citation
Chiu, Kuo-cheng, "Digital Control of Complex Systems Based on Simple Models." (1971). LSU Historical Dissertations and Theses.
2114.
https://digitalcommons.lsu.edu/gradschool_disstheses/2114

This Dissertation is brought to you for free and open access by the Graduate School at LSU Digital Commons. It has been accepted for inclusion in
LSU Historical Dissertations and Theses by an authorized administrator of LSU Digital Commons. For more information, please contact
gradetd@lsu.edu.
INFORMATION TO USERS

This dissertation was produced from a microfilm copy of the original document.
While the most advanced technological means to photograph and reproduce this
document have been used, the quality is heavily dependent upon the quality of
the original submitted.

The following explanation of techniques is provided to help you understand


markings or patterns which may appear on this reproduction.

1. The sign or "target" for pages apparently lacking from the document
photographed is "Missing Page(s)". If it was possible to obtain the
missing page(s) or section, they are spliced into the film along with
adjacent pages. This may have necessitated cutting thru an image and
duplicating adjacent pages to insure you complete continuity.

2. When an image on the film is obliterated w ith a large round black


mark, it is an indication that the photographer suspected that the
copy may have moved during exposure and thus cause a blurred
image. You will find a good image of the page in the adjacent frame.

3. When a map, drawing or chart, etc., was part of the material being
photographed the photographer followed a definite method in
"sectioning" the material. It is customary to begin photoing at the
upper left hand corner of a large sheet and to continue photoing from
left to right in equal sections with a small overlap. If necessary,
sectioning is continued again — beginning below the first row and
continuing on until complete.

4. The majority of users indicate that the textual content is of greatest


value, however, a somewhat higher quality reproduction could be
made from "photographs" if essential to the understanding o f the
dissertation. Silver prints of "photographs" may be ordered at
additional charge by writing the Order Department, giving the catalog
number, title, author and specific pages you wish reproduced.

University Microfilms
300 North Zeeb Road
Ann Arbor, Michigan 48106

A Xerox Education Company


72-17,752

CHIU, Kuo-Cheng, 19J+0-


DIGITAL CONTROL OF COMPLEX SYSTEMS BASED ON
SIMPLE MODELS.

The Louisiana State University and Agricultural


and Mechanical College, Ph.D., 1971
Engineering, chemical

University Microfilms, A XEROX Com pany, Ann Arbor, Michigan


DIGITAL CONTROL OF COMPLEX SYSTEMS BASED

ON SIMPLE MODELS

A Dissertation

Submitted to the Graduate Faculty of the


Louisiana State University and
Agricultural and Mechanical College
in partial fulfillment of the
requirements for the degree of
Doctor of Philosophy

in

The Department of Chemical Engineering

by
Kuo-Cheng Chiu
B.S., University of Rangoon, 1963
M.S.E., University of Florida, 1968
December, 1971
I

P L E A S E NOTE:

S o m e p a ge s m a y have

i n d i s t i n c t print.

F i l m e d as received.

U n i v e r s i t y M i c r o f i l m s , A X e r o x E d u c a t i o n C o mpany
ACKNOWLEDGEMENT

The author wishes to express his sincere appreciation to

Dr. Cecil L. Smith, Chairman of his Advisory Committee for his

consideration, interest, advice and helpful suggestions given

during the course of this work. He wishes to thank the other

members of his committee, Dr. Jessie Coates, Dr. Albert H. Wehe,

Dr. Jerome A. Planchard and Dr. Philemon S. H. Baw for their

cooperation in serving on the committee. He is also obliged to

Dr. Armando B. Corripio for his suggestions and advice.

The author owes his gratitude to the support of this work

by an Air Force Office of Scientific Research Contract No. F44620-

68-C-0021. The author also gratefully acknowledges the use of

the XDS Sigma 5 computer provided by this contract and the IBM-360

computer in the Computer Research Center at Louisiana State

University.

Thanks are extended to Mrs. Beanie Thibodeaux and Mrs. Sonja

Hartley for their assistance with the preparation of the manuscript.

He is also indebted to the Dr. Charles E. Coates Memorial Fund,

donated by Mr. George H. Coates, for financial assistance in the

preparation of t^is manuscript.

ii
TABLE OF CONTENTS

Page

ACKNOWLEDGEMENT ii

LIST OF FIGURES vii

LIST OF TABLES xii

ABSTRACT xiv

CHAPTER

I INTRODUCTION 1

II MODELING AND STANDARD CONTROL ALGORITHM 6

Introduction 6

Process Description 6

Process Characteristics 12

Model Development 13

Conventional Control Algorithm 23

Tuning Control Algorithms 24

Tuning Techniques 26

Optimization of Parameters 28

Summary and Observations 28

Literature Cited 40

ill
Page

Ill DEADBEAT AND KALMAN ALGORITHM 42

Introduction 42

Design of Control Algorithm Using


z-Transform 42

Deadbeat Algorithm for Step Change in


Set Point 44

Handling of Saturation 50

Ringing 50

Comparison of Algorithms 56

Kalman Algorithm 64

Comparison of First and Second Order


Algorithms 69

Sumnary 82

Literature Cited 84

IV DAHLIN'S CONTROL ALGORITHM 85

Introduction 85

Design of Control Algorithm Using


z-Transform 85

Dahlin's Method 87

Handling of Saturation 88
Ringing Analysis 95

First Order Lag Plus Dead Time Model 95

Second Order Lag Plus Dead Time Model 96

Ringing and Ringing-Free Algorithms 102

iv
Page

Tuning Procedure 111

Summary 120

Literature Cited 123

V ADVANCED CONTROL TECHNIQUES 124

Introduction 124

Cascade Control Systems 124

Cascade Control of Process 128

Inner Loop Tuning 13®


130
Outer Loop Tuning

Cascade Responses 133

Multivariable Control 1^®

Gain Matrix l^2

Relative Gain Matrix 1^9

Significance of the Relative Gain Matrix 152

Tuning 153

Tuning Concentration Loop (FCIM Tuning) 156

Tuning Concentration Loop (FCIA Tuning) 156

Multivariable Control Responses 159

Literature Cited 169

VI CONCLUSIONS 171

v
Page

APPENDIX A DERIVATION OF STANDARD ALGORITHM 173


PARAMETERS FOR. SET POINT CHANGE

VITA 178

vl
LIST OF FIGURES

Figure Page

2.1 The jacketed backmix reactor 10

2 .2a Response to change of controller output 14

2 .2b Responses to changes in feed 15

2.3 Ziegler-Nichols model 16

2.4 Miller's first-order plus dead time model 18

2.5 Sten's second-order model plus dead time 20

2.6 Meyer's second-order model plus dead time 21

2.7 Comparison of model responses to the system


response 22

2.8 Control loop with PI controller and first


order lag plus dead time system 25

2.9 Definition of common response criteria 27

2.10 Optimization of controllers parameters 29

2.11 Comparison of PID load change algorithms 33

2.12 Comparison of PID set point change 34


algorithms

2.13a Comparison of PI and FID load algorithms 35

2.13b Comparison of Pi and PID set point 36


algorithms

2.14a Comparison of PID algorithms on load 37


change

2.14b Comparison of PID algorithms on set point 38


change

vii
Page

Typical digital process control loop 43

First order ringing deadbeat algorithm 51


(Equation 3.8)

Second order ringing deadbeat algorithm 52


(Equation 3.10)

Pole-zero location for equation 3.7 54

First order deadbeat algorithm with complex 55


ringing poles removed (Equation 3.12)

First order deadbeat algorithm with all 57


ringing poles removed (Equation 3.13)

Pole-zero location for Equation 3.9 58

Second order deadbeat algorithm with z=-l 60


removed (Equation 3.15)

Second order deadbeat algorithm with all 61


ringing poles removed (Equation 3.17)

Comparison of algorithms for load change 62

Comparison of algorithms on set point change 63

Second order ringing and ringing-free 65


algorithms on load change

Desired response characteristics used to 66


design Kalman's algorithm

First order Kalman algorithm (Equation 3.25) 72

Second order Kalman algorithm 73


(Equation 3.27)

Pole-zero location for Equation 3.24 74

Pole-zero location for Equation 3.26 75

First order Kalman's algorithm with all 78


ringing poles removed (Equation 3.31)

Second order Kalman algorithm with 79


ringing poles removed (Equation 3.33)

viii
Comparison of algorithms on load change 80

Comparison of algorithms on set point 81


change

Comparison of Kalman ringing and ringing- 83


free algorithms on load change

Typical digital process control loop 86

First order Dahlin algorithm (Equation 4.8) 93

Second order Dahlin algorithm


(Equation 4.10) 94

Pole-zero location of Equation 4.8 97

First order Dahlin algorithm with complex 100


poles removed (Equation 4.15)

First order Dahlin algorithm with all 101


ringing poles removed (Equation 4.17)

Pole-zero location for Equation 4.9 103


Second order Dahlin algorithm with z=<l-e
"JlT)
removed (Equation 4.23) 106

Second order Dahlin algorithm with


removed (Equation 4.21) 107

Second order Dahlin algorithm with all


ringing poles removed (Equation 4.25) 108

First order Dahlin ringing algorithm


(Equation 4.8) on load change 109

Second order Dahlin ringing algorithm


(Equation 4.10) on load change 110

First order Dahlin ringing and ringing-free


algorithms on load change (1=1)

First order Dahlin ringing and ringing-free


algorithms on set point change (1=1)

Ix
Figure Page

4.14a Second order Dahlin ringing and ringing-free 115


algorithms on set point change (X“l)

4.15 First order Dahlin algorithm (Equation 4.17)


with varying X on load change 116

4.16 Second order Dahlin algorithm (Equation


4.25) with varying X on load change 117

4.17 First order Dahlin algorithm (Equation 4.17)


with varying X on set point change 118

4.18 Second order Dahlin algorithm (Equation


4.25) with varying X on set point change 119

4.19 Comparison of algorithms on load change 121


(X=l)

4.20 Comparison of algorithms on set point change


(X-l) 122

5 .1a Classical feedback 125

5.1b Cascade control 125


5.1c Equivalent feedback loop 125
5.1 Cascade control system 125

5.2 Cascade control with different entry of 127


disturbance

5.3 Cascade control of a stirred jacketed 129


reactor

5.4 Model fitting of jacket temperature with slave 131


controller in manual

5.5 Jacket temperature response with slave 132


controller in automatic

5.5a Model fitting of reactor temperature with


134
slave controller in automatic

5.6 Comparison of cascade PI and conventional PI 136


controller for feed rate change

x
Figure Page

5.7 Comparison of cascade PI and conventional PI 137


controller temperature set point change

5.8 Comparison of cascade PI and conventional PI 138


controller for step change in cooling water
temperature

5.9 Response for step change in feed concentration 141

5.10 Response for step change in temperature set 143


point

5.11 Schematic interacting action of multivariable 145


control system

5.12 Open loop response to step change in feed rate 147

5.13 Open loop response to set change in cooling 148


water rate

5.14 Configuration to evaluate bT/QW In 151


c | a
5.15 Two different means of tuning multivariable 155
control systems

5.16 Model and true concentration response with 157


temperature controller in manual

5.17 Model and true concentration response with 160


temperature controller in automatic

5.18 Responses for step change in feed concentration; 161


FCIM tuning; temperature controller on manual

5.19 Response for step change in feed concentration; 162


FClA tuning; temperature control on manual

5.20a Multivariable control responses to step change 164


in feed concentration FCIM tuning

5.20b Multivariable control responses to step change 165


in feed concentration FCIA tuning

5.21a Multivariable control responses to step change 166


in temperature set point, FCIM tuning

5.21b Multivariable control responses for temperature 167.


set point change, FCIA tuning

xi
LIST OF TABLES

Table Page

II-l Process Model Equations 8

II-2 Parameters and Steady-State Values 11

II-3 Tuning Results 3Q

III-l Model Transfer Function 45

III-2 Deadbeat Algorithm for First Order Plus 47


Dead Time Model

III-3 Deadbeat Algorithm for Second Order Plus 48


Dead Time Model

III-4 Deadbeat Algorithms for Second Order Plus 59


Dead Time Model

III-5 Kalman Algorithms for First Order Model 70

111-6 Kalman Algorithms for Second Order Model 71

III-7 First Order Kalman Algorithm with Ringing 76


Poles Removed

III-8 Second Order Kalman Algorithm with 77


Ringing Poles Removed

IV-1 Dahlin Algorithm for First Order Model 89

IV-2 Dahlin Algorithm for Second Order Model 91

IV-3 First Order Dahlin Algorithm 98

IV-4 Second Order Dahlin Algorithm 104

xii
Tables Page

V-l Cascade Control Models and Algorithm 135


Settings

V-2 Response for Cascade and Standard Controllers 139

V-3 Gain Matrix and Relative Gain Matrix 150

V-4 Concentration PID Controller for Load Change 158

V-5 Comparison of Two Tuning Methods 163

xiii
ABSTRACT

The primary purpose of this dissertation is to study the

use of simple models in the design of control strategies for

complex processes. The models utilized were limited to first

and second order models with dead time. Control algorithms

were designed by various methods and applied to control of both

set point changes and changes in unmeasured disturbances.

Advanced control schemes were also investigated to see what

advantages they produced over simple feedback schemes. While

all results obtained are for a specific system, they at least

provide an indication for the expected return for other similar

systems.

Specifically, the study is based on a jacketed, well-mixed

chemical reactor in which an exothermic second order chemical

reaction is proceeding. The controlled variable is the

reactor's temperature; the manipulated variable is the cooling

water flow rate. The unmeasured disturbance is the reactant's

flow rate, which was chosen because of its significant effect on

the dynamics of the model.

Various control schemes such as standard FI and FID, dead­

beat, Kalman and Dahlin algorithms are implemented on the system.

xiv
In all cases, algorithms derived on second order plus dead time

models behave much better than the first order algorithms,

in either step change in disturbance or step change in set point.

The ringing phenomenon is studied for deadbeat* Kalman and Dahlin

algorithms. It is concluded ringing-free algorithms should be

used for step change in set point while ringing algorithms are

better for changes in disturbances. For this specific system,

ringing-free algorithms designed by all three techniques

reduce to either the standard PI or PID algorithms.

Advanced control techniques such as cascade and multivariable

control are investigated. Cascade control with a PI master

controller and a proportional slave controller performed

better than the simple feedback techniques. Gain matrix,

relative gain matrix, interaction and pairing of manipulated

and controlled variables are studied. Two different tuning

techniques for the concentration loop are investigated. It is

concluded in this particular system, the two tuning techniques

give virtually identical results.

xv
Chapter I

INTRODUCTION

In recent years the digital computer has had a real impact

in the design and implementation of control systems in all

areas of technology. Improved computer technology including

analog-to-digital interface now makes it possible to directly

use the digital computer to control virtually any system.

The memory, logic capabilities, speed of computation and general

flexibility of the digital computer over conventional analog

systems offers the control engineer more power and freedom in

both design and implementation.

This dissertation will examine and study several aspects

cf the direct control of chemical processes using digital

computation. The system upon which this study is made is a

water-jacketed backmix reactor in which an exothermic second

order chemical reaction is taking place. The controlled variable

is either product temperature or concentration or both. Control

algorithms for load change and set point change are studied.

In Chapter II, the system is modeled to a set of differ­

ential equations and is simulated using Euler's forward difference


method. Temperature response of the reactor due to a step

change in the cooling water rate is approximated to either a

first or second order plus dead time model by different techniques

(1,2,3). For each model, optimum settings for the standard FI

or PID control algorithms are sought by the use of Pattern Search

(4) to minimize the Integral of Time and Absolute Error (ITAE)

criterion. The minimization is made for load disturbances as

well as for set point disturbances. Results obtained by imple­

menting these controller settings on the reactor are compared

and studied.

In Chapter III, the control algorithm is designed so that

the system will respond to a specific disturbance in a desired

manner. Deadbeat and Kalman algorithms are designed for set

point change on the system. The continuous transfer functions

derived from models in Chapter II are transformed into pulse

transfer functions by z-transforms (5) using two different sampling

times. Ringing of the manipulated variables for both load changes

and set point changes are studied. Techniques to reduce ringing

and the resulting response of the controlled variable in com­

parison with the previous response are also studied. A surprising

result is that the ringing-free algorithm is of the same form as

the standard FI or PID controller, but the tuning constants are

slightly different from those obtained in Chapter II.


In Chapter IV, Dahlin's algorithm, which has one adjustable

parameter, Is examined. Ringing of the manipulated variable

for both load changes and set point changes are studied.

Ringing, ringing-reduced or ringing-free algorithms are compared

for a "good" value of the adjustable parameters. Comparisons are

made on various values of the parameters. The ringing-free

algorithms again reduce to the standard PI or PID controllers.

In Chapter V, advanced control techniques such as cascade and

multivariable control are considered. All the models required

for tuning the algorithms in this chapter are obtained by a

least square fit to the true response. A proportional controller

is employed as the slave controller and a PI controller is

employed as the master controller of the cascade arrangement.

The resulting response is quite superior to that of the algorithms

based on a second order plus dead time model in Chapter II.

In multivariable control algorithms both temperature and concen­

tration of the product stream are controlled. The significance

of gain matrix and relative gain matrix in pairing the manipulated

and controlled variables is presented. Two different tuning

methods of the concentration loop are studied. The first method

is to tune with the temperature controller in manual and the other

method is with the temperature controller in automatic. Results

obtained by both the methods are compared.


In summary, the purpose of this dissertation is to discuss

the application of the various digital control algorithms to a

specific highly nonlinear chemical process system. One of the more

interesting results obtained from this study is that, for the

system studied, all ringing-free z-transform algorithms reduced

to the conventional PI or PID algorithms. This appears to

seriously question the contention that algorithms designed by

z-transforms are far superior to conventional algorithms, at

least for systems without large dead times such as the one studied.

The use of an advanced control technique, especially cascade,

gives considerable improvement over simple feedback control.


Literature Cited

Miller, J. A. et.al, "A Comparison of Controller Tuning


Techniques", Control Engineering, (Dec. 1967), p. 72.

Meyer, J. R. et.al., "Simplifying Process Response Approxi­


mation", Instruments and Control Systems. Vol., 40, (Dec.
1967), p. 76.

Sten, J. W., "Evaluating Second-Order Parameters", Instrument


Technology. (Sept. 1970), p. 39.

Hooke, R., and Jeeves, T. A., "Direct Search Solution of


Numerical and Statistical Problem", J. Assn. Comp. Math.,
8 , 2 (April 1961), pp. 212-229.

Kuo, B. C., Analysis and Synthesis of Sampled-Data Control


Systerns, Prentice Hall, Inc., Englewood Cliffs, N.J., 1963.
Chapter II

MODELING AND STANDARD CONTROL ALGORITHM

Introduction

The primary purpose of this chapter is to investigate the

suitability of various simple models as a basis for controller

tuning. The system to be controlled is a water-jacketed backmix

reactor in which a second order chemical reaction is taking place.

The product is to be controlled at a particular temperature for

either a step change in feed rate or step change in temperature

set point. First and second order plus dead time models obtained

by different modeling techniques are compared. Optimum setting

for the standard PI or PID control algorithms are evaluated by

minimizing the Integral Time of Absolute Error (ITAE) criterion.

Results obtained by implementing these controller settings on

the system are compared.

Process Description

The system chosen for this study of digital control algorithms

is the control of reactants temperature in a jacketed backmix

reactor in which an exothermic second order reaction takes place.


3
Reactants of concentration C^Clbs/ft ) and temperature ^ ( “F) are

fed at the rate of W(lbs/min). The contents of the reactor are

assumed to be perfectly mixed at concentration C and temperature


&
T. The products are removed at the rate W, with the same concen­

tration and temperature as the contents In the reactor. The resulting


3 3
mixture has volume V(ft ), density p(lbs/ft ) and specific heat

Cp (Btu/lb-°F).

Cooling water of T (°F) enters the jacket at a rate of W


w c
(lbs/min) and leaves at T (°F). The heat capacity of the walls
c
of the reactor and of the jacket Is combined with that of the

water In the jacket and denoted by M (Btu/°F). Specific heat


c
of water Is taken as unity and heat losses to the surroundings are

negligible. The overall heat transfer coefficient between the

reactor contents and the jacket Is assumed constant at U(Btu/


2 2
min-ft -°F). The heat transfer area of the jacket is A(ft ).

The equations describing this system are derived hy making

an unsteady-state mass balance on reactant A, and energy balance

on the reacting mixture* and an energy balance on the jacket.

The resulting equations are given in Table II-l.

As shown in Figure 2.1, the objective of this study is to

control the temperature of the reacting mixture, T, by manipu­

lating the cooling water rate, Wc> via a digital computer.

The temperature transmitter and the cooling water flow controller

are assumed to behave like first order lags with time constants

•r and t minutes respectively. The corresponding equations


t V

are also given in Table 11-1.

The numerical values for the parameters and the initial

steady state conditions are shown in Table 11-2,


8

Table II-l

Process Model Equations

Mass Balance on Reactant A


dC
Pv -dT " W(Caf- Ca> ' k°a2 <2 - ^

-a
(T+460)
k « k0e (2 .2)

Energy Balance on Reacting Mixture

dT 2
pvc - WC ( T r T )-UA(T-T )+(-AH)VkC (2.3)

Energy Balance on Cooling Water

dT
Mc -jf - UA(T-Tc) -Wc (Tc-Tw ) (2.4)

3
k = reaction rate constant, ft /lb-mln
3
k “ Arrhenius rate constant, ft /lb-min
o
a = Arrhenius temperature constant, °R

-AH “ heat of reaction, Btu/lb of A consumed

dT _
T — - T - T (2.5)
Tt dt c cR
Table II-l (cont.)

dW _
cR _ „
V “dt" " Wc " cR (2.6)

TcR = Delayed temp, °F

W = Delayed cooling water rate, lb/min


CK
CONTROLLER
SET POINT FEED

^ COOLING
** WATER

COOLING
WATER

PRODUCT

Figure 2.1. The jacketed backmlx reactor.


Table II-2

Parameters and Steady-State Values

Steady-State Conditions System Parameters

C = 3 . 6 lb/ft3 a = 2560°R
a
Ca£ = 9.0 lb/cu.ft. C = 0.9 Btu/lb-°F
P
k = 0.0278 cu.ft/lb-mln AH = 867 Btu/lb

T = 190°F k = 1 . 4 3 ft3/lb-min
o
T = 120°F M = 6000 Btu/°F
c c
Tf = 150°F UA = 600 Btu/min-°F

T = 80°F V = 250 cu.ft.


w
W = 1000 lb/mln p = 60 lb/cu.ft.

W = 1050 lb/mln *r = 1 min


c t
t = 0.1 min
Process Characteristics

The disturbances considered in this study are a step change

in load (specifically feed flow rate W) and a step change in

set point. Although the assumptions made in the derivation of

the system equations simplify the problem somewhat, the most

important non-linearities of the system are still included in the

mathematical model. Of these, the one that makes this system

especially difficult to control is related to the operation of

the jacket. As more cooling is required, the jacket water

temperature Tfi must decrease so as to increase the rate of heat

transfer from the contents of the reactor. This decreases the

temperature rise of the cooling water (Tc - T^). Since this

term is multiplied by the flow rate of cooling water,Wc (equation

2.4) the net effect is a decrease in the effective gain of the

control action. This change in gain causes the control response

to be sloppy at high cooling rates, while unstable at low cooling

rates. Corripio and Smith (1) show that the response of the

reactor temperature for a step change in set point from 190°F to

185°F is different from that of 185°F to 190°F, a result of

these nonlinearities.

Although the difference between the reactor temperature

and the jacket temperature could possibly be used to adjust

the controller gain in an adaptive fashion, this was not attempted.

Also, using a cascaded controller for the jacket temperature should

provide some improvement.


The system in question is self-regulated by nature. With

small disturbances in load, such as feed flow rate, feed tempera­

ture or cooling water rate,the process seeks another steady-state

in the manner illustrated in Figure 2.2. Of these, the response

to the disturbance in feed flow rate is most interesting. With

a sudden increase in feed, the temperature in the reactor

initially decreases ( T ^ C T ) . Nevertheless, the concentration

in the reactor tends to rise, causing the reaction rate to increase.

More heat is produced and the oveall temperature in the reactor

will eventually rise to a new steady-state.

Model Development

Response to changes in the manipulated variable W c is employed

for model development. Owing to the nonlinearities of the system

as discussed in the last section, the gains are different for

different changes in the manipulated variable. In this study,

a decrease in cooling water from 1350 to 1050 (lbs/min) is ar­

bitrarily chosen.

In many applications it is only necessary to have a rough

approximation of a mathematical model for a system; the simpler

the model the better. The first tuning technique presented in

the literature was that of Ziegler and Nichols (3), who charac­

terized the process by two parameters and

in Figure 2.3. Ziegler and Nichols developed empirical equations

to relate the tuning parameters to the constants LRand R^. As

illustrated in Figure 2.3, their construction can be readily


o RESPONSE CURVES
Q

o>

05

Q_

CD
CD.

AW = -300 lb/min.

03
0.00 10.00 20.00 30.00 50.00 00
TIME IN MIN.
Figure 2-2a. Response to Change of Controller Output
.
O
RESPONSE CURVES
Q
=a*
o>

o Feed Temp.
o
(AT, = 5°F)

CD

rxj Feed Rate


o>_
IN F

(AW = 200 lb/min.)

CD
o
TEMP

o
05

O
Cl)
CO.
10.00 30.00 U0.00 50
T I M E IN M I N

Figure 2 .2b. Responses to changes in feed


o
o

Slope = IL 2.3 min

Slope = 0.18oF/min
3 09
1'
3'qq = -0.0103 °F/pound per minute

2.3 min

17.2 min

“2.3s
T= 17.2 -0.0103 e
17.2 s + 1

0.00 1 0 .0 0 20.00 30.00 HO. 00 50.00 60.00


T I M E IN MIN.
Figure 2.3. Ziegler-Nichols Model.
extended to obtain a first order lag plus dead time model of

the form

K 33 process gain

0 = dead time

T 9 time constant

But since empirical equations were used, Ziegler and Nichols'

method does not depend upon this procedure giving a good fit.

The method suggested by Miller (2) illustrated in Figure 2.4

generally produces an approximation that "looks" better; and thus

would be expected to give better results in any model-based

tuning methods such as those suggested by Lopez (4,5).

As the order of the model is increased, its ability to

describe the process improves. The next logical step is to a

second order plus dead time model (6):

G(s) = (Tis+1)(t2s+1) “ 2

where

K = process gain

6 = dead time

t^,T2 “ time constants

0)^ =» natural frequency

£ “ damping ratio
o>

CD

o>_

-0.0103 °F/pounds per minute

63.2% of change 12.5 min

Q. 2.3 min.
Z
LUo -0.0103 e”2 *3s
G(s)
12.5s+l (2.8)
CO
CO-

o T = 12.5

p*
CO.
0.00 1 0. 0 0 20.00 50.00 60

Figure 2.4. Miller's First-Order Plus Dead Time Model.


The main deterrent to considereing second order models is the

difficulty of evaluating the parameters. However, several

techniques have appeared in recent years.

Oldenbourg and Sartorius (7) proposed a technique for deter­

mining the time constants for overdamped systems without dead time.

Smith (8) and later Cox et.al. (9) extended this method to systems

with dead time. Meyer et.al. (10) developed a quick and easy

technique for evaluating both over-damped and under-damped

second order plus dead time models. Sten (11) published a

graphical method for overdamped systems. An improved pulse

testing method for second order plus dead time model is developed

by Pemberton (12).

The derivation of the model via Sten's method is illustrated

in Figure 2.5; via Meyer’s method in Figure 2.6. Miller's,

Sten's and Meyer's models together with the system response

are shown in Figure 2.7. Meyer's model appears to approximate

the process response slightly better than Sten's model, whereas

Miller's first order model is poorest of all. For study of control

algorithms, only Meyer's second order and Miller's first order

models will be considered. However, that Meyer's method is always

better than Sten's method should not be concluded; for other systems

the converse may be true. In fact, the two second order models

are so nearly the same that little difference in the results

presented subsequently would be expected had Sten's model been

used instead of Meyer's.


o

CD

= 0.054

o
o>_
'rom Figure 2 in the original artic .e (11)

1.72 min
CL. 13.14 min

0.93 min
CD
co_ 1.37 min
Gf.i --o-oioae'1-373 (2.9)
= 17.2 K J (1.72S+1)(13.14s + 1)
3.09
-300
o
-0.0103 °F/pound per min.
03
0 .00 10.00 20.00 50.00

Figure 2.5. Sten's Second-Order Model Plus Dead Time.

c
o>
0 = 1.0 min

o 27.8 min
o>. _9 _

t„/tft = 0.1728

Fran Figure 5 > in the original article (10)

0.18 min
UJo
T.= 11.84 min
CD
00- T0= 2.61 min

K = -0.0103°F/pound per min


to
G M = -0-0103 e-1-0s
v^ s' (11.84s+l)(261s+l) (2 .10)
0 00 . 10.00 20.00 30.00 0.00 60.00
T I M E IN MIN
Figure 2.6. Meyer.!s Second-Order Model Plus Dead Time.
o RESPONSE CURVES
o

o>.

o Miller
o>_ System

Meyer
ll-C3

4z-ai
»■ »co_
Sten

CD
00-

O
O

03
0.00 5.00 1 0.0 0 15.00
T I M E IN M I N .
Figure 2.7. Comparison of Model Responses to the System Response.

77
Conventional Control Algorithm

The most common algorithm is the discrete equivalent of the

proportional-plus-integral-plus-derivative (PID) controller. The

PID algorithm is derived by simply replacing the integral by a sum

and the derivative by a finite difference. Hence,

m
n
(2 .11)

where

K = proportional gain
c
T^ = reset time

T = derivative time
d
m^ = initial valve position

T = sampling time

m = manipulated variable

e = error

This is known as the position form of the PID algorithm

since the actual valve position is calculated from the error

sequence.. An alternative is to calculate the change in valve

(actuator) position rather than its actual value via the velocity

algorithm:

Both of these algorithms are commonly used in control systems (13).


Tuning Control Algorithms

The critical phase of the implementation of the PI or PID

algorithm is the selection of the numerical values of the constants

Kc , T^, and Td in the algorithm. The steps in tuning a digital

control algorithm parallel the steps in tuning an analog controller.

These are:

1. Approximate the process with a model

a. First-order plus dead time model

b. Second-order plus dead time model

2. Select the optimum constants for controlling the chosen

model.

3. Apply the results to the physical process

For a PI controller, the control loop is shown in Figure 2.8.

For a selected input the response is a function of only the

controller parameters, i.e., Kc and T^. The objective of the

tuning process is to select the combination of and T^ that

gives the responses C(t) that most nearly satisfies the desired

criteria. This is expressed as optimize[fc(t)} Kc ,tJwhich reads

"Optimize the response C(t) over Kc and T^". Although there are

many criteria upon which the selection is based, the commonly

used criteria can be separated into two classes: (1) the simple

but approximate criteria based on only a few points on the response

and (2) the more exact (but more difficult to evaluate) criteria

based on the entire response.

In the first category fall such criteria as settling time,


Load or
Disturbance

Controller Plant

Ke Output,C(t)
t s +1

Figure 2.8. Control loop with PI controller and first order lag plus
dead time system.
percent overshoot, rise time, decay ratio, etc., which are defined

in Figure 2.9. Of these, the one-quarter decay ratio has been

recognized as a reasonable trade-off between a fast rise time and

a reasonable settling time.

In the second category fall such criteria as integral criteria,

whose general form is

Of these, the one that is probably most appropriate for

controller tuning is the integral of time multiplied by the

absolute value of the error (ITAE):

00

Tuning Techniques

The primary difference between tuning continuous control

loops and digital control loops is that in the latter case an

additional parameter, the sampling time, must be taken into

consideration. Lopez (4,5) developed tuning graphs based on inte­

gral criteria, but this additional parameter made it impossible

to reduce them to a convenient equation form. Moore (14) examined

the possibility of using the dead time equivalent to half the

sampling time in the zero-order hold. The continuous tuning

techniques could be applied by using 0 + T/2 instead of 0.


Decay Ratio

105%
100% •
Output

95%

Time t:o~firs-t "peak'


settling time
1— Rise Time
Time

.Figure 2.9. Definition of common response criteria


The first tuning techniques to appear were for continuous

systems based on the one-quarter ratio criterion. The first was

proposed by Ziegler and Nichols (3); Cohen and Coon (15) presented

a later version; and finally Smith and Murrill (16, 17). The

control parameters that minimize the integral criteria for load

changes were developed by Lopez (18) and later Rovira (19)

extended this to set point changes. Lopez (20) developed tuning

techniques for a second order plus dead time model with disturbance

in load. Rovira (21) suggested a modified algorithm if the

controller's primary function is to compensate for load changes.

Optimization of Parameters (22)

The ITAE criterion is chosen in evaluating the optimal

parameters for the first order plus dead time model, the second

order plus dead time model, and the process. The pattern search

technique (23) is employed for both PI and PID algorithms. The

optimization is schematically shown in Figure 2.10.

Table II-3 gives the results from applying the optimal

parameters for PID and PI algorithms to the original process.

The step change in load is from W = 1000 to W ■ 12000 (lbs/min).

The step change in set point is from T = 190°F to T = 185°F.

The sampling time is one minute.

Summary and Observations

From Table II-3, several interesting observations can be

made. First, the Ziegler-Nichols technique works ajLsrost perfectly


OPTIMIZATION CRITERION

Set
Point STANDARD
CONTROLLER SYSTEM

Figure 2.10. Optimization of Controllers Parameters.


Table II-3

Tuning Results

ITAE ITAE Decay


Method K T. Set Point Load Ratio
c i Td

PID Tuning, Load Change

Ziegler-Nichols Tuning -520 5.6 1.4 289.3 68.5 .022

First-order Tuning -543 4.92 1.075 355.7 105.2 0.104

Second-order Tuning -700 42605 2.1 259.5 38.9 .004

True Optimum -710 2.9 3.0 335.4 36.2 .085

PID Tuning, Set Point Change

Ziegler-Nichols Tuning -520 5.6 1.4 289.3 68.5 .071

First-order Tuning -337 16.4 0.96 118.1 376.6 -

Second-order Tuning -679.6 23.3 2.0 65.5 248.0 -

True Optimum -890 27.1 1.5 54.8 201.8 -


Table II-3 (cont.)

ITAE ITAE Decay


Method K T. Set Point Load Ratio
c i Td

PI Tuning, Load Change

Ziegler-Nichols Tuning -390 9.32 - 290.4 180.5 0.152

First-order Tuning -360 6.71 - 400.3 218.3 0.139

Second-order Tuning -390 10.1 - 276.5 189.3 0.156

True Optimum -410 9.1 - 243.7 178.4 0.157

PI Tuning, Set Point Change

Ziegler-Nichols Tuning -390 9.32 - 290.4 180.5 .079

First-order Tuning -225 12.6 - 200.6 400.5 -

Second-order Tuning -582.5 20.0 - 163.6 248.8 0.164

True Optimum -390 17.3 - 148.0 329.5 .022


for the PI control, load change case, but for PID control, load

change, a 50% improvement can be obtained. However, in both

cases and especially for PID control, the decay ratio is less

than the objective of one quarter. For both set point cases,

the Ziegler-Nichols technique was poorest of all. The first-

order tuning method was poorest of all on the PI and PID load

changes cases. An indication of the quality of the various

settings can be obtained from Figure 2.11 for the PID, load

case and Figure 2.12 for the PID, set point case.

On the question of PI and PID control, the improvement of the

PID over PI was by a factor of about 5 for load changes and almost

3 for set point changes. This is illustrated for both cases in

Figure 2.13.

The values in Table II-3 also indicate that optimum settings

for load changes are not optimal for set point changes, and vice

versa. This situation is much more pronounced for the PID case

than for PI, as is illustrated in Figure 2.14.

In each of the four cases presented in Table II-3, the second

order model produced settings that were quite close to the true

optimum. In each case, the improvement over the first order

model was substantial, and in fact one could question whether

the first order model adequately represents this particular

system. The main deterrent to the use of a second order model is

that it is considerably more difficult - in many cases virtually


jj

192 . 0 0

first-order
Optimum
190.00

second-order
IN r.
168.00
TEMP
155.00

I D . 00 0.00 1 0 .0 0 20.00
TIME FM M

Figure 2.11. Comparison of FID Load Change Algorithms.


195.00
.190.00
IN F.

first-order

seqond-order
185,00
TEMP

optimum
IBD.tJO

.
0 00 10.00 30.00
T I M E IN MIN.

Figure 2.12. Comparison of PID Set Point Change Algorithms


192.00

PID
190.00
IN F.
186.00
.TEMP
86.00

- 10.00 0 . 00 10.00 20.00 30

Figure 2.13a. Comparison of PI and PID load algorithms.


36

o
o
in
03

o
03
Ll.

1— 4
O

a
o
oo
• ID.00 0.00 10.00 30.00
T I M E IN MIN.

Figure 2.13b. Comparison of PI and PID set point algorithms.


j/

192.00

Set point
190.00
IN F.

Load
1 8 6 . DO
TEMP
1 8 6 . DO

0.00 0 .00 10.00 20.00 30.00


T I M E IN MIN.

Figure 2.14a. Comparison of PID algorithms on load change.


195.00
190.00
IN F.
1 8 5 . DO

Set point
TEMP
180.00

- 10.00 .
0 00 1 0 . OU 20.00 30.00

Figure 2.14b. Comparison of PID algorithms on set point change.


impossible to obtain accurate parameter values from normal

chart records. However, in any computer-based or automated

tuning procedure, use of the second order model would be

preferable.
Literature Cited

1. Corripio, A. B., Smith, C. L., "Computer Simulation to Evaluate


Control Strategies", Instruments and Control System, January
1971, p. 87.

2. Miller, J. A., et.al., "A Comparison of Controller Tuning


Techniques", Control Engineering. (Dec.1967), p. 72.

3. Ziegler, J. G., and Nichols, N B., "Optimum Setting for


Automatic Controller", Transactions ASME» (November 1942),
p.759.

4. Lopez, A. M . , Murrill, P. W., and Smith, C. L., "Optimal


Tuning of Proportional Digital Controllers", Instruments
and Control Systems, (Oct. 1968), p. 97.

5. Lopez, A. M., et.al., "Tuning PlandPID Digital Controllers",


Instruments and Control Systems, (Feb. 1969), p. 89.

6 . Harriott, P., Process Dynamics, McGraw-Hill Book Co., New


York, 1964.

7. Oldenbourg, R. C., and Sartorius, H., "The Dynamics of Auto­


matic Controls", ASME, New York (1948), p. 276.

8. Smith, 0. J. M., "A Controller to Overcome Dead Time",


ISA Journal, Vol. 6 , No. 2 ( February 1959 ),
pp. 28-33.

9. Cox, J. B., et.al, "A Practical Spectrum of DDC Chemical-


Process Control Algorithms", ISA Journal, Vol. 13, No. 10,
(Oct. 1966), pp. 65-72.

10. Meyer, J. R., et.al., "Simplifying Process Response Approxi­


mations", Instruments and Control Systems, Vol. 40, (Dec.
1967), p. 76.

11. Sten, J. W., "Evaluating Second-Order Parameters",


Instrumentation Techn61ogy, (Sept. 1970), p. 39.
12. Pemberton, T. J., "An Improved Pulse Testing Method",
Instrumentation Technology,(Dec. 1970), p. 51.

13. Fertik, H. A., and Ross, C. W., "Direct Digital Control


Algorithm with Anti-Windup Feature", Leeds & Worthrup
Technical Journal, (Jan. 1968), p. 1.

14. Moore, C. F., Smith, C. L., and Murrill, P. W., "Simplifying


Digital Control Dynamics for Controller Tuning and Hardware
Lag Effects", Instrument Practice, (Jan. 1969), p. 45.

15. Cohen, G. H., and Coon, G. A., "Theoretical Considerations of


Retarded Control", Taylor Instrument Companies Bulletin
#TDS-10A102.

16. Smith, C. L., and Murrill, P. W., "Analytical Tuning of


Underdamped Systems", ISA Journal, Vol. 13, No. 9, p. 40.

17. Smith, C. L., and P. W. Murrill, "Controllers - Set Them


Right", Hydrocarbon Processing and Petroleum Refiner, Vol.
45, No. 2, (Feb. 1966).

18. Lopez, A. M., et. al., "Controller Tuning Relationships


Based on Integral Performance Criteria", Instrumentation
Technology, Vo . 16, No. 11, Nov. 1967, p. 57.

19. Rovira, A. A., Murrill, P. W., and Smith, C. L., "Tuning


Controllers for Set point Changes", Instruments and Control
Systems, (Dec. 1969), p. 67.

20. Lopez, A. M. Murrill, P. W. , and Smith, C. L., "An Advanced


Tuning Method", British Chemical Engineering, Vol. 14, No. 11,
(Nov. 1969), pp. 1552-1555

21. Rovira, A. A., "Modified PI Algorithm for Digital Control",


Instruments and Control Systems, Vol. 43, No. 8 (Aug. 1970),
pp. 101-102.

22. Wilde, D. J., and Beightler, C. S., Foundations of Optimization,


Prentice Hall, Inc., Englewood Cliffs, New Jersey.

23. Hooke, R. and T. A. Jeeves, "Direct Search Solution of


Numerical and Statistical Problem,"Journal of the Association
for Computing Machinery, 8 , 2 (April 1961), pp. 212-229.
Chapter III

DEAD-BEAT AND KALMAN ALGORITHM

Introduction

The conventional PI and PID controllers are tuned to

satisfy a particular criterion such as ITAE in Chapter II.

As far as the algorithm is concerned the standard algorithm

stays the same except for the coefficients, which makes it

quite restricted.

In order for a particular system to respond to a specific

disturbance in a desired manner, the controller has to be

designed for each individual system. Dead-beat and Kalman

algorithms are considered in this chapter.

Design of Control Algorithm Using z-Transform (1)

The general digital control loop considered in this study

is illustrated in Figure 3.1. The objective is to design the

controller D(z) so that the desired loop performance is obtained.

From block diagram analysis with negative feedback (2),

Figure 3.1 can be represented by the following equation:

C(z) « HG(z)D(z) |R(z)-C(z2 (3.1)


Set Zero-order
Point Controller Hold Plant C(s)
D(z) 1-e -sT G(s)
- Error Plant
Manipulated Output
Variable

Feedback Signal C(z)

Figure 3.1. Typical digital process control loop.


where

HG(z) = Z --- G(s)J (3.2)

is the pulse transfer function of the process.

So in order to design D(z), an expression for HG(z) must

be obtained in some manner. Once HG(z) is specified, the only

unknowns in equation 3.1 are D(z) and C(z)/R(z). If the desired

loop performance characteristics could be used to specif^

C(z)/R(z), then equation 3.1 could be solved for D(z):

K } 777
HG(z) 1-C(z)/R(z) 0.3)

Unfortunately, C(z)/R(z) cannot be specified without giving some

attention to HG(z). For example, if HG(z) contains a time delay,

then C(z)/R(z) must contain the same time delay.

One approach to obtain expressions for HG(z) is to use the

modeling techniques described in Chapter II, where expressions

for HG(z) corresponding to first and second order plus dead time

models are derived (3). Table III-l gives numerical values for

HG(z) corresponding to two different sampling times.

Deadbeat Algorithm for Step Change in Set Point

A deadbeat or minimal response is one that satisfies the

criteria that (1) the settling time must be finite, (2) the rise

time should be a minimum, and (3) the steady-state error should


Table III-l

Model Transfer Function

First order plus Second order plus


dead time model dead time model

Continuous Transfer
■0.0103 e-2 -3!_ ,0i0103 e-ls
Function
12.5s + 1
(11.84s+l)(2.61s+l)

Pulse Transfer
[b.0545-fq.02_24Z-.1] 0139+0.0119z_1
-o.oios*’2 [ M i 2
Function T = 1 -3
-0.0103z |
L „1 - « no„_-l1 JI
0.9231Z- I,-, ''>
L(l-0 .9i90z'1)(l-0.687z"1X

Pulse Transfer
-2 [~0.1272 + 0.0207z~1 -1 [~0.0139+0.06
0619z’1+0.1074z"2
Function T = 2 -0.0103z -0.0103z
L 1 - 0.8521z_1 . L(l-0.8446z' 1)(l-0.4647z"1)
One specific case that satisfies the above criteria is that

the response to a step change in set point should have zero

error at all sampling instants after the first.

R(z) = — * . (step input)


1-z

C (z) = z"1 R(z) (3.4)

C(z)/R(z) = z *

Substituting into equation 3.3 gives

-1 .
D(z)\ = Z 1
1_z-l HG(z) (3.5)

The control algorithm will be physically realizable only

if the time delay in HG(z) does not exceed the sampling time.

So in general cases

-N-l x
D(z) = 1_z-N-l HG(z) <3 *6)

where NBnearest Integer number of sampling times in dead time.


The controller D(z) and the equivalent difference equations

for equations for first and second order models are given in

Table III-2 and Table III-3 respectively with sampling time

T “ 1 minute.
Table III-2

Dead-beat Algorithm for

First order Plus Dead Time Model

D(z) = (3.7)

(3.8)

where
1
a, = -
1 T
-a.mT
C = 1 - e
6
-a-jmT -a-T
C? = e - e

0 = N T + 6

and

T = 1 min. = sampling time K = process gain

9 = dead time E = error

t “ time constant M = manipulated variable

(3.8a)
Table III-3

Deadbeat Algorithm for

Second order Plus Dead Time Model

-aT -1W 1 -bT -1.


D(z) = 0 - 1 ^ ----» -? q - e 5-i. (3.9)
-2 2 -1
KC-Cl-ss ^ ( 1 + —
I cx z )

or

M, = J L fE . (.- « + e-bI)E ,+ e-<a+b>T E ,1


n KC^ L n ' n-1 n-2J

C C
E7 V l + Mn-2 + 57 “a-3 (3-10>

E = K C 1 fM + ~ M M „ - 7—- M J + (e"aT+ e“bT)E ,


n 1 L n Cj n-1 n-2 n-3J ' n-1

- e"<a+b)T En_2 (3.10a)

where
1
a = --
T1

b = ^
T2

9 - NT

T = sampling time = 1 min.

be"aT- ae"bT
C1 ° 1 + ' a - b
Table III-3 (cont.)

c _ -(a+b)T be~bT - ae~aT


2 _ + a - b

and

0 = dead time

t^»T2 = time constants

K = process gain

E = error

M = manipulated variable
Handling of Saturation

Physical constraints are imposed upon the system. The fully

opened valve position corresponds to 6,000 lb/min of cooling

water rate, whereas the fully closed valve position corresponds

to no flow of cooling water. When saturation occurs, the error

in equations 3.8a and 3.10a is recalculated from the control

equation 3.8 or 3.10 by employing either equal to 6,000 or 0.

This practice is analogous to the use of the velocity form

of the PID algorithm as opposed to the position form. Applying

the control algorithms in Tables III-2 and III-3 to the original

process for a step change in set point, the manipulated and the

controlled variables are as shown in Figures 3.2 and 3.3.

Ringing

Judging the performance of a controller solely from the

response is generally not sufficient. The manipulated variable

(valve signal) m(t) in Figures 3.2 and 3.3 exhibits excessive

valve movement, a phenomenon known as ringing. Obviously the

controller outputs are ringing. In other words, the algorithm

acts like an on-off controller, which is not the objective of our

design.

The ringing amplitude is defined as the difference between

the first two controller outputs, and can be related to the

algorithm's pole - zero locations. The z ** -1 point is called

the ringing node, and the pole at z = -1 is known as the ringing


o
o
u>
r-
Ol
O

CCo
UJ •
a ”

10.00 0.00 1 0.0 0 20.00 30.00


T I M E IN MIN.
o
o
lO
o>,

o>.

Q
CO,
-10.00 0.00 10.00 20.00 30.00
T I M E IN MIN.

Figure 3.2. First order ringing deadbeat algorithm (Equation 3.8).


a
0

«l
O

CCo
LJ •
n 10

CO
Q
2 2

- 10.00 0.00 1 0 .0 0 30.00


T I M E IN MIN.
o
o
U>

o
o
O
U-i05
z
o

o
CJ
CD.
- 10.00 0.00 1 0 .0 0 20.00 30.00
TI M E IN MIN;
Figure 3.3. Second order ringing deadbeat algorithm (Equation 3.10).
pole. Moving the pole away from the ringing node decreases

the ringing amplitude. A pole in the right-half plane reduces

the ringing amplitude, while zeros in the right-half plane

aggrevate it (4).

For the algorithms derived from the first-order model,

there are one real and two complex conjugate poles in the left-

half plane as shown in Figure 3.4. In fact, the two complex

poles are on the unit circle. Dahlin (4) suggests the ringing

pole be simply eliminated and the gain adjusted accordingly.

Since the complex conjugate poles are on the unit circle,

it seems reasonable to remove them first, giving the algorithm

(3.11)

or
1
M
n 3KC,
6
(3.12)

The controller output of this algorithm, shown in Figure 3.5,

is improved considerably.

By removing the pole z = -C_/C, from equation 3.11, we obtain


7 o
the algorithm
1
Imaginary

1.0

923 Real

Figure 3.4. Pole-zero location for Equation 3.7.


o
CD

*— i

oi

21 CD

-ID.00 0.00 10.00 20.00 30.00


TIME IN MIN.
in
03

Li_

•— 1

o
to
" - ' 0.00 0.01 10.00 20.00 30.00
TIME IN MIN
Figure 3.5. First order deadbeat algorithm with complex ringing
poles removed (Equation 3.12).
Using this algorithm gives virtually the same process response,

as shown in Figure 3.6, as was obtained in Figure 3.5. The

controller output is also very similar. This algorithm happens

to be a PI controller with a gain of -389.0 and a reset time of

12.0. For PI control of step changes in set point, minimization

of the ITAE criterion gives the values -225 and 12.6, respectively.

Similar results are obtained for the algorithm derived

from the second order model. The pole-zero locations are shown

in Figure 3.7. Removing the ringing pole at z = -1 gives the

control algorithm in equation 3.14; removing the ringing poles

at z = -1 and z = gives the algorithm in equation 3.16

as can been seen in Table III-4. As shown in Figure 3.8 and

3.9, the algorithm with both poles removed exhibits much less

ringing, whereas the responses of the two algorithms are virtually

identical. However, equation 3.16 is identical to the conven­

tional PID algorithm with Kc= -654.8, T^ = 13.49, and T^ = 1.80.

In Chapter II, minimization of the ITAE criterion for a step

change in set point gave Kc = -679.6, T^ 83 23.3, and T^ ° 2.0.

Comparison of Algorithms

From Figures 3.10 and Figure 3.11 it is obvious, that the

second order model algorithm is always better than the first

order model algorithm for either load or set point changes.

For load change, responses using algorithms without the ringing

poles removed give less overshoot and less oscillatory behavior


o
c=>

•— i

0.00 10.00 20.00 30.00


TIME IN MIN.
CD
O
urj
o>

CD

CD
0- -

o
oo
- 10.00 .
0 00 10.00 20.00 30.00
TIME IN MIN.
Figure 3.6. First order deadbeat algorithm with all ringing
poles removed (Equation 3.13).
z-Plane

Imaginary

0.86 682
919

Figure 3.7. Pole-zero location for Equation 3.9.


Table III-4

Deadbeat Algorithms for

Second order Plus Dead Time Model

Removing z = -1:

-aT -1N .. -bT -1N


D(z) - Cl - e z )(1 - e ) (3 .14)

2KC1(1 - z"1)Cl + 57 z_1>

or
__ 1 r„ / -aT , -bT. -(a+b)T i
M = IE - (e + e )E ,+ e E 0
n 2KC^ L n n-1 n-2J

Removing both z= "^/C^ an(* z=

D(,J = Cl - e ' " . ' 1) (1 - e ^ V 1)


2K(Cj+C2) (1 - z’1) (3'16)
or

«n - 2K(C*+ C2) [En-(e'aT+ e'M ) E n-l+

(3.17)
60

O
o

JUlXLTur-u/

- 10.00 0.00 1 0 .0 0 20.00 30.00


T I M E IN MIN.
o
o
in
co,

o
oo
- 10.0 0 0.00 10.00 20.00 30.00
T I M E IN MIN.
Figure 3.8. Second order deadbeat algorithm with z=»l removed
(Equation 3.15).
o
C3
ID.

w
o
X q

CCo
IxJ -
0.2 -

cn
a
2 S

- 1 0 .DO 1 0 .0 0 20.00 30.00


T I M E IN MIN
e>
Cl
in
o>,

Z
I— <
o
o

a
CO
- 10.00 10.00 20.00 30.00
T I M E IN MIN.
Figure 3.9. Second order deadbeat algorithm with all ringing poles
removed (Equation 3.17)>
192
second order PID first order dead­
beat (Equation 3 .1;
190.00
IN F.

second order deadbeat


186.00

(Equation 3.17)
TEMP
186.00

0.00 1 0 .0 0 30
TI M E IN MIN.

Figure 3.10. Comparison of algorithms for load change.


195.00
190.00
IN F.

PID second order


185.00

second order deac


TEMP

beat (Equation ^
3.17)
180.00

first order deadbeat


(Equation 3.13)

10. 0.00 10.00 S O . 00 30


TI M E IN MIN.

Figure 3.11. Comparison of algorithms on set point change.


as compared to those with ringing poles removed, as shown in

Figure 3.12. That is, in the case of a set point change the

controlled variable responds almost identically, but this is

not quite true in the case of a load change.

Kalman Algorithm (5)

Instead of specifying C(z)/R(z) as in the previous section,

Kalman obtained an algorithm by requiring the response to a step

change in set point to reach the final value in two sampling

times and remain at the final value thereafter, as illustrated

in Figure 3.13. Therefore, the expression for C(z) is

o/ \ -1 , -2 , -3 ,
C(z) = c^z + z + z + ...

No restrictions are placed on the value of c^. In order

to accomplish the above, the manipulated variable M(z) will

assume two intermediate values and then assume its final

values thereafter, as shown in Figure 3.13. The expression

for M(z) is

-1 -2 -3
M(z) = m + m,z + m_z + m_z + ...
o 1 f f

Where m^ is the final value which is the reciprocal of the process

gain in a linear system.

For a unit step change in set point,


Q
o
cu
03
* -*

second order deadbeat ringing pole removec


O
O (Equation 3.17)^7

second order deadbeat(Equation 3.10)

CD
D
to
lO.
- 10.00 0.00 10. D U
TIME IN MIN

Figure 3.12. Second order ringing and ringing-free algorithms


on load change.
time

m
o

1 1
| |
1
n»(t) . 1 1
I
ml 1 1 1
i 1
. 1 i
0 T 2T 3T 4T

time

Figure 3.13. Desired response characteristics used to design


Kalman's algorithm.
o/

R(z) = —
1 - z’1

c (z) = /i - z
= (1 X c ^ “I j+. z
"2 +, z
-3 +, ...).

= c^z -1 +, ,, - c^)z
v -2
(1

= P jZ -1 + P2z~2 = P (z) (3 .1 8 )

and

M(z) _ „ -lw . -1 -2 -3, v


= (1 - z )(n^+ m^z + mfz + mfz +...)

= q0 + + q2z "2 = Q(z > (3 .1 9 )

From Elgure 3.1, the process pulse transfer function HG(z)

is the ratio of C(z) to M(z), i.e.,

HG(z) = = ^X.2) (2 20)


} M(z) Q(z)

Therefore, it is obvious that the coefficients in P(z) and Q(z)

must equal to the coefficients in the process pulse transfer

function HG(z).

The following relationships among the coefficients should

be noted:
2

i? ! Pi = + P2 = cx + ” ci^ = 1 (3 .2 1 )

2
- - - + m, - m + m - m, = m_ = 1/K
i=l\ - 9o + ^1 + ^2 " mo + ml" V V ml “ ”f
(3 .2 2 )
where K is the process gain for a linear system. Although these

relationships do not directly hold for the pulse transfer

functions in Table III-l, simply dividing by the sum of the

numerator coefficients will give the desired results.

Now that P (2) and Q(z) are known, the control algorithm

D(z) can be derived from equation 3.3.

J)(z) s — -—
' HG(z) 1-C(z)/R(z)

o Slzj PCz)
P(z) l-P(z)

D(z) <3 -23>

Hence, the coefficients in the control algorithm are directly

related to the coefficients in the process pulse transfer

function.

Two intermediate switches in the. manipulated variable

were assumed in the derivation of equation 3.21. It was found

that Q(z), the denominator of HG(z), must contain terms through


-2
z . In other words, the plant should be second order, since

a third order process would require three switches.


-2
If the process contains a dead time z , equation 3.23

still holds. That is, the process dead time presents no problems

in deriving the control algorithm.


The controller D(z) and the equivalent difference equations

for first and second order models are given in Tables III-5 and

III-6, respectively, with sampling time T = 1 min.

Comparison of First and Second Order Algorithms

Figure 3.14 and Figure 3.15 are responses to step change

in set point for equations 3.25 and 3.27, respectively. Even

though the controlled variable responses look pretty good, too

much ringing is experienced in both algorithms. Pole-zero

relationships for equations 3.24 and 3.25 are shown in Figures

3.16 and Figure 3.17, respectively. It is anticipated that both

algorithms will ring.

Tables III-7 and III-8 give the algorithms with ringing

poles removed for first and second order algorithms, and the

respective responses are shown in Figures 3.18 and 3.19. It

is obvious that the algorithms with ringing poles removed

perform very well. However, equation 3.31 happens to be a PI

controller with a gain of -354.0 and a reset time of 12.0.

Similarly, equation 3.33 happens to be a PID controller with a

gain of -531.0, a reset time of 13.49, and a derivative time

of 1.80.

Comparison of Kalman's first order model algorithm and second

order model algorithm are shown in Figure 3.20 and Figure 3.21

for load and set point changes respectively. Definite conclu­

sions can be drawn about the superiority of the second order


Table III-5

Kalman Algorithms for

First Order Model

a.T ,
1 (1 - e z )
D (z) = (3.24)
K(C6+C?) r -3 -4 i
C,z C,z
1 6 7
C,+C_ C,4C
6 7 6 /J

or
-ajT C
1 r_
M =
K(C6+C?) L^n c "n-lj ' C6+ Mn-3 + C,+C, Mn-4
o /

(3.25)

where

a = —
1 T
-a^mT
C6 ‘ 1 ' c
-a.mT -a T
C? = e - e

6
m
1 ” T

e = NT + 6

ana T = 1 mln = sampling time K = process gain

0 a dead time E = error

t ** time constant M = manipulated variable


Table III-6

Kalman Algorithm for

Second Order Model

i fi / -aT. “bT. -1 . - (a+b)T -2]


.1____ Ll ~ ( e + e )z + e z J
D(z) =
k (c 1+ c 2)
. C1 -2 C2 -3
■ C1+C2 z "Cj+Cg 2
(3.26)

or
1 r -aT, "bT._ . -(a+b)T „ i
n K(C1+C2)LEn- (e +e )En-l En-2]

c c
+ 5 ^ «n-2 + c j f q “a-3 <3 -27>

where

a = — 9 = dead time
T1

b-i- t^, t2 = time constants


t2

0 = NT K = processgain

T = sampling time *= 1 min E = error

D6 -aT — £L6-bT
= 1 + a - b M = manipulated variable

-(a+b)T . be ae
2 = e + a - b ----
/1.

o
o
in

CD
in_
•=*

- 10.00 0.00 10.00 30.00


TIM E IN MIN.
CD
O
in
03

CD

t— 4

a
--

o
oo
- 10.00 0.00 10.00 20.00 0.00
TI ME IN MIN.
Figure 3.14. First order Kalman algorithm (Equation 3.25).
CD
O

0.00 1 0 .0 0 20.00 30.00


TI ME IN MIN.
o
o
Ift

CD
O

CD,
- 10.00 0.00 1 0 .0 0 20.00 30.00
TIM E IN MIN.

Figure 3.15. Second order Kalman algorithm (Equation 3.27).


z-Plane

Imaginary

1.0

Real
382 923

309-*8161

Figure 3.16. Pole-zero location for Equation 3.24.


/3

z-Plane

919

-.5-.461

Figure 3.17. Pole-zero location for Equation 3.26.


/o

Table III-7

First Order Kalman Algorithm

With Ringing Poles Removed

Removing the complex poles

”al^
-1
D (z) - “ r f ™ (1 V E -1 :-! - (3-28)
.10JK. (1_z J-)(1+>382 z A)

or
_a x
“n ' l k K ' 6 1 En-l] + -618Mn-l + -382Mn-2 <3 ‘29>

Removing all the ringing poles

-aiT
_, „ 1 (1 - e z )
D <*> ■ K(3C6« C 7) . ,-1, (3-30)

or
“3 T
(3.31)
Mn = K(3C6+4C?) [En " 6 1 En-l] + Mn-1
Table III-8

Second Order Kalman Algorithm

With Ringing Poles Removed

Removing the complex poles

[l -
( ' " K(2C1+3C2)- - ( 1 . z-l)
(3.32)

or

Mn " ^ W ) [En- (e'aT'Hi'bT)En-l+ e-(att)TE ,1 + M ,


1 2 n-2J n-1

(3.33)

where

a = — 9 = dead time
T1
1
b = —- T|»To = time constants
2
0 = NT K = process gain

T = sampling time = 1 min e = error

_aT jjT M = manipulated variable


c , - 1 + be -■ ^
1 a - b

_ -(a+b)T be bT - ae aT
2 + a - b
78

o
CD
tn

CO
O
2 5

- 10.00 .
0 00 10.00 20.00 30.00
TTI M E IN MIN
CD
o
U3
05

CD
CD

CD
- 10.00 0.00 10.00 20.00 30.00
TIM E IN M N
Figure 3.18. First order Kalman's algorithm with all ringing poles
removed (Equation 3.31).
79

a
CD
to
O

CO
o
*- J—k

-10.00 0.00 10.00 20.00 30.00


T I M E IN MIN.
CD
a
to
o>

<— i
CD
O
cl : .

luS“

CD
o
o
DO
0.00 10.00 20.00 30.00
TI M E IN MIN.

Figure 3.19. Second order Kalman algorithm with ringing poles


removed (Equation 3.33).
ov

O
o
OJ
o>
Kalman's first ordc
(Equation 3.31).
PID second order
o
o
,o
ox

2
Kalman's second order
(Equation 3.33).

CJ
o
to

0.00 10.00 20.00 30.00


T I M E IN MIN.

Figure 3.20. Comparison of algorithms on load change.


U JL

U>
o?

PID second order

i— i
C3
* O
0.-; *

u j» -

C3
(Equation 3.33) (Equation 3.31)

CO
- 1 0 .0 0 0.00 10.00 20.00 30.00
T I M E IN MIN.

Figure 3.21. Comparison of algorithms on set point change.


model algorithm in which the rising time is faster, less over­

shoot, faster settling time and less oscillatory.

In many cases, algorithm designed for a set point change

perform very well for a load change. From Figure 3.22 it is

clear that, for a load change, the algorithm without removing

the ringing poles yields a better response compared to the

one with the ringing poles removed.

Summary

Algorithms derived from a second order model always

performed better than algorithms derived from a first order

model in either the deadbeat or Kalman method. Superiority

in the second order model algorithm over the first order

model algorithm is experienced in both load and set point

changes, whereas it is advisible to employ the original

algorithm for load changes.

In all cases considered in this chapter, removal of the

ringing poles from the algorithm designed from a first order

model produced a PI algorithm; removal of the ringing poles

from the algorithm designed from the second order model

produced a PID algorithm. This is surprising since one of the

advantages generally cited for use of the z-transform algorithms

is that they should give superior performance to that of PI and

PID algorithms.
09

o Equation 3.33)

»— i

(Equation 3.27)

CD
to
to
- 10.00 0.00 10.00 30
TIME IN MIN

Figure 3.22. Comparison of Kalman ringing and ringing-free


algorithms on load change.
Literature Cited

Kuo, B . C ., Analysis and Synthesis of Sampled-Data Control


Systems, Prentice-Hall, Englewood Cliffs, N.J., 1963.

Koppel, L. B., Introduction to Control Theory. McGraw-Hill


Book Company, Inc.

Smith, C. L., and Merrill, P. W., "The Dynamics of Spot


Samples," Hydrocarbon Processing, Vol. 46, No. 12, (Dec.1967);
Vol. 47, No. 1, (Jan.1968).

Dahlin, E. B., "Designing and Tuning Digital Controllers",


Instruments and Control Systems. (June 1968), p. 77.

Kalman, R. E., discussion following article entitled


"Sampled-Data Processing Techniques for Feedback Control
Systems", by A. R. Bergen and J. R. Ragazzini, Transactions
of the AIEE. (Nov. 1954), pp. 236-247.
Chapter IV

DAHLIN'S CONTROL ALGORITHM

Introduction

The deadbeat algorithm and Kalman's algorithm require the

process output to move from the old set point to the new set

point in one and two sampling times, respectively (Chapter III),

which is far too streneous for most industrial processes.

The problem is not with the design method, but with the chosen

expression for C(z), the process output. For both deadbeat

and Kalman algorithms there is no parameter that can be adjusted.

The Dahlin method has such a parameter and therefore may be more

attractive for process control.

Design of Control Algorithm Using z-Transform (1)

The general digital control loop is illustrated in Figure 4.1.

From block diagram analysis with negative feedback (2), the

controller D(z) is derived as

1 C(z)/R(z)
(4.1)
HG(z) 1-C(z)/R(z)
where

HG(z) = Z F - — - G(s) (4.2)


Zero-order
E(z) Controller Hold Plant
C(s)
R(z). -sT
D(z) 1-e
G(s)
Plant
"- Error
Set Output
Manipulated
Point
Variable

Feedback Signal C(z)

Figure 4.1. Typical digital process control loop.


is the pulse transfer function of the process. Several expressions

for HG(z) are given in Table ITI-1 in Chapter III.

Dahlin's Method

Dahlin (3) suggests that the closed loop system should behave

like a continuous first-order lag with dead time, i.e.,

Xe~08
c (s ) = —
s+X (4.3)

where X = reciprocal of the time constant.

The \ of the closed loop response can be best visualized

as a tuning parameter. Large values of X give tight control;

small values of X give loose control.

It should be pointed out that the process model dead time

6 should be included in equation 4.3. If situations such as this

are not recognized, the controller D(z) would require future

values of error in order to calculate the current value of the

output, which cannot physically be accomplished.

In discrete form
(1 - e”^T)z"N_1
C O O = <■' •li > -i (4.4)
(1-z ) (1-e KTz L)

where N = nearest integer number of sampling times in 6

T = sampling time

If R(z) is the unit step change, then

C(z) (l-e"XT)z~N ~1
(4.5)
R<z> " (l-e^V1)
Substituting into equation 4.1, we have

D(z) =
i . - M z -1- (1-e
1-e ,, -\T.)z -N-l HG(z)

(4.6)

Pulse transfer functions HG(z) for first order plus dead

time model and for second order plus dead time model are

derived in Chapter III.

The controller D(z) and the equivalent difference equations

for first and second order models are given in Table IV-1 and

IV-2 respectively for sampling time T = 1 min.

Handling of Saturation

Applying the control algorithms in Table IV-1 and Table IV-2

to the original process for a step change in set point for X=l,

the manipulated and controlled variables are as shown in

Figures 4.2 and 4.3. Physical constraints have been imposed

upon the system. Fully-opened valve position corresponds to

6,000 lb/min of cooling water rate, whereas fully-closed valve

position corresponds to no flow of cooling water. When saturation

occurs, the errors in Equations 4.8a and 4.10a are recalculated

from the control equation 4.8 or 4.10 .by employing either

equal to 6,000 or 0. This practice is analogous to the use

of the velocity form of the PID algorithm as opposed to the

position form.
oy

Table IV-1

Dahlln Algorithm for

First Order Model

-XT "ajT ,
d /8) * d - e XT) (1-e 1 z 1) ..
-XT -1 -XT. -31 fn _ -1. (4,7)
K |_l-e z -(1-e )z J (Cg+C^z )

1 -XT -a,T
M = [ V e ’1‘ V , ] +(e'XI-C7/C6)Mn.1

+ 57 •'X\ - 2 + <1- ” >,W C 7/C6 (1'e"XT>Mn-4 (4 -8>


6

KC
\ = [Mn-(e'"T-C7/C6)M„ - r C7/C6e' " \ - 2 - < 1-e'W )M„-3

->T -a-T
- C7/C6 (l'e )Mn-4] + 6 En-1 (4‘8a)

where
1
ai T
-a irfT
C6 = X - e

-a-rrfT -ajT
C = e - e
7

m = 1 - 6/T
90

Table IV-1 (cont,)

0 = NT + 6

T = 1 min = sampling time

0 = dead time

T = time constant

K = process gain

E = error

M = manipulated variable

X = tuning parameters
Table IV-2

Dahlln Algorithm for

Second Order Model

K [l-eW z'1-a-e' W )z'2 ] (Cj +Cj Z-1)

- (CJ /C1- - M ) ^ + r(c /c l)e- w +1] M


u / I * n-2

+ C2/Cl ( l - ’« ) H n_3

KC
En = 7 7 x f [ V (C2/Cr e'W ) M n - r [<C2/Cr l)e'XT+l]

- C2/c1(1V « ^ - 3] + (e-aT+e-bI)En.1-e-<a« ) IEr

where
92

Table IV-2 (cont.)

9 = NT

T = sampling time = 1 min

-aT -bT
C 1 + be - ae
'1 a - b

c = e-(a+b)T + be~bT - ae~aT


2 a - b

9 = dead time

= t:^me constants

K - process gain

E = error

M = manipulated variable

X = tuning parameter
o
o
in.
(M
o
X q
©

CC©
LJ '
0-S

1 0 .0 0 0,00 10.00 20.00


T I M E IN MIN.
CD
CD
in
o>

»—«

X=1

oo.
-10.00 0.00 1 0 .0 0 20.00
T I M E IN MIN.

Figure 4.2. First order Dahlin algorithm (Equation 4.8).


CD
O

u>.
r**

o g

Xt»-|
=*•

UJ
I—
CE^
tn°
l/5_| LnJlrun.
cc ~
UJ
f—
CCo
120
in
I
- 1 0 .0 0 0.00 1 0 .0 0 2 0 .0 0 30.00
T I M E IN MIN.
o
o
in
05

oo
- 1 0 .0 0 0.00 1 0.0 0 20.00 30.00
T I M E IN MIN.
Figure 4.3. Second order Dahlin algorithm (Equation 4.10).
Ringing Analysis

Judging the performance of a controller merely from the

response is inadequate. The manipulated variable (valve signal)

m(t) in Figure 4.2 and 4.3 exhibits excessive valve movement, a

phenomenon known as ringing. The ringing amplitude is defined

as the difference between the first two controller outputs,

and can be related to the algorithm's pole-zero locations.

First Order Lag Plus Dead Time Model

Rewriting equation 4.7

(1-z’1)(l+(l-e"XT)z-1+(l-e"XT)z“2) (1+ )

(4.11)

where
-\T

The two varying conjugate complex poles are

-(l-e~XT)± iV/£a-e"XT)-(l-e"XT)2
z (4.12)
2

and the magnitude is


It is very interesting to note that deadbeat is a special

algorithm of equation 4.11 when X approaches infinity. The pole-

zero locations are shown in Figure 4.4 for \=1 and (deadbeat).

Ringing is anticipated and the tighter the control, the greater

the ringing amplitude as the result of equation 4.13.

Poles on the left half plane contribute ringing (3).

For the most part, the ringing phenomenon is generally super­

imposed upon the useful output of that algorithm. Dahlin

suggested that the ringing pole be simply eliminated and the

gain adjusted accordingly.

The various algorithms with ringing poles removed are

shown in Table IV-3. Since the magnitude of the complex

ringing poles dominate, the complex poles will be removed

first, which considerably improves loop performance as shown

in Figure 4.5. As Cy/Cg <»5, it does not contribute too much

ringing, the algorithm with all the left half poles removed

(Figure 4.6) behaves only slightly better than Figure 4.5.

This algorithm happens to be a PI controller with a gain of

-325.0 and reset time of 12.0.

Second Order Lag Plus Dead Time Model

;Rewriting equation 4.9

(4.18)

1
s/

z-Plane

Imaginary

X=1
\=

-.412 923 1.0

Real

X=1

Figure 4.4. Pole-zero location of Equation 4.8.


Table IV-3

First Order Dahlln Algorithm

With the conjugate complex poles removed

-IT **al^ -1
D(z) = (1 -,.e„XT)(l - e .z )
-1 ^7 -1
KC6 (3-2e KL)(l-z ) (1+- (T * A)
6

-XT -a,T
1-e
M =
n
KC6 (3-2e"X’T)
[ V 6 1 Vl] +(1-°7/C6>Mn-

^7
+ 7T M o
6 n"2

With all the ringing poles removed


Table IV-3 (cont.)

where

-a.mT
C6 ’ 1 - 0

-a.mT -a.T
C? = e - e 1

m = 1 - 6/T

e = NT + 6

T = sampling time = 1 min

0 = dead time

T = time constant

K = process gain

E = error

M = manipulated varialbe

\ = tuning parameter
IUU

o
o
in

OCo
UJ •
0-2“

- 10.00 0.00 10.00 20.00 30.00


TI ME IN MIN.
o
o
LO
CD.

X=1

a
co
-10.00 0.00 10.00 30.00
TI ME IN MIN.

Figure 4.5. First order Dahlin algorithm with complex ploes


removed (Equation 4.15).
o
o
LO

X o
o
0
in.

OCo
LU •
CUS­
CO
Q
2 5

10.00 0.00 10.00 Z0.00 30.00


T I M E IN MIN.
o
o
in
o>

21
V—I

X=1

o
a
03

-10.00 0. 00 10.00 • 20.00 30.00


T I M E IN MIN.

Figure 4 .6 . First order Dahlin algorithm with all ringing poles


removed (Equation 4.17).
where
1 - e’XT
K" = 6
“ l

There is one real varying pole on the left half plane

z = -(1 - e’XT) (4.19)

As o o ., the algorithm becomes the second order deadbeat

algorithm in Chapter III. The pole-zero locations are shown

in Figure 4.7 for X=1 and X=0° . Ringing is anticipated as


-XT
poles are located at -0.855 and -(1-e ).

The various algorithms with ringing poles removed are shown

in Table IV-4. From Figure 4.8, considerable ringing still

exists with z= -(1-e removed. Algorithm with z=-C2/C^

removed behaves better than Figure 4.8 as 02/0 ^ = .855 and


-IT
(1-e ) = .632, as shown in Figure 4.9. Algorithm with both

the ringing poles removed works beautifully as shown in

Figure 4.10. However, this happens to be a PID controller

with a gain of -507, a reset time of 13.49, and a derivative

time of 1.80.

Ringing and Ringing-Free Algorithms

Algorithms derived from set point change can be used for

load change as in the case of Kalman and deadbeat algorithms in

Chapter III. Figures 4.11 and 4.12 are responses for first and
z-Plane

Imaginary

-.855 919 0
Real
682

Figure 4.7. Pole-zero location for Equation 4.9.


iwo

Table IV-4

Second Order Dahlin Algorithm

With z = removed

i "XT /1 -aT -1 .. -bT -1


n f z) = _____ (1-e z )(l-e z_J_ (4.20)
K(C1+C2) (1 - z " 1) (l+(l-e"XT)z"1)

m - 1-e~XT |\, / _aT . -bT. „ . - (a-t-b)T


n K(C1+C2) I n ” n-1 En-2]

+ e"XTM ,+(l-e”*-T)M 9 (4.21)


n-1 n-2

with z = -(l-e”^X) removed

, XT ,, -aT -1W 1 -bT -1N


D«) = .I > (1-e .« > (4.22)
KC1(2-e A i ) (1-z i)(l-K]2/C1z 1)

M = -e
l-eXT r„ , -ax. -oi\„
r. ./.-aT^-bT^ , -(a+DjT„
^-(a-A)!, t
— — ^ [ E n-(e +e ^ n-1 V l ]
n KC

(4.23)
+ ^ - V W l W n - a
Table IV-4 (cont.)

“XT
with z 5 -(1-e ) and removed

1 n -aT -1. -bT -1N


D <Z> " - L ^ S <1-e Z >P~e Z >
K(C1+C2)(2-e Ai) (1 - z )

-XT
M„ L ~-- |E -(e'aT+e‘bT)E .+e-<a+b>TE
“ k (c 1-
k ;2) (2-e XT) L n n '1
o
o

tn,
p-

(M

2 §
Xin4
sa*

Lul
h-
CEO
0Cl
in.
i
QC-
LU

<Xo
3 0
in
i
10.00 0.00 10.00 20.00 30.00
TI M E IN MIN.
o
o
in

o
o
GO
- 10.00 0.00 10.00 20.00 30.00
TI M E IN MIN.

Figure 4.8. Second order Dahlin algorithm with z^/l-e-^ )


removed (Equation 4.23).
107

o
o
in.

UJ

- 10.00 0.00 1 0. 0 0 20.00 90.00


TIME IN MIN.
o
o
in
03

UJ
U_

03
- 10.00 0.00 10.00 20.00 90.00
TIME IN MIN.
Figure 4.9. Second order Dahlin algorithm with z=-C0/C, removed
(Equation 4.21). 2 1
I v/o

o
o
u>

- 1 0.0 0 0.00 1 0 .0 0 20.00 30.00


T I M E IN MI N
o
Q
io

o
o
o
05
UJ
Ll.

o
CD

Q
CO
- 10.00 0.00 10.00 20.00 30,00
TI M E IN MI N
Figure 4.10. Second order Dahlin algorithm with all ringing
poles removed (Equation 4.25).
- 1 0.0 0 0.00 1 0 .0 0 20.00 30,00
TIM E IN MIN.
o
o
Cvl
Oi

=1
190,00
IN PER*
186.00
TEMP
85.00

- 1 0.0 0 0.00 10.00 20.00 30.00


TI M E IN MIN.

Figure 4.11. First order Dahlin ringing algorithm (Equation 4.8)


on load change.
o
o
in

in
- 1 0.0 0 0.00 10.0 0 20.00 30.00
T I M E IN MIN.
Cj
O
rO
cn

to
03
0.00 1 0 .0 0 20.00 30.00
TI M E IN MIN.
Figure 4.12. Second order Dahlin ringing algorithm (Equation 4.10)
on load change.
second order algorithms respectively without the ringing poles

removed. It is obvious that ringing is not significant in case

of load change as the error imposed is gradual in contrast to

the set point change. From Figure 4.13 and Figure 4.14, we

notice that the ringing algorithms give a faster rise time, less

overshoot, smaller settling time and less oscillatory compared

to the ringing-free algorithms. It is interesting to note that

it is more so in the first order algorithms rather than the

second order algorithm, probably because the first-order

ringing-free algorithm is only a PI controller whereas the second

order ringing-free algorithm is a PID controller.

Compensation of ringing at the expense of response has to

be justified in a particular situation. In this system it is

advisible to use the ringing algorithms for the load change and

the ringing-free algorithms for set point change.

Tuning Procedure (3)

As mentioned before X is a tuning parameter in Dahlin1s

algorithm and deadbeat is a particular algorithm when X tends

to infinity. The larger value of \ the tighter is the control.

This is obviously illustrated in Figure 4.15,4.16,4.17, and 4.18.

Higher X is desirable but highest possible X is recognized by

overshoot cauccd by mismatch.between actual and assumed

parameters, nonlinearities and ignored dynamic characteristics.


i
.lZ

132,00

Equation
190.00

4.17
IN F.

Equation 4.8
166.00
TEMP
166.00

10.00 0.00 10.00 e o .o o 30.00


TIME IN MIN.

Figure 4.13. First order Dahlin ringing and ringing-free


algorithms on load change (X=l).
e>
o
to
co

CO,

*—I Equation 4.8

Equation 4.17

co,
- 10.00 0.00 10.00 30.00
T I M E IN M I N

Figure 4.13a. First order Dahlin ringing and ringing-free


algorithms on set point change (X=l).

o

CVJ
05

Equation 4.25
O
Q
05

Equation 4.10

to

10.00 0.00 20.00


T I M E IN MIN.

Figure 4.14 . Second order Dahlin ringing and ringing-free


algorithms on load change ( 1=1 )
o>

Equation 4.10

Equation 4.25

o
00.
- 1 0 .0 0 0.00 1 0 .0 0 30,00
T I M E IN M I N

Figure 4.14a. Second order Dahlin ringing and ringing-free


algorithms on set point change (\=1)•
132.00
100.00

1.0
IN F.

\=
186.00
TEMP
186.00

1 0 .0 0 30.00
T I M E IN MIN.
Figure 4.15. First order Dahlin algorithm (Equation 4.17) with
varying X on load change.
CD
O
CM

O
=1
O
07
TEMP .IN F

\= 5

o
CO
ID.
10.00 0. DO 10.00 20.00 30.00
TI M E IN MIN.

Figure 4.16. Second order Dahlin algorithm (Equation 4.25)


with varying \ on load change.
o
o
to
o>.

o
o
O
CD

\=5

CO,
- 10.00 0.00 10.00 20.00 30.00
T I M E IN MIN.

Figure 4.17. First order Dahlin algorithm (Equation 4.17) with


varying \ on set point change.
o
o
in
o>.
IN F
TEMP

,=l

X=5

o
00,
10.00 0.00 10.DD
T I M E IN MIN.
Figure 4.18. Second order Dahlin algorithm (Equation 4.25) with
varying ^ on set point change.
Gradually Increasing X and observing response to small set point

changes yields the optimum value. A one parameter

optimization can be made to obtain the desired response

characteristics for the particular application (4). Out of

the three values of X in this study X=5 looks the best.

Summary

From Figures 4.19 and 4.20 it is obvious that algorithms

derived from the second order model are superior to those

derived from the first order model. The second order algorithms

give shorter rise time, less overshoot, and faster settling

time compared to those of the first order. Comparison with

the standard PID algorithm is also shown.

Algorithms designed for set point change perform very

well for load disturbances, but the ringing versions perform

noticeably better than the ringing-free versions.


192.00

Equation
190.00

4.17 A
IN F.

/ PID
Equation 4.25
186.00
TEMP
85.00

.
0 00 10.00 30.00
T I M E IN M I N

Figure 4.19. Comparison of algorithms on load change =1)


03
130.00
IN F.
185.00
TEMP

PID

Equation 4.25
Equation 4.17
180.00

- 10.00 0.00 10.00 20.00


T I M E IN MIN.
Figure 4.20. Comparison of algorithms on set point change (X-l)•
Literature Cited

1. Kuo, B. C., Analysis and Synthesis of Sampled-Data Control


Systems, Prentice-Hall, Englewood Cliffs, N.J., 1963.

2. Koppel, L. B., Introduction to Control Theory. McGraw-Hill


Book Company, Inc.

3. Dahlin, E. B., "Designing and Tuning Digital Controllers",


Instruments and Control System (June 1968), pp. 77.

4. Wilde, D. J. and Beightler, C. S., Foundations of Optimization,


Prentice-Hall, In., Englewood Cliffs, N. J.

(
Chapter V

ADVANCED CONTROL TECHNIQUES

Introduction:

The fact that a digital control system Is more costly than

a conventional analog control system is generally accepted

today. In perhaps the majority of the cases, the justification

for this extra capital outlay comes from improved control

using advanced control techniques that are either impossible

or impractical to implement under the hardware constraints

of the analog system. This chapter considers the use of cascade

control and multivariable control systems on the system dis­

cussed in the previous chapters. The results obtained with these

techniques are compared to those obtained with the simple feedback

scheme examined in earlier chapters. All control algorithms

considered will be digital with a sampling time of one minute.

Cascade Control Systems (1)

Suppose we have a process essentially consisting of two

lags in series as illustrated in Figure 5.1. The usual feedback

arrangement \tould be to measure the output and adjust the

input as illustrated in Figure 5.1a.


Controller
s+1 t 0 s +1

Figure 5.1a. Classical feedback

Master Slave Process 1 Process 2


ontroller ontroller
T.s+l

Figure 5.1b. Cascade Control.

Process 1 Process 2

Controller

Figure 5.1c. Equivalent feedback loop.

Figure 5.1. Cascade control system.


Assuming the dynamics of Process 1 are somewhat faster than

that of Process 2, i.e., is smaller than T2 , the performance

of this control system can be improved by somehow making

Process 1 even faster dynamically. One approach by which this

can be accomplished is to add a feedback loop around Process 1,

giving the cascade arrangement in Figure 5.1b. The output of

the outer or master controller is the set point to the inner

or slave controller.

Assuming a proportional controller Kc is employed in the

slave controller, applying the principle of negative feedback

to the inner loop (2).

Since t./(1+K )is less than -r,> Process 1 becomes faster


JL C J*

due to cascading as shown in Figure 5.1c.

This approach to the development of the cascade control

concept was chosen in an attempt to show that cascading

improves the basic performance of the loop. The real payout

comes when the disturbances are considered, as illustrated in

Figure 5.2. Some disturbances enter the inner loop, and others

enter the outer loop. It is obvious that compensation of those

entering the inner loop is greatly improved, but cascading also


Master Slave

Controller ontroller Process 1 Process 2

Figure 5.2. Cascade control with different entry of disturbance.


improves the compensation of disturbances entering the outer

loop. Utopia is when all major disturbances enter a fast

inner loop.

Cascade Control of Process

Figure 5.3 illustrates the cascade control of the process

studied in the previous chapters. The jacket outlet water

temperature is the intermediate variable. The main obstacle to

obtaining good performance from a cascade control arrangement

is the tuning of the two controllers. The basic approach is

simple (3):

1. Place the master controller on manual and tune the

inner loop using the method of your choice (4-8). As

the inner loop must respond well to the set point

changes from the master controller, it should probably

be tuned to give good set point response.

2. Place the inner loop on automatic and tune the outer

loop. It should be noted that the "process", i.e.,

the entire entity controlled by the master controller,

contains the slave controller. It is suggested that

the settings for the slave controller not be changed

while tuning the master controller.

Tuning is certainly a factor to consider in the selection

of the modes for the various controllers. For the inner loop,

a pure proportional controller is frequently recommended as

offset is not significant for the intermediate variable (water


(set point)
Feed

Master '.TRC jacket


overflow
i (set point)

Slave (TRC O

Products
cooling
water

Figure 5.3. Cascade control of a stirred jacketed reactor.


outlet temperature). Use of proportional action in the inner

loop makes the inner loop faster, which should lead to an easier

tuning problem for the outer loop. Reset is sometimes added

but rarely is derivative found to be beneficial. For the outer

loop, a PI controller is typically used, with a rate added in

some cases.

Inner Loop Tuning

Minimum ITAE criterion ( 9) is employed in tuning for set

point change (IQ). In order to tune the slave controller a first

order plus dead time model is fitted to the response of the outlet

water temperature to a step change in water feed rate . A two-

parameter search is made with the Pattern Search technique (11)

to minimize the ISE criterion. In other words, least square

fitting is employed. The model is found to be

re \ - -0.0237 e~°,34s .
G(s) " 4.67 s + 1 (5'2)

Comparison of the true response and the one by the model is

shown in Figure 5.4. The optimum proportional setting was

evaluated as in Chapter II, and found to be K = -1324. The


c
jacket temperature response is shown in Figure 5.5.

Outer Loop Tuning

A first order plus dead time model is fitted to the response

of the reactor temperature to a step change in the set point

for the inner loop with the slave controller (proportional) in


1 22 .0 0
120.00
RESPONSE CURVES

True response

Model 5.2
116.00
IN F.

- 300
116.00
TEMP
1 R .00
112.00

00 10.00 30.00 50.00


T I M E IN MIN.

Figure 5.4. Model fitting of jacket temperature with slave


controller in manual.
o
o
rvi

o
o
CM-

Q
CO
U_

C.00 10.00 20.00 30.00 50. D D


TIME IN MIN.

Figure 5.5. Jacket temperature response with slave controller in


automatic.
automatic. Another two-parameter search is made to minimize

ISG. The model is found to be

, 0.3234 e"°-94s
G(s) ° ~ 8~ 8"s +"l <5 '3>

The model matched the true response very well, as can be

seen in Figure 5.5a. The model and tuning parameters that

minimize the ITAE for both PI and PID control (12) are given in

Table V-l.

Cascade Responses

Step change in feed rate (disturbance entering outside the

inner loop) is simulated when both the master and slave

controllers are in automatic. The response is compared to

the one with PI algorithm tuned to a second order plus dead time

model in Chapter II, as shown in Figure 5.6. Similar results

are shown in Figure 5.7 for temperature set point change. As

is mentioned before, cascade control is best appreciated when

disturbance enters the inner loop. A step change in inlet water

temperature is simulated as shown in Figure 5.8. It is noted

that cascade response yields shorter rise time and less

overshoot.

From Table V-2, it is seen that a PI cascade control

behaves much better than a simple feedback loop with a PI

controller tuned on a second order model. Although PID cascade


o RESPONSE CURVES
o
05,

ll_
True response

Model 5.3

ATSP = 7 . 1

00-

o
o
(0
03
0.00 1 0 .0 0 30.00 50.00
T I M E IN MIN.

Figure 5.5a. Model fitting of reactor temperature with slave


controller in automatic.

i
Table V-l

Cascade Control Models and

Algorithm Settings

Inner Loop Model

' . -0.0237 e"°*34s


G(s) =■
4.67 s + 1

Proportional controller Kc= - 132.4

Outer Loop Model with Slave Controller In Automatic

03234 e“0*9^8
G(s) = 6
8.68 s + 1

Setting for Load Change

PI Kc 12.678 ph , Kc = 9.0

Tj*3 3.661 T± = 1.5

Td = 0.1

Setting for Set Point Change


o JPI
FI cascade
IN
TEMP

to

o
to
to-
0.00
T I M E IN MIN.

Figure 5.6 . Comparison of cascade PI and conventional PI controller


for feed rate change.
OJ,

•— I
o
I cascade

PI

a
to-
1Q.QD 20.00
TIME IM MIN.
Figure .7. Comparison of cascade PI arid conventional PI controller
temperature set point change.
ea
o
m
CM
o>

PI

PI cascade

- 1 0 . DO 0 .00 30. 00
T I M E IN M I N .

Figure 5.8. Comparison of cascade PI and conventional PI controller


for step change in cooling water temperature.
Table V-2

Response for Cascade and

Standard Controllers

Load Change: Am = +200 ITAE

Cascade FI 58.8

Standard PI 189.3

Cascade PID 35.7

Standard PID 38.9

Load Change: ATw= +10

Cascade PI 9.4

Standard PI 110.0

Cascade PID 18.6

Standard PID 23.4

Set Point Change: ATSP = -5

Cascade PI 107.5

Standard PI 163.6

Cascade PID 56.6

Standard PID 74.6


control behaves better than a simple feedback loop with a PID

controller, the Improvement is not as significant as with PI.

It must be borne in mind that the cascade tuning is based

on a first order model whereas the simple feedback tuning is

based on a second order model. Furthermore, PID tuning is much

more difficult than PI. Therefore, utilization of a higher

order model to improve the cascade PID tuning would probably

yield much better performance.

Multivariable Control (13.1Z^

Up to this point the control loops were designed to maintain

one variable (reactor temperatue) at a desired value (the

set point) by manipulating one process input (cooling water

rate). In fact many practical control systems do not exactly

fit this description. This section considers a control system

for maintaining both the reactor temperature and concentration

near their desired values by manipulating both the cooling water

rate, Wc » and the reactor feed rate, W. Every controlled

variable needs a manipulated variable (15) and is thus a multi-

variable control system.

Previously the load change to the system was the change

in feed rate, which is now a manipulated variable. Instead,

the change in concentration of the feed stream is taken as the

load change to the system. Using the PID algorithms in Chapter II,

the responses are shown in Figure 5.9 for a step change in feed
to
CD
cn

li­

en

- 1 0 .0 0 1 0 .0 0
T I M E IN MIN.
tZJ
ta

CD

to
to
cm
CD.

to
to 0.5
af
CD

10.00 0.00
T I M E IN MIN.

Figure 5.9. Response for step change in feed concentration.


3 3
concentration from 9.0 lb/ft to 9.5 lb/ft and Figure 5.10

for step change in set point from 190°F to 185°F. Since it is

a self-regulating system, the concentration seeks another


3
steady-state which is quite different from 3.6 lb/ft . In most

cases where the concentration is to be controlled as well as

the temperature, multivariable control comes into play.

Gain Matrix

It is obvious from Figures 5.9 and 5.10 both the temperature

and concentration are functions of cooling water rate W c and

feed rate W.

Hence T= ^ ( W ,W) (5.4)

ca= f2 < V W) (5’5)

One aspect of the design of a multivariable control system is

to decide which manipulated variables (W, or W) will control

which controlled variables (T and C ). Since interaction exists


ci
in the system, it is a matter of common sense to say that the

manipulated variable which has the most influence on the con­

trolled variable should be paired. This is where the gain

matrix plays a significant role. Expressing equations 5.4 and

5.5 as changes about the operating point gives


s
cn

23
Ota
Oin
m
1 0 .0 0 10.00 30.00
T I M E IN MIN.
o
ca
in
07,

ATSP “ -5

10*00
T I M E IN MTN
Figure 5.10. Response for step change in temperature set point.
The matrix representation of these equations is
I

“ br bT
m
8* c
W W_
(5.8)
be
AC __a m
a
w VT
o

■» * m «

or in matrix form

C = M m (5.9)

£h
where i,j element of M is bC^/brn^ J This is typically

represented schematically in Figure 5.11.

The matrix of the partials in equation 5.8 is called the

gain matrix, each element being equivalent to the process gain

in a single-input, single-output control loop. For the two-

input, two-output system there are four gains; for an n-input,


2
n-output system there will be n of them.

These gains could be determined experimentally with n

process tests, e.g., impose a step change in one manipulated

variables while holding all others constant. The response of

each controlled variable could be recorded for evaluation of the


Figure 5.11. Schematic interacting action of multivariable
control system.
partial of each controlled variable with respect to the manipulated

variable for which the step change was made.

That is, suppose we place all control systems on the reactor

on manual and impose a change AW on the feed rate into the reactor

whlile holding W constant. The resulting steady state change in


c
C& and T, say, A a n d AT, can be obtained from the response of

the unit as shown in Figure 5.12. Since W has been held


c
constant, the partials with respect to W can be evaluated as

follows:

bC | AC
aI a
■ I, m <5ao)
Wc

AT
- W (5.11)
W
c

The"approximately equal to" notation is necessary because

most systems are nonlinear which means that the final answer

will vary somewhat with the magnitude of Small changes in

the manipulated variable should be undertaken. In the

limiting case, the above approximations become more exact.

Similarly from Figure 5.13, &Ca /8Wc J ^ , bT/tWc | w can be

evaluated.

During the calculation of the gain matrix, the units utilized

determine the final number obtained. Thus it is not wise to

compare elements of the gain matrix to decide the pairing of


- 20.00 0.00 20.00 10.00 60.00
T I M E IN MIN.
CM
o>,

,o
JN F

o
TEMP

tO
m * -loo

Vi

T I M E IN MIN.
Figure 5.12. Open loop response to step change in feed rate
o

u_

a.o
lO
*3»- “■

- 20.00 .
0 00 HO. 00
T I M E INI MIN.
o
o
CVJ
03

- 300
u>

CQ.OQ
T I M E IN M IN .

Figure 5.13. Open loop response to step change in cooling


water rate.
variables. From Table V-3, bT/tHf| ^ is much greater than
c
6T/6tfc | yet it will be wrong to pair the temperature with the

feed rate. Therefore, the gain matrix is not unique in pairing

variables.

Bristol (16) introduced the technique of normalizing the gain

matrix making it independent of units and other considerations.

It has been proven to work in several industrial applications

(17,18).

Relative Gain Matrix

For multivariable systems, the above gains are not the only

ones that can be defined (19). Suppose a step change in the

rate of is made in the configuration in Figure 5.14. As

the control system will vary the reactor feed rate W to maintain

constant concentration C , the change in temperature T due to a


£L
change in W c will not be given by tfT/tWc J as developed above.

Instead, this test can be used to evaluate another gain, namely

bT/£W I p , the change in temperature with respect to jacket


c ! ca
flow rate at constant reactor concentration.

Obviously one can evaluate three more gains of this type,

namely The notation for these

will be The relative gain matrix is defined as


&C.,
150

Table V-3

Gain Matrix and Relative Gain Matrix

Gain

- I “ - 0.0101 °F/pound per min.


c IW

.000078 pound/pound per min.


W

.020869 °F/pound per min. feed


W

.001256 pound/pound per min. feed


W

Gain Matrix
W w
c
-.0101 .020869
M =
.000078 .001256

Relative Gain Matrix

W W
c
887 .113"! T

2 'I 113 •887J C.


twC _

8
REACTOR a

I
Controller

Figure 5.14. Configuration to evaluate b T / W


In principle, experimental tests (Figure 5.9) analogous

to that illustrated in Figure 5.14 could be used, but such tests

are not really very convenient. Once the response'matrix or

gain matrix is obtained, the relative gain matrix is evaluated

by matrix manipulation (16). Denoting the gain matrix by M, an

intermediate matrix L is formed by inverting and then trans­

posing the gain matrix

L = (M'V (5.13)

The relative gain matrix U is determined by multiplying

corresponding terms of matrices M and L. The relative gain

terms are the elements of the matrix and are found from:

uij = ^“i j ^ i j * (5.14)

Significance of the Relative Gain Matrix

The utility of the relative gain is that it indicates how

manipulated and controlled variables should be paired. For

example, should the rate of cooling water to the reactor

be used to control reactor temperature or concentration? In

some cases the pairing is obvious, but in some cases it is not.

Occasionally it makes no difference (equal coupling), while

it is also possible for the proper pairing to change with operating

level. In these cases, the relative gains can be indispensable.


To facilitate the pairing of manipulated variables and

controlled variables, it is convenient to arrange the relative

gains as follows:

2 n

In
c
2 2n

c
n nn

The pairing is via the combinations whose relative gains

are the largest positive values. For example, the manipulated

variable for c^ would be the one corresponding to the largest

positive gain on the first row of the relative gain matrix.

A useful property of the relative gain matrix is that each

row and each column sums to unity. Thus, in a two-by-two

matrix as in this case, only one relative gain must be

determined.

The gain matrix and the relative gain matrix for this

system is shown in Table V-3. It is obvious that the reactor

temperature and the cooling water rate should be paired.

Tuning

Even though the above analysis indicates the proper pairing

of controlled and manipulated variables, connecting simple


feedback controllers often proves to be unsatisfactory. In

tuning the controllers, the general approach is to first

tune one of the loops and then tune the other. Generally no

problems arise in tuning the first controller (Chapter II), giving

the configuration in Figure 5.15a. In this configuration, the gain

of the system is . The second controller is tuned

with

1. The First Controller In Manual (FCIM), in Figure 5.15b.

2. The First Controller in Automatic (FCIA),in Figure 5.15c.

The gain of the first loop is now different due to the

feedback of the second loop. The value of the first loop gain

is now bc^/bm^| c . If this value is significantly different

from the original gain for which the loop was tuned, its per­

formance deteriorates significantly. These loops are then said

to be interacting.

Hence relative gain is a measure of interacting

in multivariable control. Thus, the relative gain matrix is

called the "interaction measure" by Bristol(16). In this

system, if the concentration of the reactor feed increases,

the composition controller would decrease the reactor feed

rate. This would tend to raise the reactor temperature (feed

temperature is below reactor temperature), calling for an increase

in the rate of cooling water. But this temperature rise would


Controller
SYSTEM

Figure 5.15a. Tuning of controller 1.

Controller

SYSTEM

Controller

Figure 5.15b. Tuning of controller 2 with First Controller


in Manual (FCIM).

ontroller
k > 4

SYSTEM

"2
Controller

Figure 5.15c. Tuning of controller 2 with First Controller


in Automatic (FCIA).

Figure 5.15. Two different means of tuning multivariable control


systems.
increase the reaction rate* calling for more feed to maintain

the composition. Thus, a control action in one loop is a

disturbance to the other. The two kinds of tuning mentioned

above are studied.

Tuning Concentration Loop (FCIM Tuning)

In this case it is an ordinary tuning for the concentration

loop ( 9,12). An underdamped second order is approximated to

the concentration response to a step change in feed rate. Using

least square criterion gives the following model:

0.001256
0 0 " UT s +v 2 t£ s + 1 ' <5-«>

where

t - 2.361

c = 0.772

fits the actual response quite well as shown in Figure 5.16.

The standard FID settings determined from the model are given

in Table V-4.

Tuning Concentration Loop (FCIA Tuning)

In this case, the response of the concentration to a step

change in feed rate (W®1200 pound per min.) is fitted to an

overdamped second order plus dead time model. Using the same

procedure as above gives the model


RESPONSE CURVES
O
CO

CO

U.

Model 5.15

True response

- 100

o
m
n
.DC 60.00
T I M E IN MIN.

Figure 5.16. Model and true concentration response with


temperature controller in maual.
Table V-4

Concentration PIP Controller for

Load Change

Underdamped Model

G(s) 0.001256
2.36 12s 2+2x 2 .361x0.772s+l

Optimum PIP (FCltfl

K = 2619.4
c
- 1.488
Ti
« 1.571
Td

Overdamped Model

rf ^ = 0.001247 e"°*lls
(3.66s+l)(0.72s+l)

Optimum PIP (FCIA)

K - 2488.4
c
Tt - 1.453

Td - 0.788
0.001247 e"0,lla
(5.16)
G(8) " (3.668+1)(0.728+1)

which fits the true response extremely well as can be seen In

Figure 5.17. The standard PID settings determined from the model

are given in Table V-4.

Multivariable Control Responses

It is very interesting to note that the PID settings for

the two models are quite different as tabulated in Table V-4,

but the response of the two controllers on concentration alone,

illustrated in Figures 5.18 and 5.19, look similar. Even though

the ITAE for both the methods are the same as shown in Table V-5

the FCIA tuning gives a slightly higher overshoot but a much

faster settling time compared to the FCIM tuning in this par­

ticular system, as can be seen in Figure 5.20 for a step change

in feed rate. Multivariable control to a step change in temper­

ature set point is shown in Figure 5.21.

For this system, the controllers for both temperature and

concentration perform very well to load or set point distur­

bances. Among the two tuning methods, it appears that FCIA

tuning has taken the interaction factor into consideration

whereas FCIM does not. Because of low interaction in this

particular system, both tuning methods seem satisfactory; if

the interaction were more, neither would probably have worked.


RESPONSE CURVES
o
CD
CO

O
CO True response and Model 5.16
co

U.

Q
C L-o
co
Zco"
M

00 10.00 20.00 30.00 50.00


TI M E IN MIN.

Figure 5.17. Model and true concentration response with temperature


controller in automatic.
o
CD
to

- 10.00 0.00 10.00


T I M E IN MIN.

03

O
o

-IU.00 0.00
T
I

Figure 5.18. Responses for step change in feed concentration; FCIM


tuning; temperature controller on manual.
Otn

- I '■'.00 0.00 10.00 20.00


T I M E IN M I N . -

*—I
D
fl_ .

111°>-

to
03
- ; a . oa 1C.QQ

Figure 5 19. Responses for step change in feed concentration;


FCIA tuning; temperature, control on manual.
Table V-5
Comparison of Two Tuning Methods

Setting for Temperature Controller in manual:

Load Change (flC-.*=+.5) JTAE *TAE


___________ v af ' Temp. Cone.

Temperature Control 28.7

Concentration Control — 0.325

Multivariable Control 20.2 0.216

Set Point Change (AT=-5°F)

Temperature Control 74.6

Multivariable Control 141.3 0.111

Setting for Temperature Controller in Automatic:

Load Change (AC ..=+.5)


_________& N af Temp. Cone.

Temperature Control 28.7

Concentration Control — 0.3286

Multivariable Control 20.7 0.163

Set Point Change (AT— S 0?)

Temperature Control 74.6

Multivariable Control 141.4 0.109


o
CD
CO

U_

Uncoi

FCIM controller

- 10.00 0.00 10.00 20.00 30.00


T I M E IN MIN.
O
O
=r
o>.

Single variable control

Multivariable control

Q
to
CD'

T I M E IN MIN.

Figure 5.20a. Multivariable control responses to step, change


in feed concentration FCIM tuning.
Uncontrolled

CLo FCIA controlled


co
z cn

- 10.00 0.00 30.00


T I M E IN M I N
O

zr
03,

Single variable control


fl_ -
11103"
Multivariable control

to

10.00
T i m e i n m i i '.
Figure 5.20b. Multivariable control responses to step change in
feed concentration FCIA tuning.
166

co

3 o
rj r*
^co'
Q
21
3 Uncontrolled
O
0.0
to
Z to‘
FCIM controller

O o
O m
co
- 10.00 0.00 10.00 20.00 30.00
T I M E IN MIN.
o
o
in
o>.

o
aft Single variable control
2 2
I ■ |W -

Multivariable control

o
o
ID'
. o . oa ea.aa
time in m i n .
Figure 5.21a. Multivariable control responses to step change
in temperature set point, FCIM tuning.
167

co
to

U.
3 0
O ^
*n .co “
Q

Uncontrolled

cn
FCIA. controller

D o
CJin
ro
10.00 30.00
T I M E IN MIN.
o
ca
in
o».

M
Single variable control

Multivariable control

ca
o
o
to
10.00 30.00
IME IN MIN.
Figure 5.21b. Multivariable control responses for temperature
set point change, FCIA tuning.
In this particular case, much of the discrepancy between

the tuning methods can be traced to the fact that the model

for the FCIM case was underdamped, whereas the model for the

FCIA case was overdamped and apparently gave a better fit.

That this would be true in general is doubtful.

In summary, the over-riding consideration in this situa­

tion is the degree of interaction. Only when the interaction

is low will either FCIM or FCIA work.


Literature Cited

1. Wills, D. M . , "Cascade Control Applications and Hardware",


Technical Bulletin Number TX119-1, Minneapolis, Honeywell
Philadelphia, 1960.

2. Coughanowr, D. R. and Koppel, L. G., Process Systems Analysis


and Control. McGraw-Hill, New York.

3. Murrlll, P. W., Automatic Control of Processes. International


Textbook Co., Scranton, Pa., 1967.

4. Ziegler, J. G., and Nichols, N. B., "Optimum Setting for


Automatic Controller", Transactions AS ME . (Nov. 1942), p.759.

5. Lopez, A. M . , Murrlll, P. W., and Smith, C. L., "Optimal


Tuning of Proportional Digital Controllers", Instruments and
Control Systems, (Oct. 1968), p. 97.

6. Lopez, A. M., et.al., "Tuning PI and PID Digital Controllers",


Instruments and Control Systems (Feb. 1969), p. 89.

7. Lopez, A. M., et.al, "An Advanced Tuning Method", British


Chemical Engineering. Vol. 14, No. 11 (Nov. 1969), pp. 1552-1555.

8. Rovira, A. A., Murrlll, P. W., and Smith, C. L., "Tuning


Controllers for Set Point Changes", Instruments and Control
Systems (Dec. 1969), p. 67.

9. Lopez, A. M., et.al., "Controller Tuning Relationships


Based on Integral Performance Criteria", Instrumentation
Technology. Vo.. 16, No. 11, Nov. 1967, p. 57.

10. Rovira, A. A., "Modified PI Algorithm for Digital Control",


Instrument and Control Systems. Vol. 43, No. 8 (Aug. 1970),
pp. 101-102.

11. Hooke, R. and Jeeves, T. A., "Direct Search Solution of


Numerical and Statistical Problems", J.A.C.M.. no. 2, Vol. 8,
1961.
12. Adkison, B. M. and Kohler, G. K., "Optimum Tuning of a
Generalized Three-mode Digital Control Algorithm", Paper
70-512. Proceedings of the 25th Annual ISA Conference.
Advances in Instrumentation. Vol. 25, 1970.

13. Rijnsdrop, J. E., Van Kampen, J. A, and Bollen, H. "Automatic


Feedback Control of Two Product Qualities of a Distillation
Column", Paper 32.b, Third IFAC, London, 1966.

14. Dahlin, E. G., "Interactive Control of Paper Machines",


Control Engineering. Jan. 1970.

15. Nisenfeld, A. E., and Schultz, H. M. , "Interaction Analysis in


Control System Design", Paper 70-562. Proceedings of the 25th
Annual ISA Conference. Advances in Instrumentation. Vol. 25,
Part 1, 1970.

16. Bristol, E. H.,"0n a New Measure of Interaction for Multi-


variable Process Control," IEEE Transactions on Automatic
Control. Jan. 1966, pp. 133-134.

17. Shinskey, F. G, "Material Balance Control on Multi-Product


Towers", Oil and Gas Journal. July 20, 1969.

18. Nisenfeld, A. E.and Stavinski, C., "Feed-Forward Control for


Azeotropic Distillation", Chemical Engineering. Sept. 23,1968.

19. Shinskey, F. G., Process Control Systems. McGraw-Hill Book Co.,


New York, 1967.
Chapter VI

CONCLUSIONS

The primary purpose of this dissertation is to study the

use of simple models in the design of control strategies for

complex processes. In this chapter, the conclusions drawn from

the results presented in the preViow chapters are sumnarized.

The major contribution of this work is the presentation

of the control parameters of a PI or PID algorithm in closed

form for set point changes. Ringing-free algorithms derived

from a first order model invariably reduced to a standard PI.

Similarly, ringing-free algorithms derived from a second order

model reduced to a standard PID. The reset time and derivative

time were identical for deadbeat, Kalman and Dahlin algorithms,

and are functions of the time constants of the models and the

sampling time adopted in a discrete control system. The

proportional gain is a function of model parameters and the sampling

time. The gain of the ringing-free Dahlin algorithm can be tuned

to become the gain of either deadbeat or Kalman algorithm.

While the results obtained in Chapter III and IV (specifically,

that the sampled-data algorithms always reduced to the standard

PI or PID) are specific to the system studied, Appendix A shows

that this result applies to a more general model, even one with
large dead time. While the reset time and and derivative time are

independent of the dead time, the controller gain decreases with

increasing dead time. This also assumes that all poles in the

left-half plane are removed, which may not always be desirable.

In effect, removal of all of these poles deletes the dead-time

compensation from the algorithm, which is not good. If dead time

compensation is not wantedrfor set point change it is not neces­

sary to design the control algorithm.

As the second order model approximates the system better

than the first order model, tuning using the second order model

is always superior than the first order model. The FID algorithm

performs better than the FI algorithm especially when a second

order model is employed.

Deadbeat, Kalman and Dahlin algorithms perform very well

for load disturbance. Ringing-free algorithms should be employed

for set point changes whereas, the original algorithms are more

suitable for load changes. Algorithms derived from second order

model always perform better than algorithms derived from first

order model.

As expected, cascade control performs better than the simple

feedback system especially for disturbance entering the inner loop.

Relative gain is unique in pairing of controlled and manipulated

variables. Tuning of the concentration loop with the temperature

controller in manual and in automatic give virtually the same result

for this particular system, which has low interaction.


APPENDIX A

DERIVATION OF STANDARD ALGORITHM

PARAMETERS FOR SET POINT CHANGE


First Order Plus Dead Time Model:

Continuous Transfer Function

K e ' es
G<*> - H > T

Pulse Transfer Function

HG (z ) “ “ --------------------^
(1-vz )

where

K = gain

0 = dead time

T = time constant

0 = (N + 6 )T

T - sampling time

f = 1 - exp(-hmT)

g = exp(-hmr)-exp(-hT)

h * 1/t

m “ 1 - 6

v ■ exp(-hT)
Deadbeat Algorithm

-N-l
D(z) - 1 • Z
HG (z) . -N-1
1 "Z

(l-vz . 1
mm 1 1
K(f+gz A) 1-z W 1

D(z)
K(f+gz-1)(l-z"1)(l+z_1+z“2+. ..-z"N)

Removing all the poles on the left side of the z-plane, we have

1 • tt-v s'1)
D<Z> " K(W-l) (ftg) (1_2-1)

This is equivalent to a PI with

KK
c (N+l) (ffg)

1 - v
T/T
i v

Second Order Pl»s Dead Model

Continuous Transfer Function

G(s) - K e~°S
(Tl8+1) (T2 s +1)
Pulse Transfer Function

HB-W ( c ^ - V e z - 2)
(1-xz )(1-yz )

where

t^»T2 ° time constant

x « exp (-aT)

y =• exp(-bT)

a = 1/Tj

b - 1/t2

xx = exp(-amT)

yy = exp(-bmT)

c = [1 + bxx-ayy J
a - b

d = fayy(l+x) - bxx(l-fy) - (a-b) (x+y)J / (a-b)

e •» £(a-b)xy+bxxy-ayyxj|/(a-b)

Deadbeat Algorithm
D(z) „ Is"* ±.£f < 2)
' ' -1 -2 -N-l
K(c+dz +ez )(l-z w )

Removing all the poles on the left side of the z-plane

D(Z) * --- 1---- &zl2S±Zl2li±2E3Cslf7


° (Z) K(N+1)(c+d+e) (1-z“l)

This is equivalent to a FID with

(x+y)-2xy
KK
c (N+l)(c+d+e)

t/t . lzfr*7)*EL
i (x+y)-2xy

t /t « .*y
Ad/A (x+y)-2xy
VITA

Kuo-Cheng Chiu was born on May 18, 1940 in Mogok, Burma

and received all his early education in Rangoon, Burma. His

undergraduate work was at the University of Rangoon, graduating

with a Bachelor of Science degree in Chemical Engineering in 1963.

From 1963 to 1966 he served on the faculty of the Department of

Chemical Engineering at Rangoon Institute of Technology. In

1968 the author received a Master of Science in Engineering

degree in Chemical Engineering from the University of Florida

at Gainesville, Florida. He held the position of Development

Engineer at the Louisiana Division of tlje Dow Chemical Company,

where he worked from 1968 to 1970.

The author is married to the former Chu-Ling Chang alias

Khin Mar Wai.


EXAMINATION AND THESIS REPORT

Candidate: Kuo-Cheng Chiu

Major Field: Chemical Engineering

Title of Thesis: Digital Control of Complex Systems Based on Simple Models

Approved:

M ajor Professor and Chairman

Dean of the Graduate School

EXAMINING COMMITTEE:

Date of Examination:

November 22, 1971