®

Control Station
Innovative Solutions from the Process Control Professionals

Practical Process Control

using LOOP-PRO Software

Douglas J. Cooper

Practical Process Control® by Douglas J. Cooper

Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Practical Process Control Using LOOP-PRO®

Copyright © 2006 by Control Station, Inc.

All rights reserved. No portion of this book may be reproduced in any form
or by any means except with explicit, prior, written permission of the author.

Doug Cooper, professor of chemical engineering at the University of Connecticut, has been teaching
and directing research in process control since 1985. Doug's research focuses on the development of
advanced control strategies that are reliable, and perhaps most important, easy for practitioners to
use. He strives to teach process control from a practical perspective. Thus, the focus of this book is on
proven control methods and practices that practitioners and new graduates can use on the job.

Author Publisher
Prof. Douglas J. Cooper Control Station, Inc.
Chemical Engineering Dept. One Technology Drive
University of Connecticut, Unit 3222 Tolland, CT 06084
Storrs, CT 06269-3222
Email: [email protected]
Email: [email protected] Web: www.controlstation.com

Practical Process Control® by Douglas J. Cooper

Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Table Of Contents
Page

Practical Control 8
1. Fundamental Principles of Process Control 8
1.1 Motivation for Automatic Process Control 8
1.2 Terminology of Control 10
1.3 Components of a Control Loop 11
1.4 The Focus of This Book 12
1.5 Exercises 13

2. Case Studies for Hands-On and Real-World Experience 15

2.1 Learning With a Training Simulator 15
2.2 Simulating Noise and Valve Dynamics 15
2.3 Gravity Drained Tanks 15
2.4 Heat Exchanger 16
2.5 Pumped Tank 17
2.6 Jacketed Reactor 18
2.7 Cascade Jacketed Reactor 19
2.8 Furnace Air/Fuel Ratio 19
2.9 Multi-Tank Process 21
2.10 Multivariable Distillation Column 22
2.11 Exercises 23

3. Modeling Process Dynamics - A Graphical Analysis of Step Test Data 24

3.1 Dynamic Process Modeling for Control Tuning 24
3.2 Generating Step Test Data for Dynamic Process Modeling 26
3.3 Process Gain, KP, From Step Test Data 26
3.4 Overall Time Constant, τP , From Step Test Data 28
3.5 Apparent Dead Time, θP , From Step Test Data 30
3.6 FOPDT Limitations - Nonlinear and Time Varying Behaviors 32
3.7 Exercises 35

4. Process Control Preliminaries 37

4.1 Redefining “Process” for Controller Design 37
4.2 On/Off Control – The Simplest Control Algorithm 38
4.3 Intermediate Value Control and the PID Algorithm 39

5. P-Only Control - The Simplest PID Controller 41

5.1 The P-Only Controller 41
5.2 The Design Level of Operation 42
5.3 Understanding Controller Bias, ubias 42
5.4 Controller Gain, KC , From Correlations 43
5.5 Reverse Acting, Direct Acting and Control Action 44
5.6 Set Point Tracking in Gravity Drained Tanks Using P-Only Control 44
5.7 Offset - The Big Disadvantage of P-Only Control 45
5.8 Disturbance Rejection in Heat Exchanger Using P-Only Control 46
5.9 Proportional Band 48

Practical Process Control® by Douglas J. Cooper

Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 5.10 Bumpless Transfer to Automatic 48
5.11 Exercises 48

6. Automated Controller Design Using Design Tools 51

6.1 Defining Good Process Test Data 51
6.2 Limitations of the Step Test 52
6.3 Pulse, Doublet and PRBS Test 52
6.4 Noise Band and Signal to Noise Ratio 54
6.5 Automated Controller Design Using Design Tools 55
6.6 Controller Design Using Closed Loop Data 59
6.7 Do Not Model Disturbance Driven Data! 60
6.8 FOPDT Fit of Underdamped and Inverse Behaviors 62

7. Advanced Modeling of Dynamic Process Behavior 64

7.1 Dynamic Models Have an Important Role Beyond Controller Tuning 64
7.2 Overdamped Process Model Forms 65
7.3 The Response Shape of First and Second Order Models 66
7.4 The Impact of KP, τP and θP on Model Behavior 67
7.5 The Impact of Lead Element τL on Model Behavior 70

8. Integral Action and PI Control 72

8.1 Form of the PI Controller 72
8.2 Function of the Proportional and Integral Terms 72
8.3 Advantages and Disadvantages to PI Control 74
8.4 Controller Bias From Bumpless Transfer 75
8.5 Controller Tuning From Correlations 75
8.6 Set Point Tracking in Gravity Drained Tanks Using PI Control 76
8.7 Disturbance Rejection in Heat Exchanger Using PI Control 79
8.8 Interaction of PI Tuning Parameters 80
8.9 Reset Time Versus Reset Rate 81
8.10 Continuous (Position) Versus Discrete (Velocity) Form 81
8.11 Reset Windup 82

9. Evaluating Controller Performance 83

9.1 Defining “Good” Controller Performance 83
9.2 Popular Performance Criteria 83

10. Derivative Action, Derivative Filtering and PID Control 86

10.1 Ideal and Non-interacting Forms of the PID Controller 86
10.2 Function of the Derivative Term 87
10.3 Derivative on Measurement is Used in Practice 87
10.4 Understanding Derivative Action 88
10.5 Advantages and Disadvantages of PID Control 89
10.6 Three Mode PID Tuning From Correlations 89
10.7 Converting From Interacting PID to Ideal PID 90
10.8 Exploring Set Point Tracking Using PID Control 91
10.9 Derivative Action Dampens Oscillations 92
10.10 Measurement Noise Hurts Derivative Action 93
10.11 PID With Derivative Filter Reduces the Impact of Noise 94
10.12 Four Mode PID Tuning From Correlations 95
10.13 Converting From Interacting PID with Filter to Ideal PID with Filter 96

Practical Process Control® by Douglas J. Cooper

Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 10.14 Exploring Set Point Tracking Using PID with Derivative Filter Control 96

Practical Theory 99
11. First Principles Modeling of Process Dynamics 99
11.1 Empirical and Theoretical Dynamic Models 99
11.2 Conserved Variables and Conservation Equations 99
11.3 Mass Balance on a Draining Tank 100
11.4 Mass Balance on Two Draining Tanks 103
11.5 Energy Balance on a Stirred Tank with Heater 104
11.6 Species (Component) Balance on a Stirred Tank with Reaction 106
11.7 Exercises 108

12. Linearization of Nonlinear Equations and Deviation Variables 110

12.1 The Linear Approximation 110
12.2 Linearization for Functions of One Variable 111
12.3 Linearization for Functions of Two Variables 112
12.4 Defining Deviation Variables 113
12.5 Deviation Variables Simplify the Equation Form 114
12.6 Exercises 116

13. Time Domain ODEs and System Behavior 117

13.1 Linear ODEs 117
13.2 Solving First Order ODEs 117
13.3 Deriving "τp = 63.2% of Process Step Response" Rule 119
13.4 Solving Second Order ODEs 121
13.5 The Second Order Underdamped Form 126
13.6 Roots of the Characteristic Equation Indicate System Behavior 127
13.7 Exercises 130

14. Laplace Transforms 133

14.1 Laplace Transform Basics 133
14.2 Laplace Transform Properties 135
14.3 Moving Time Domain ODEs into the Laplace Domain 138
14.4 Moving Laplace Domain ODEs into the Time Domain 141
14.5 Exercises 143

15. Transfer Functions 144

15.1 Process Transfer Functions 144
15.2 Controller Transfer Functions 146
15.3 Poles of a Transfer Function and Root Locus 148
15.4 Poles as Complex Conjugates 150
15.5 Poles of the Transfer Function Indicate System Behavior 151
15.6 Exercises 153

16. Block Diagrams 155

16.1 Combining Transfer Functions Using Block Diagrams 155
16.2 The Closed Loop Block Diagram 158
16.3 Closed Loop Block Diagram Analysis 159
16.4 Simplified Block Diagram 161

Practical Process Control® by Douglas J. Cooper

Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 16.5 The Padé Approximation 161
16.6 Closed Loop Analysis Using Root Locus 162
16.7 Exercises 166

17. Deriving PID Controller Tuning Correlations 168

17.1 The Direct Synthesis Design Equation 168
17.2 Deriving Controller Tuning Correlations Using Direct Synthesis 170
17.3 Internal Model Control (IMC) Structure 173
17.4 IMC Closed Loop Transfer Functions 174
17.5 Deriving Controller Tuning Correlations Using the IMC Method 175
17.6 Exercises 178

Combining Theory and Practice 180

18. Cascade Control 180
18.1 Architectures for Improved Disturbance Rejection 180
18.2 The Cascade Architecture 180
18.3 An Illustrative Example 181
18.4 Tuning a Cascade Implementation 184
18.5 Exploring the Jacketed Reactor Process 184
18.6 Single Loop Disturbance Rejection in the Jacketed Reactor 185
18.7 Cascade Disturbance Rejection in the Jacketed Reactor 187
18.8 Set Point Tracking Comparison of Single Loop and Cascade Control 192
18.9 Exercises 193

19. Feed Forward Control 194

19.1 Another Architecture for Improved Disturbance Rejection 194
19.2 The Feed Forward Architecture 194
19.3 An Illustrative Example 196
19.4 Feed Forward Control Design 198
19.5 Feed Forward Control Theory 198
19.6 Limits on the Form of the Feed Forward Model 200
19.7 Feed Forward Disturbance Rejection in the Jacketed Reactor 202
19.8 Static Feed Forward Control 206
19.9 Set Point Tracking Comparison of Single Loop and Feed Forward Control 207

20. Multivariable Controller Interaction and Loop Decoupling 209

20.1 Multivariable Process Control 209
20.2 Control Loop Interaction 210
20.3 Decouplers are Feed Forward Controllers 211
20.4 Distillation Study - Interacting Control Loops 214
20.5 Distillation Study - Decoupling the Loops 218

21. Modeling, Analysis and Control of Multivariable Processes 223

21.1 Generalizing 2x2 Multivariable Processes 223
21.2 Relative Gain as a Measure of Loop Interaction 224
21.3 Effect of KP on Control Loop Interaction 224
21.4 Effect of τP on Control Loop Interaction 228
21.5 Effect of θP on Control Loop Interaction 230
21.6 Decoupling Cross-Loop KP Effects 232

Practical Process Control® by Douglas J. Cooper

Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 21.7 Decoupling Cross-Loop τP Effects 235
21.8 Decoupling Cross-Loop θP Effects 236

22. Model Based Smith Predictor For Processes with Large Dead Time 238
22.1 A Large Dead Time Impacts Controller Performance 238
22.2 Predictive Models as Part of the Controller Architecture 239
22.3 The Smith Predictor Control Algorithm 239
22.4 Exploring the Smith Predictor Controller 241

23. DMC - Single Loop Dynamic Matrix Control 248

23.1 Model Predictive Control 248
23.2 Dynamic Matrix Control 248
23.3 The DMC Process Model 251
23.4 Tuning DMC 252
23.5 Example Implementation 253
23.6 Chapter Nomenclature 260
23.7 Tuning Strategy for Single Loop DMC 262

25. Non-Self Regulating (Integrating) Processes 263

25.1 Open Loop Behavior of Integrating Processes 263
25.2 The FOPDT Integrating Model 264
25.3 Modeling Data From Integrating Processes 265
25.4 Designing and Implementing Controllers for Integrating Processes 268

Appendix A: Derivation of IMC Tuning Correlations 270

A.1 Self Regulating Processes 270
A.1.a Ideal PID Control 270
A.1.b Interacting PID Control 272
A.1.c Ideal PID with Filter Control 274
A.1.d Interacting PID with Filter Control 276
A.2 Non-Self Regulating Processes 279
A.2.a Ideal PID Control 279
A.2.b Interacting PID Control 281
A.2.c Ideal PID with Filter Control 283
A.2.d Interacting PID with Filter Control 285

Appendix B: Table of Laplace Transforms 289

Appendix C: DMC Controller Tuning Guides 290

C.1 DMC Tuning Guide for Self Regulating (Stable) Processes 290
C.2 DMC Tuning Guide for Integrating (Non-Self Regulating) Processes 291

Appendix D: PID Controller Tuning Guides 292

D.1 PID Tuning Guide for Self Regulating (Stable) Processes 292
D.2 PID Tuning Guide for Integrating (Non-Self Regulating) Processes 293

Practical Process Control® by Douglas J. Cooper

Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Practical Control

1. Fundamental Principles of Process Control

1.1 Motivation for Automatic Process Control

Safety First
Automatic control systems enable a process to be operated in a safe and profitable manner. They
achieve this by continually measuring process operating parameters such as temperatures, pressures,
levels, flows and concentrations, and then making decisions to, for example, open valves, slow down
pumps and turn up heaters so that selected process measurements are maintained at desired values.
The overriding motivation for modern control systems is safety, which encompasses the
safety of people, the environment and equipment. The safety of plant personal and people in the
community is the highest priority in any plant operation. The design of a process and associated
control system must always make human safety the prime objective.
The tradeoff between safety of the environment and safety of equipment is considered on a
case by case basis. At the extremes, a nuclear power plant will be operated to permit as much as the
entire plant to be ruined rather than allowing significant radiation to be leaked to the environment. On
the other hand, a fossil fuel power plant may be operated to permit an occasional cloud of smoke to
be released to the environment rather than permitting damage to a multimillion dollar process unit.
Whatever the priorities for a particular plant, safety of both the environment and the equipment must
be specifically addressed when defining control objectives.

The Profit Motive

When people, the environment and plant equipment are properly protected, control objectives can
focus on the profit motive. Automatic control systems offer strong benefits in this regard. Plant-wide
control objectives motivated by profit include meeting final product specifications, minimizing waste
production, minimizing environmental impact, minimizing energy use and maximizing overall
production rate.
Process: Gravity Drained Tank Controller: Manual Mode

4.4
70 operating constraint
OperationOperation

4.2
60
More Profitable

4.0
50

poor control means

More Profitable

3.8
40 large variability, so
the process must be
operated in a less
3.6
30
process variable profitable region

60 80 100 120 140

Time (mins)

Figure 1.1 - Process variability from poor control means lost profits

8
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Product specifications set by the marketplace (your customers) are an essential priority if
deviating from these specifications lessens a product's market value. Example product specifications
range from maximum or minimum values for density, viscosity or component concentration, to
specifications on thickness or even color.
A common control challenge is to operate close to the minimum or maximum of a product
specification, such as a minimum thickness or a maximum impurities concentration. It takes more raw
material to make a product thicker than the minimum specification. Consequently, the closer an
operation can come to the minimum permitted thickness constraint without going under, the greater
the profit. It takes more processing effort to remove impurities, so the closer an operation can come to
the maximum permitted impurities constraint without going over, the greater the profit.
All of these plant-wide objectives ultimately translate into operating the individual process
units within the plant as close as possible to predetermined values of temperature, pressure, level,
flow, concentration or other of the host of possible measured process variables. As shown in Fig. 1.1,
a poorly controlled process can exhibit large variability in a process measurement over time. To
ensure a constraint limit is not exceeded, the baseline (set point) operation of the process must be set
far from the constraint, thus sacrificing profit.
Process: Gravity Drained Tank Controller: Manual Mode

4.2
70 operating constraint
OperationOperation

4.0
60

tight control permits

More Profitable

3.8
50
operation near the
process variable
constraint, which
means more profit
More Profitable

40
3.6

30
3.4

80 100 120 140 160

Time (mins)

Figure 1.2 - Well controlled process has less variability in process measurements

Figure 1.2 shows that a well controlled process will have much less variability in the
measured process variable. The result is improved profitability because the process can be operated
closer to the operating constraint.
Automatic Process Control
Because implementation of plant-wide objectives translates into controlling a host of individual
process parameters within the plant, the remainder for this text focuses on proven methods for the
automatic control of individual process variables. Examples used to illustrate concepts are drawn
from the LOOP-PRO® software package.
The Case Studies module presents industrially relevant process control challenges including
level control in a tank, temperature control of a heat exchanger, purity control of a distillation column
and concentration control of a jacketed reactor. These real-world challenges will provide hands-on
experience as you explore and learn the concepts of process dynamics and automatic process control
presented in the remainder of this book.

9
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
1.2 Terminology of Control
The first step in learning automatic process control is to learn the jargon. We introduce some basic
jargon here by discussing a control system for heating a home as illustrated in Fig. 1.3. This is a
rather simple automatic control example because a home furnace can only be either on or off.
As we will explore later, the challenges of control system design increase greatly when
process variable adjustments can assume a complete range of values between full on and full off. In
any event, a home heating system is easily understood and thus provides a convenient platform for
introducing the relevant terminology.
The control objective for the process illustrated in Fig. 1.3 is to keep the measured process
variable (house temperature) at the set point value (the desired temperature set on the thermostat by
the home owner) in spite of unmeasured disturbances (heat loss from doors and windows opening;
heat being transmitted through the walls of the house).
To achieve this control objective, the measured process variable is compared to the
thermostat set point. The difference between the two is the controller error, which is used in a
computation by the controller to compute a controller output adjustment (an electrical or pneumatic
signal).

thermostat
controller
set point
TC TT
temperature heat loss
sensor/transmitter (disturbance)
control
signal

fuel flow
furnace
valve

Figure 1.3 - Home heating control system

The change in the controller output signal causes a response in the final control element (fuel
flow valve), which subsequently causes a change in the manipulated process variable (flow of fuel to
the furnace). If the manipulated process variable is moved in the right direction and by the right
amount, the measured process variable will be maintained at set point, thus satisfying the control
objective. This example, like all in process control, involves a measurement, computation and action:

Measurement Computation Action

house temperature, THouse is it colder than set point ( TSetpoint − THouse > 0 )? open fuel valve

is it hotter than set point ( TSetpoint − THouse < 0 )? close fuel valve

Note that computing the necessary controller action is based on controller error, or the difference
between the set point and the measured process variable.

10
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
1.3 Components of a Control Loop
The home heating control system of Fig. 1.3 can be organized in the form of a traditional feedback
control loop block diagram as shown in Fig. 1.4. Such block diagrams provide a general organization
applicable to most all feedback control systems and permit the development of more advanced
analysis and design methods.

controller manipulated house

controller output fuel flow to temperature
error signal furnace
Set Point Home Heating
+ Thermostat Fuel Valve
- Process

Heat Loss
house temperature Disturbance
measurement
signal

Temperature
Sensor/Transmitter

Figure 1.4 - Home heating control loop block diagram

Following the diagram of Fig. 1.4, a sensor measures the measured process variable and
transmits, or feeds back, the signal to the controller. This measurement feedback signal is subtracted
from the set point to obtain the controller error. The error is used by the controller to compute a
controller output signal. The signal causes a change in the mechanical final control element, which in
turn causes a change in the manipulated process variable. An appropriate change in the manipulated
variable works to keep the measured process variable at set point regardless of unplanned changes in
the disturbance variables.
The home heating control system of Fig. 1.4 can be further generalized into a block diagram
pertinent to all control loops as shown in Fig. 1.5. Both these figures depict a closed loop system
based on negative feedback, because the controller works to automatically counteract or oppose any
drift in the measured process variable.
Suppose the measurement signal was disconnected, or opened, in the control loop so that the
signal no longer feeds back to the controller. With the controller no longer in automatic, a person
must manually adjust the controller output signal sent to the final control element if the measured
process variable is to be affected.
It is good practice to adjust controller tuning parameters while in this manual, or open loop
mode. Switching from automatic to manual, or from closed to open loop, is also a common
emergency procedure when the controller is perceived to be causing problems with process operation,
ranging from an annoying cycling of the measured process variable to a dangerous trend toward
unstable behavior.

11
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 manipulated measured
controller controller process process
error output variable variable
Final
Set Point
+ Controller Control Process
- Element

Disturbance

feedback
signal

Measurement
Sensor/Transmitter

Figure 1.5 - General control loop block diagram

1.4 The Focus of This Book

Although an automatic control loop is comprised of a measurement, computation and action, details
about the commercial devices available for measuring process variables and for implementing final
control element actions are beyond the scope of this text. For the kinds of process control applications
discussed in this book, example categories of such equipment include:

Sensors to Measure: temperature, pressure, pressure drop, level, flow, density, concentration
Final Control Elements: solenoid, valve, variable speed pump or compressor, heater or cooler

The best place to learn about the current technology for such devices is from commercial
vendors, who are always happy to educate you on the items they sell. Contact several vendors and
learn how their particular merchandise works. Ask about the physical principles employed, the kinds
of applications the device is designed for, the accuracy and range of operation, the options available,
and of course, the cost of purchase. Keep talking with different vendors, study vendor literature, visit
websites and participate in sales demonstrations until you feel educated on the subject and have
gained confidence in a purchase decision. Don't forget that installation and maintenance are important
variables in the final cost equation.
The third piece of instrumentation in the loop is the controller itself. The automatic
controllers explored in some detail in this book include:

Automatic Controllers: on/off, PID, cascade, feed forward, model-based Smith predictor,
multivariable, sampled data, parameter scheduled adaptive control

Although details about the many commercial products are beyond the scope of this text, fortunately,
the basic computational methods employed by most all vendors for the controllers listed above are
remarkably similar. Thus, the focus of this book is to help you:

12
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 • learn how to collect and analyze process data to determine the essential dynamic behavior of a
process,

• learn what "good" or "best" control performance means for a particular process,

• understand the computational methods behind each of the control algorithms listed above and
learn when and how to use each one to achieve this best performance,

• learn how the different adjustable or tuning parameters required for control algorithm
implementation impact closed loop performance and how to determine values for these
parameters,

• become aware of the limitations and pitfalls of each control algorithm and learn how to turn this
knowledge to your advantage.

1.5 Exercises

Q-1.1 New cars often come with a feature called cruise control. To activate cruise control, the driver
presses a button while traveling at a desired velocity and removes his or her foot from the gas
pedal. The control system then automatically maintains whatever speed the car was traveling
when the button was pressed in spite of disturbances. For example, when the car starts going
up (or down) a hill, the controller automatically increases (or decreases) fuel flow rate to the
engine by a proper amount to maintain the set point velocity.

a) For cruise control in an automobile, identify the:

- control objective
- measured process variable
- manipulated variable
- set point
- two different disturbances
- measurement sensor
- final control element

b) Draw and properly label a closed loop block diagram for the cruise control process.

Q-1.2 The figure below shows a tank into which a liquid freely flows. The flow of liquid out of the
tank is regulated by a valve in the drain line. The control objective is to maintain liquid level
in the tank at a fixed or set point value. Liquid level is inferred by measuring the pressure
differential across the liquid from the bottom to the top of the tank. The level
sensor/controller, represented in the diagram as the LC in the circle, continually computes
how much to open or close the valve in the drain line to increase or decrease the flow out as
needed to maintain level at set point.

13
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Flow In

LSetPoint
LC

Flow Out

Draw and label a closed loop block diagram for this level control process.

14
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 2. Case Studies for Hands-On and Real-World Experience

2.1 Learning With a Training Simulator

Hands-on challenges that demonstrate and reinforce important concepts are crucial to learning the
often abstract and mathematical subject of process dynamics and control. The Case Studies module, a
training simulator that challenges you with real-world scenarios, provides this reinforcement. Use
Case Studies to explore dozens of challenges, brought to life in color-graphic animation, to safely and
inexpensively gain hands-on experience.
The Case Studies module contains several simulations for study. You can manipulate process
variables to obtain step, pulse, ramp, sinusoidal or PRBS (pseudo-random binary sequence) test data.
Process data can be recorded as printer plots and as disk files for process modeling and controller
design studies. The Design Tools module in LOOP-PRO is well suited for this modeling and design
task. After designing a controller, return to the Case Studies simulation to immediately evaluate and
improve upon the design for both set point tracking and disturbance rejection.
The processes and controllers available in Case Studies enable exploration and study of
increasingly challenging concepts in an orderly fashion. Early concepts to explore include basic
process dynamic behaviors such as process gain, time constant and dead time. Intermediate concepts
include the tuning and performance capabilities of P-Only through PID controllers and all
combinations in between. Advanced concepts include cascade, decoupling, feed forward, dead time
compensating and discrete sampled data control.

2.2 Simulating Noise and Valve Dynamics

For all processes, changes in the controller output signal pass through a first order dynamic response
element before impacting the manipulated or disturbance variables. This simulates the lag associated
with the mechanical movement of a process valve.
Normally distributed random error is added to the measured process variable for all processes
to simulate measurement noise. The value entered can be user-adjusted, is in the units of the measured
process variable, and represents ±3 standard deviations of the normal distribution of the random error.

2.3 Gravity Drained Tanks

The gravity drained tanks process, shown in Fig. 2.1, is two non-interacting tanks stacked one above the
other. Liquid drains freely through a hole in the bottom of each tank. As shown, the measured process
variable is liquid level in the lower tank. To maintain level, the controller manipulates the flow rate of
liquid entering the top tank. The disturbance variable is a secondary flow out of the lower tank from a
positive displacement pump. Thus, the disturbance flow is independent of liquid level except when the
tank is empty.
This is a good process to start your studies because its dynamic behavior is reasonably intuitive.
Increase the liquid flow rate into the top tank and the liquid level rises. Decrease the flow rate and the
level falls. One challenge this process presents is that its dynamic behavior is modestly nonlinear. That is,
the dynamic behavior changes as operating level changes. This is because gravity driven flows are
proportional to the square root of the hydrostatic head, or height of liquid in a tank.

15
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 manipulated variable
controller output
.

measured
process variable level sensor
& controller

disturbance
variable

Figure 2.1 - Gravity drained tanks case study

2.4 Heat Exchanger

The heat exchanger, shown in Fig. 2.2, is a counter-current, shell and tube, lube oil cooler. The measured
process variable is lube oil temperature exiting the exchanger on the tube side. To maintain temperature,
the controller manipulates the flow rate of cooling liquid on the shell side. The nonlinear character of the
heat exchanger, or change in dynamic process behavior evident as operating level changes, is more
pronounced compared to that of the gravity drained tanks. The significance of nonlinear dynamic
behavior on controller design will become apparent as we work through this book.

disturbance variable
cooling flow exit
.

manipulated variable

controller output

temperature sensor
& controller
measured
process variable

Figure 2.2 - Heat exchanger case study

16
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 This process has a negative steady state process gain. Thus, as the controller output (and thus
flow rate of cooling liquid) increases, the exit temperature (measured process variable) decreases. Another
interesting characteristic is that disturbances, which result from changes in the flow rate of warm oil that
mixes with the hot oil entering the exchanger, cause an inverse or nonminimum phase open loop response
in the measured exit temperature.
To understand this inverse response, consider that an increase in the warm oil disturbance flow
increases the total flow rate of liquid passing through the exchanger. Liquid already in the exchanger
when the disturbance first occurs is forced through faster than normal, reducing the time it is exposed to
cooling. Hence, the exit temperature initially begins to rise. Now, the warm oil disturbance stream is
cooler than the hot oil, so an increase in the disturbance flow lowers the mixed stream temperature
entering the exchanger.
Once the new cooler mixed liquid works its way through the exchanger and begins to exit, it will
steady out at a colder exit temperature than prior to the disturbance. Thus, an increase in the disturbance
flow rate causes the measured exit temperature to first rise (from faster flow) and then decrease (from
cooler mixed liquid entering) to a new lower steady state temperature.

2.5 Pumped Tank

The pumped tank process, shown in Fig. 2.3, is a pickle brine surge tank. The measured process
variable is liquid brine level. To maintain level, the controller manipulates the brine flow rate out of
the bottom of the tank by adjusting a throttling valve at the discharge of a constant pressure pump.
This approximates the behavior of a centrifugal pump operating at relatively low throughput. The
disturbance variable is the flow rate of a secondary feed to the tank.

disturbance variable

level sensor .

& controller

measured controller output

process variable

manipulated variable

Figure 2.3 - Pumped tank case study

Unlike the gravity drained tanks, the pumped tank is not a self regulating process (it does not
reach a natural steady state level of operation). Consider that the discharge flow rate is regulated
mechanically and changes only when the controller output changes. The height of liquid in the tank
does not impact the discharge flow rate. As a result, when the total flow rate into the tank is greater
than the discharge flow rate, tank level will continue to rise until the tank is full, and when the total
flow rate into the tank is less than the discharge flow rate, the tank level will fall until empty.

17
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 This non-self-regulating dynamic behavior is associated with integrating processes. The
pumped tank appears almost trivial in its simplicity. Its integrating nature presents a remarkably
difficult control challenge.

2.6 Jacketed Reactor

The jacketed reactor, shown in Fig. 2.4 for the single loop case, is a continuously stirred tank reactor
(CSTR) in which the first order irreversible exothermic reaction A→B occurs. Residence time is
constant in this perfectly mixed reactor, so the steady state conversion of reactant A from the reactor
can be directly inferred from the temperature of the reactor product stream. To control reactor
temperature, the vessel is enclosed with a jacket through which a coolant passes.
The measured process variable is the product exit stream temperature. To maintain
temperature, the controller manipulates the coolant flow rate through the jacket. The disturbance
variable is the inlet temperature of coolant flowing through the jacket.

manipulated
variable

disturbance
.

variable controller output

temperature sensor
& controller
measured
process variable

Figure 2.4 – Single loop jacketed reactor case study

This process has an upper and lower operating steady state. The process initializes at the
upper steady state, indicated by high values for percent conversion in the exit stream. You can move
the process to its lower steady state by dropping the cooling jacket inlet temperature to low values
(for example, with the controller output at the startup value of 42%, change the jacket inlet
temperature to 30°C and the reactor will fall to the lower operating steady state).
The process is modeled following developments similar to those presented in many popular
chemical engineering texts. Assuming an irreversible first order reaction (A→B); perfect mixing in
the reactor and jacket; constant volumes and physical properties; and negligible heat loss, the model
is expressed:
dC A F
Mass balance on Reactor A: = (C A0 − C A ) − kC A
dt V

dT F ∆H R UA
Energy balance on reactor contents: = (T0 − T ) − kC A − (T − TJ )
dt V ρCP Vρ C P

18
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 dT J UA F
Energy balance on reactor jacket: = (T − TJ ) + J (TJ 0 − TJ )
dt V J ρ J C PJ VJ

Reaction rate coefficient: k = k 0 e − E / RT

2.7 Cascade Jacketed Reactor

The open loop process behavior of the cascade jacketed reactor, shown in Fig. 2.5, is identical to the
single loop case. The only difference between the two is the controller architecture. A control cascade
consists of two process measurements, two controllers, but only one final control element - the same
final control element as in the single loop case study. The benefit of a cascade architecture, as
explored later in Chapter 18, is improved disturbance rejection.
For the cascade case, the secondary measured process variable is the coolant temperature out
of the jacket. The secondary or inner loop controller manipulates the cooling jacket flow rate. As in
the single loop case, the primary or outer measured process variable is still the product stream
temperature. As with all cascade architectures, the output of the primary controller is the set point of
the secondary controller.

secondary
manipulated variable

secondary
disturbance process variable
variable
.
secondary sensor
& controller .

primary controller output

(manipulate secondary set point)

primary sensor
& controller

primary
process variable

Figure 2.5 - Cascade jacketed reactor case study

2.8 Furnace Air/Fuel Ratio

The furnace air/fuel ratio case study is a process in which a furnace burns natural gas to heat a process
liquid flowing through tubes at the top of the fire box. As shown in Fig. 2.6, the measured process
variable is the temperature of the process liquid as it exits the furnace. To maintain temperature,
controllers adjust the feed rate of combustion air and fuel to the fire box using a ratio control strategy.
The flow rate of the process liquid acts as a load disturbance to the process.
A ratio controller measures the flow rate of an independent or "wild" stream. It then adjusts
the flow rate of a second dependent stream to maintain a user specified ratio between the two.
Blending operations are a typical application for ratio control, and here we are blending air and fuel in
a desired fashion.

19
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 For the furnace, the independent stream is the combustion air flow rate and the dependent
stream is the fuel flow rate. Note that while air flow rate is considered the independent stream for
ratio control, its flow rate is specified by the temperature controller on the process liquid exiting the
furnace.
Air/fuel ratio control provides important environmental, economic and safety benefits.
Running in a fuel-rich environment (too little air to complete the combustion reaction) permits the
evolution of carbon monoxide and unburned hydrocarbons in the stack gases. These components are
not only pollutants, but unburned hydrocarbon represents wasted energy. A fuel-lean environment
(air in excess of that needed to complete combustion) results in economic loss because it takes extra
fuel to heat extra air that is then just lost up the stack. Also, excess air can lead to the increased
production of nitrogen oxides that promote the formation of smog. In general, it is desirable to
maintain the ratio of air to fuel at a level that provides a small excess of oxygen, say 2-5%, compared
to the stoichiometric requirement.
The ratio controller also provides an important safety benefit. As shown in Fig. 2.6, it is the
air flow rate that is adjusted by the temperature controller on the liquid exiting the furnace. When the
liquid temperature falls below set point, the temperature controller raises the set point flow rate of the
combustion air. As the air flow ramps up, the air flow transmitter detects the change and sends the
new flow rate to the ratio controller. The ratio controller responds by increasing the set point to the
fuel flow controller in order to maintain the specified ratio.

measured
disturbance
process variable
variable

temperature sensor
& controller

.
dependent
manipulated variable .

ratio set point high select

controller output
independent
manipulated variable

Figure 2.6 -Furnace Air/Fuel Ratio case study

To understand why this is the safer alternative, consider what would happen if the plant were
to lose combustion air (e.g. the air compressor dies). The air flow transmitter will immediately detect
the falling flow rate and send the information to the ratio controller. The ratio controller will respond
by cutting the fuel flow. As a result, the hazardous situation of fuel without sufficient combustion air
in the furnace fire box will be averted.

20
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Now suppose we make fuel the independent stream and control air flow rate in ratio to the
fuel. If we were to lose air as the liquid temperature controller calls for more fuel, the ratio controller
will not be aware that a problem exists because it will only be receiving information from the fuel
flow transmitter. The ratio controller will continue to send fuel to the furnace in an attempt to raise
temperature, and an explosive environment will develop.
An additional piece of the safety control strategy lies with the high select component (the HS
with the circle around it on the process graphic that appears when the temperature controller is in
automatic). As shown in Fig. 2.6, the high select receives two candidate set points. One is the air flow
rate set point requested by the temperature controller. The other is the minimum air flow rate
permitted given the current flow rate of fuel. The minimum air/fuel ratio permitted is 10/1, and to go
below this value means there will be more fuel in the fire box than air needed to combust it. The high
select sends the highest of the two candidate set points to the air flow controller, thus ensuring that a
fuel rich environment is not created. This intervention in the control loop creates interesting control
challenges.

2.9 Multi-Tank Process

The multi-tank process is a multivariable version of the single loop gravity drained tanks process. As
shown in the process graphic of Fig. 2.7, there are two sets of freely draining tanks positioned side by
side. The two measured process variables are the liquid levels in the lower tanks. To maintain level,
the two level controllers manipulate the flow rate of liquid entering their respective top tanks.

controller
output 2
manipulated variable 1 manipulated variable 2

controller upper tank 2

output 1

upper tank 1 lower tank 2

lower tank 1 measured

.
process variable 2
.

level sensor
level sensor & controller 2
& controller 1

disturbance
variable 1
disturbance
variable 2

Figure 2.7 – Multi-Tank case study

21
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Similar to the single loop case, the gravity driven flows are proportional to the square root of
the height of liquid in the tank (hydrostatic head), so this process displays moderately nonlinear
behavior. Each lower tank has a secondary flow out of the lower tank that is independent of liquid level
and acts as an operating disturbance.
An important feature of this process is that each of the upper tanks drains into both lower
tanks. This creates a multivariable interaction because actions by one controller affect both measured
process variables. Note that the characteristics of each drain stream are all different so there is no
symmetry between feed flow rate and steady state tank level.
To gain a better understanding of the multivariable loop interactions, suppose the measured
level in lower tank 1 is below set point. The tank 1 level controller responds by increasing F1, the
flow rate entering upper tank 1. While this action will cause the level to rise in lower tank 1, the side
drain out of upper tank 1 will increase, causing the level in lower tank 2 to rise also.
As the level in lower tank 2 is forced from set point, the tank 2 level controller compensates
by decreasing the flow rate of F2, the flow rate entering upper tank 2. This decreases the level in tank
2, but the decrease in the side drain rate out of upper tank 2 causes the level in lower tank 1 to
decrease. The tank 1 level controller "fights back" with more corrective actions, causing challenging
multivariable loop interactions.
Decouplers are simple models of the process that can be designed into a controller architecture to
minimize such multivariable interaction. As we will learn in Section 20.3, decouplers are feed forward
elements that treat the actions of another controller as a measured disturbance.

2.10 Multivariable Distillation Column

The distillation column case study, shown in Fig. 2.8, consists of a binary distillation column that
separates benzene and toluene. The objective is to send a high percentage of benzene (and thus low
percentage of toluene) out the top distillate stream and a low percentage of benzene (and thus high
percentage of toluene) out the bottoms stream. The column dynamic model employs tray-by-tray mass
and energy calculations similar to that proposed by McCune and Gallier [ISA Transactions, 12, 193,
(1973)].
To separate benzene from toluene, the top controller manipulates the reflux rate to control the
distillate composition. The bottom controller adjusts the rate of steam to the reboiler to control the
bottoms composition. Any change in feed rate to the column acts as a disturbance to the process.
To understand the loop interaction in this multivariable process, suppose the composition (or
purity) of benzene in the top distillate stream is below set point. The top controller will respond by
increasing the flow rate of cold reflux into the column. This increased reflux flow will indeed increase the
benzene purity of the distillate stream. However, the additional cold liquid will work its way down the
column, begin to cool the bottom, and as a result permit more benzene to flow out the bottoms stream.
As the bottoms composition moves off set point and produces a controller error, the bottom
controller will compensate by increasing the flow of steam into the reboiler to heat up the bottom of the
column. While this works to restore the desired purity to the bottoms stream, unfortunately, it also results
in an increase of hot vapors traveling up the column that eventually will cause the top of the column to
begin to heat up.

22
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 As the top of the column heats up, the purity of benzene in the distillate stream again becomes too
low. In response, the top controller compensates by further increasing the flow of cold reflux into the top
of the column. The controller “fight," or multivariable interaction, begins. Like the multi tank process,
decouplers can minimize such multivariable loop interaction.

manipulated variable

measured
process variable

disturbance composition sensor

& controller
.

variable .

manipulated variable

composition sensor
& controller

measured
process variable

Figure 2.8 - Distillation column case study

2.11 Exercises

Q-2.1 Measurement sensors are not discussed in this book but are a critical part of any control loop.
Among other features, consider that a sensor should:
- generate a continuous electrical signal that can be transmitted to the controller,
- be reasonably priced and easy to install and maintain,
- be capable of withstanding the rigors of the environment in which they are placed,
- respond quickly to changes in the measured process variable,
- be sufficiently accurate and properly calibrated to the application.

Using only information you obtain from the world wide web, research and then specify such
details for a:
a) liquid level sensor suitable for both the gravity drained tanks and the pumped tank process

b) temperature sensor for the heat exchanger process

c) temperature sensor for the single loop jacketed reactor process

23
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 3. Modeling Process Dynamics - A Graphical Analysis of Step Test Data

3.1 Dynamic Process Modeling for Control Tuning

Consider the design of a cruise control system for a car versus that for an eighteen wheel truck. For
the two vehicles, think about:
- how quickly each vehicle can accelerate or decelerate using the gas pedal, and
- the effect of disturbances on each vehicle such as wind, hills and other vehicles passing by.
The actions that a properly designed controller must take to control a car versus a truck, that is, how
the controller output signal should manipulate gas flow to maintain a constant set point velocity in
spite of disturbances such as wind and hills, will differ because the dynamic behavior of each vehicle
"process" is different.
The dynamic behavior of a process describes how the measured process variable changes
with time when forced by the controller output and any disturbance variables. The manner in which a
measured process variable responds over time is fundamental to the design and tuning of an automatic
controller.
The best way to learn about the dynamic behavior of a process is to perform experiments.
Whether the process to be controlled is found in LOOP-PRO, a lab or a production facility, this
experimental procedure entails the following steps:

1) The controller output is stepped, pulsed or otherwise perturbed, usually in manual mode, and
as near as practical to the design level of operation,
2) The controller output and measured process variable data are recorded as the process
responds,
3) A first order plus dead time (FOPDT) dynamic model is fit to this process data,
4) The resulting FOPDT dynamic model parameters are used in a correlation to obtain initial
estimates of the controller tuning parameters,
5) The tuning parameters are entered into the controller, the controller is put in automatic and
controller performance is evaluated in tracking set points and/or rejecting disturbances,
6) Final tuning is performed on-line and by trial and error until desired controller performance is
obtained.

Generating experimental data for step 1 above is best done in manual mode (open loop). The goal is
to move the controller output far enough and fast enough so that the dynamic character of the process
is revealed as the measured process variable responds. Because dynamic process behavior usually
differs as operating level changes (real processes display this nonlinear behavior), these experiments
should be performed at the design level of operation (where the set point will be set during normal
operation).
The response of the measured process variable during a test must clearly be the result of the
change in the controller output. To obtain a reliable model from test data, the measured process
variable should be forced to move at least 10 times the size of the noise band, or 10 times farther than
the random changes resulting from noise in the measurement signal that are evident prior to the start
of the test. If process disturbances occurring during a dynamic test influence the response of the
measured process variable, the test data is suspect and the experiment should be repeated.

24
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 It is becoming increasingly common for dynamic studies to be performed with the controller
in automatic (closed loop). For closed loop studies, the dynamic data is generated by stepping,
pulsing or otherwise perturbing the set point. Closed loop testing can be problematic because, in
theory, the information contained in the data will reflect the character of the controller as well as that
of the process. In practice, however, the controller character rarely produces significant corruption in
the final FOPDT model parameter values. In the unhappy event that you are not permitted to perform
either open or closed loop dynamic tests on a process, a final alternative is to search data logs for
useful data that have been recorded in the recent past.
Open Loop Step Test
Process: Custom Process Controller: Manual Mode

60
Step Test
Process Variable

55

50

60
Controller Output

55

50

0 5 10 15 20
Time (mins)
Time

Figure 3.1 - Controller output step with measured process variable response

The dynamic process data set is then analyzed in step 3 to yield the three parameters of a
linear first order plus dead time (FOPDT) dynamic process model:

dy (t )
τP + y (t ) = KP u(t − θP) (3.1)
dt

where y(t) is the measured process variable and u(t) is the controller output signal. When Eq. 3.1 is fit
to the test data, the all-important parameters that describe the dynamic behavior of the process result:

- Steady State Process Gain, KP

- Overall Process Time Constant, τP
- Apparent Dead Time, θP

The importance of these three model parameters is revealed in step 4, where they are used in
correlations to compute initial tuning values for a variety of controllers. We will also learn later that
this model is important because, for example:

25
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 - the sign of KP indicates the sense of the controller (+KP → reverse acting; −KP → direct acting)
- the size of τP indicates the maximum desirable loop sample time (be sure sample time T ≤ 0.1τP)
- the ratio θP /τP indicates whether a Smith predictor would show benefit (useful if θP > τP)
- the dynamic model itself can be employed within the architecture of feed forward, Smith
predictor, decoupling and other model-based controller strategies.
Thus, the collection and modeling of dynamic process data are indeed critical steps in controller
design and tuning.

3.2 Generating Step Test Data for Dynamic Process Modeling

The popular experiments to generate dynamic process data include the step test, pulse test, doublet
test and pseudo-random binary sequence (PRBS) test. Data generated by a pulse, doublet or PRBS
require a computer tool for fitting the dynamic model and are discussed later in Section 6.3.
Here we explore the fitting of a FOPDT dynamic model to step test data using a hand-
analysis of the process response plot. Our goal is to obtain values for process gain, KP, overall time
constant, τP, and apparent dead time, θP, that reasonably describe the dynamic behavior of the process
and thus can be used for controller design and tuning.
Figure 3.1 shows a typical step test response plot. As prescribed in the experimental test
procedure described earlier, the controller is in manual mode and the controller output and measured
process variable are initially at steady state. The test entails stepping the controller output to a new
constant value and collecting data as the measured process variable responds completely to its final
steady state.

3.3 Process Gain, KP, From Step Test Data

The steady state process gain describes how much the measured process variable, y(t), changes in
response to changes in the controller output, u(t). A step test starts and ends at steady state, so KP can
be determined directly from the plot axes. Specifically, KP is computed as the steady state change in
the measured process variable divided by the change in the controller output signal that forced that
change, or:

Steady State Change in the Measured Process Variable, ∆y (t )

KP = (3.2)
Steady State Change in the Controller Output, ∆u (t )

where ∆u(t) and ∆y(t) represent the total change from initial to final steady state.

Example: Gravity Drained Tanks

Figure 3.2 shows open-loop step response data from the gravity drained tanks. The process is
initially at steady state with the controller output at 50%, while the pumped flow disturbance (not
shown on plot) remains constant at 2.0 L/min. The controller output is stepped from 50% up to a
new steady state value of 60%. This step increase causes the valve to open, increasing the liquid
flow rate into the top tank.
In response, the measured variable for this process, the liquid level in the lower tank, rises
from its initial steady state value of 1.93 m up to a new steady state level of 2.88 m. Note that the
response plot of Fig. 3.2 displays data collected on either side of an average measured liquid level
of about 2.4 m.

26
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Gravity Drained Tanks - Open Loop Step Test
Process: Gravity Drained Tank Controller: Manual Mode
3.0

Process Variable
2.8
2.6
2.4 ∆y = (2.88 - 1.93) m
2.2
2.0
Controller Output 1.8

60

55 ∆u = (60 - 50) %
50

45
8 9 10 11 12 13 14 15 16 17 18
Time (mins)

Figure 3.2 - KP computed from gravity drained tanks step test plot

Using Eq. 3.2, the steady state process gain for this process is:

∆y 2 . 88 − 1 . 93 m m
KP = = = 0 . 095
∆u 60 − 50 % %

Note that KP has a size (0.095), a sign (positive, or +0.095 in this case), and units (m/%).

Õ Õ Õ

Example: Heat Exchanger

Figure 3.3 shows open loop step response data from the heat exchanger. The controller output is
initially constant at 25% while the warm oil disturbance flow rate (not shown on plot) remains
constant at 10 L/min. The controller output is stepped from 25% up to a new steady state value of
35%. This step increase causes the valve to open, increasing the flow rate of cooling liquid into
the shell side of the exchanger.
The additional cooling liquid causes the measured variable for this process, liquid exit
temperature on the tube side, to decrease from its initial steady state value of 151.2°C down to a
new steady state value of 142.6°C. Note that the response plot of Fig. 3.3 displays data collected
on either side of an average measured exit temperature of about 147°C.
Using Eq. 3.2, the steady state process gain is:

∆ y 142.6 − 151.2 ° C °C
KP = = = − 0 . 86
∆u 35 − 25 % %

Again, KP has a size (−0.86), a sign (negative, or −0.86 in this case), and units (°C/%).

27
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Heat Exchanger - Open Loop Step Test
Process: Heat Exchanger Controller: Manual Mode
152

Process Variable
150
148 ∆y = (142.6 - 151.2) oC
146
144
142
Controller Output
35

30 ∆u = (35 - 25) %

25

5 6 7 8 9 10 11 12 13
Time (mins)

Figure 3.3 - KP computed from heat exchanger step test plot

Õ Õ Õ

3.4 Overall Time Constant, τP , From Step Test Data

The overall process time constant describes how fast a measured process variable responds when
forced by a change in the controller output. The clock that measures "how fast" does not start until the
measured process variable shows a clear and visible response to the controller output step. We will
soon see in Section 3.5 that the delay that occurs from when the controller output is stepped until the
measured process variable actually starts to respond is the process dead time. To estimate τP from step
test data:

1) Locate on the plot where the measured process variable first shows a clear initial response to
the step change in controller output. The time where this response starts is tYstart.
2) Locate on the plot the time when the measured variable process reaches y63.2, which is the
point where y(t) has traveled 63.2% of the total change it is going to experience in response to
the controller output step. Label t63.2 as the point in time where y63.2 occurs.
3) The time constant, τP, is then estimated as the time difference between tYstart and t63.2. Since the
time constant marks the passage of time, it must have a positive value.

The value 63.2% is derived assuming the process dynamic character is exactly described by the linear
FOPDT model of Eq. 3.1. Details can be found in Section 13.3 (Deriving the τp = 63.2% of Process
Step Response Rule). In reality, no real process is exactly described by this model form. Nevertheless,
the procedure most always produces a time constant value sufficiently accurate for controller design
and tuning procedures.

Example: Gravity Drained Tanks

Figure 3.4 shows the same step response data from the gravity drained tanks used in the previous
KP calculation. With the pumped flow disturbance constant at 2.0 L/min, the controller output is
stepped from 50% up to 60% and the measured liquid level responds from an initial steady state
of 1.93 m up to a new steady state of 2.88 m.

28
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
Applying the time constant analysis procedure to this plot:

1) The time when the measured process variable starts showing a clear initial response to the
step change in controller output is tYstart. From the plot:

tYstart = 9.6 min

2) The measured process variable starts at steady state at 1.93 m and exhibits a total
response of ∆y = 0.95 m. Thus, y63.2 is computed:

y63.2 = 1.93 m + 0.632(∆y) = 1.93 m + 0.632(0.95 m) = 2.53 m

From the plot, y(t) passes through 2.53 m at:

t63.2 = 11.2 min

3) The overall time constant for this process response is then:

τP = t63.2 − tYstart = 11.2 min − 9.6 min = 1.6 min

Gravity Drained Tanks - Open Loop Step Test

Process: Gravity Drained Tank Controller: Manual Mode
3.0
Process Variable

2.8
y63.2 = 2.53 m
2.6
2.4 ∆y = 0.95 m
2.2
2.0
1.8
Controller Output

60

55
τP = 1.6 minutes
50

45
8 9 10 11 12 13 14 15 16 17 18
Time (mins)

tYstart t63.2

Figure 3.4 - τP computed from gravity drained tanks step test plot

Õ Õ Õ

Example: Heat Exchanger

Figure 3.5 shows the same step test data used in the previous KP calculation. With the warm oil
disturbance flow rate constant at 10 L/min, the controller output is stepped from 25% up to 35%
and the measured exit temperature displays a negative response as it moves from an original
steady state of 151.2°C down to a final steady state of 142.6°C.

29
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Heat Exchanger - Open Loop Step Test
Process: Heat Exchanger Controller: Manual Mode
152

Process Variable
150
148 y63.2 = 145.8 oC ∆y = -8.6 oC
146
144
Controller Output 142

35

30 τP = 1.0 minute
25

5 6 7 8 9 10 11 12 13
Time (mins)
tYstart t63.2

Figure 3.5 - τP computed from heat exchanger step test plot

Applying the time constant analysis procedure to this plot:

1) The time when the measured process variable starts showing a clear initial response to the
step change in controller output is tYstart. From the plot:

tYstart = 6.8 min

2) The measured process variable starts at steady state at 151.2°C and exhibits a total response
of ∆y = −8.6°C. Thus, y63.2 is computed:

y63.2 = 151.2°C + 0.632(∆y) = 151.2°C + 0.632(−8.6°C) = 145.8°C

From the plot, y(t) passes through 145.8°C at:

t63.2 = 7.8 min

3) The overall time constant for this process response is then:

τP = t63.2 − tYstart = 7.8 min − 6.8 min = 1.0 min

Õ Õ Õ

3.5 Apparent Dead Time, θP , From Step Test Data

Dead time is the time that passes from the moment the step change in the controller output is made
until the moment when the measured process variable shows a clear initial response to that change.
Dead time arises because of transportation lag (the time it takes for material to travel from one point
to another) and sample or instrument lag (the time it takes to collect, analyze or process a measured
variable sample).

30
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Dead time can also appear to exist in higher order processes simply because such processes
are slow to respond to a change in the controller output signal. Overall or apparent dead time, θP,
refers to the sum of dead times evident from all sources.
If θP > τP, tight control of a process becomes challenging, and the larger dead time is relative
to the process time constant, the worse the problem. For important loops, every effort should be made
to avoid introducing unnecessary dead time. This effort should start at the early design stages,
continue through the selection and location of sensors and final control elements, and persist up
through final installation and testing.

The procedure for estimating dead time from step response data is:

1) Locate on the plot tUstep, the time when the controller output step, ∆u(t), is made.

2) As in the previous τP calculation, locate tYstart, the time when the measured process variable starts
showing a clear initial response to this controller output step.

3) The apparent dead time, θP, is then estimated as the difference between tUstep and tYstart. Since dead
time marks the passage of time, it must have a positive value.

Example: Gravity Drained Tanks

Figure 3.6 shows the same step response data from the gravity drained tanks used in the previous
calculations. Applying the dead time analysis procedure to the step test plot of Fig 3.6:

Gravity Drained Tanks - Open Loop Step Test

Process: Gravity Drained Tank Controller: Manual Mode
3.0
Process Variable

2.8
2.6
2.4
2.2
2.0
1.8
Controller Output

60

55

θP = 0.4 minutes
50

45
8 9 10 11 12 13 14 15 16 17 18
tUstep Time (mins)

tYstart

Figure 3.6 - θP computed from gravity drained tanks step test plot

1) The step change in the controller output occurs at time tUstep (9.2 min).
2) As in the time constant calculation, the time when the measured process variable starts
showing a clear initial response to the step change in controller output is tYstart (9.6 min).

31
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 3) Thus, the apparent dead time for this process response is

θP = tYstart − tUstep = 9.6 min − 9.2 min = 0.4 min

Õ Õ Õ

Example: Heat Exchanger

Figure 3.7 shows the same step response data from the heat exchanger process used in the
previous calculations.

Heat Exchanger - Open Loop Step Test

Process: Heat Exchanger Controller: Manual Mode
152
Process Variable

150
148
146
144
142
Controller Output

35

30

25
θP = 0.3 minutes
5 6 7 8 9 10 11 12 13
tUstep Time (mins)

tYstart

Figure 3.7 - θP computed from heat exchanger step test plot

Applying the dead time analysis procedure to the step test plot of Fig 3.7:

1) The step change in the controller output occurs at time tUstep (6.5 min).

2) As in the time constant calculation, the time when the measured process variable, ystart, starts
showing a clear initial response to the step change in controller output is tYstart (6.8 min).

3) Thus, the apparent dead time for this process response is:

θP = tYstart − tUstep = 6.8 min − 6.5 min = 0.3 min

Õ Õ Õ

3.6 FOPDT Limitations - Nonlinear and Time Varying Behaviors

As detailed in this chapter, a critical step in controller design and tuning is the fitting of the linear
FOPDT model of Eq. 3.1 to dynamic process data. It may come as a surprise to learn, then, that no
real process has its dynamics exactly described by this model form. A linear FOPDT model is just too
simple to describe the higher order, time varying and nonlinear behaviors of the real world.

32
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Though only an approximation, and for some processes a very rough approximation, the
value of the FOPDT model is that it captures those essential features of dynamic process behavior
that are fundamental to control. When forced by a change in the controller output, a FOPDT model
reasonably describes the direction, how far, how fast and with how much delay the measured process
variable will respond.
The FOPDT model of Eq. 3.1 is called "first order" because it has only one time derivative.
The dynamics of real processes are more accurately described by models that are second, third or
higher order time derivatives. This notwithstanding, the simplifying assumption of using a first order
plus dead time model to describe dynamic process behavior is usually reasonable and appropriate for
controller tuning procedures. And often, a FOPDT model is sufficient for use as the model in model-
based strategies such as of feed forward, Smith predictor and multivariable decoupling control.

Time Varying Behaviors

The dynamic response of the FOPDT model does not change with time, though the dynamics of real
processes do. For a given KP, τP, θP and steady state value of u(t), the FOPDT model will compute the
same steady state value of y(t) tomorrow, next week and next year. If u(t) is stepped from this steady
state value, the FOPDT model will predict the same y(t) response with equal consistency.
Real processes, on the other hand, have surfaces that foul or corrode, mechanical elements
like seals or bearings that wear, feedstock quality or catalyst activity that drifts, environmental
conditions such as heat and humidity that changes, and a host of other phenomena that impact
dynamic behavior. The result is that real processes behave a little differently with every passing day.
The time varying (also called nonstationary) nature of real processes is important to
recognize because, for a particular operating level, the KP, τP, and θP that best describe the dynamics
will change over time. As we will see later, this means that the performance of a well tuned controller
will eventually degrade. It may take weeks or it may take months, but this result is difficult to avoid
with conventional controllers.
From a practical perspective, even if loop performance is only fair, in many instances that is
good enough. Hence, a slow degradation in performance is not always a pressing concern. For
important loops where tight control has a significant impact on safety or profitability, periodic
performance evaluation and retuning by a skilled practitioner is a justifiable expense.

Nonlinear Behaviors
Though a mathematician would classify time varying behaviors as nonlinear, here the term is reserved
to describe processes that behave differently at different operating levels. Figure 3.8 shows process
test data at two operating levels. First, a step test moves the process from steady state A to steady state
B. Then, a second step moves the process from steady state B to steady state C.
If individual FOPDT models were fit to these two data sets and the process was linear, then
the KP, τP, and θP computed for both sets would be identical. As shown in Fig 3.8, like most real
processes, the response is different at the different operating levels. If individual FOPDT models are
fit to these two data sets, one or more of the model parameters will be different. Hence, this process is
considered to be nonlinear.

33
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Example Nonlinear Behavior
Process: Custom Process Controller: Manual Mode
80
response shape is different

Process Variable
C
70 at different operating levels
60
B
50
A
70
Controller Output

65
60
55 even though controller
50 output steps are the same
2 4 6 8 10 12 14 16 18 20
Time (time units)

Figure 3.8 - A nonlinear process behaves differently at different operating levels

Figure 3.9 explores the nonlinear nature of the gravity drained tanks process. The bottom half
of the plot shows the controller output stepping in five uniform ∆u's from 40% up to 90%. The top
half of the plot shows the actual response of the measured process variable, liquid level in the lower
tank, along with the response of a linear FOPDT model.

Nonlinear Behavior of Gravity Drained Tanks

Model: First Order Plus Dead Time (FOPDT) File Name: TEST.DA
7
Measured Level (m)

6 nonlinear process
5
4 variable response
3
2 constant parameter
1 FOPDT model
Controller Output (%)

90
80
70
60
50
equal ∆u’s
40

0 5 10 15 20 25 30 35 40 45
Time

Figure 3.9 - Linear FOPDT model does not reasonably describe nonlinear behavior of
gravity drained tanks process across a range of operating levels

Each of the controller output steps are the same, so the linear FOPDT model exhibits five
identical responses across the entire range of operation. However, the model accurately describes the
behavior of the process only as it responds to the first controller output step from 40% up to 50%. By
the time the final steps are reached, the linear model and process behavior become quite different.
Thus, the gravity drained tanks process is nonlinear, and this is evident because the response of the
measured process variable changes as operating level changes.

34
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 The implication of nonlinear behavior is that a controller designed to give desirable
performance at the lower operating levels may not give desirable performance at the upper operating
levels. When designing and tuning controllers for nonlinear processes, it is important that the test
data be collected at the same level of operation where the process will operate once the control loop
is closed (where the set point is expected to be set).

3.7 Exercises
For all exercises, leave the controller in manual mode and use default values for noise level, disturbance
value and all other parameters of the case study.

Q-3.1 Click the Case Studies button on the LOOP-PRO main screen and from the pop-up list of
processes click on "Gravity Drained Tanks" to start the tanks simulation.

a) To generate dynamic process data, we need to change the controller output signal. This
moves the valve position, causing a manipulation in the flow rate of liquid into the top tank. In
this exercise we are interested in a step change in the controller output.

At the upper right of the draining tanks graphic on your screen, locate the white number box
below the Controller Output label. The most convenient way to step the controller output
value is to click once on this white number box. Click once on the controller output box and it
will turn blue. For the first step test, type 55 into the box and press Enter. This will cause the
controller output to change from its current value of 70% down to the new value of 55%,
decreasing the flow of liquid into the top tank and causing the liquid level to fall.
b) Watch as the process responds. When the measured process variable (liquid level) reaches its
new steady state, click on the “pause” icon on the toolbar above the graphic to stop the
simulation and then click on the "view and print plot" icon to create a fixed plot of the response.
Use the plot options as needed to refine your plot so it is well suited for graphical calculations
and then print it.
c) Using the methodology described in this chapter, use a graphical analysis to fit a first order plus
dead time (FOPDT) dynamic model to the process response curve. That is, compute from the
response plot the FOPDT model parameters: steady state process gain, KP, overall time constant,
τP, and apparent dead time, θP.
d) Repeat the above procedure for a second step in the controller output from 55% down to 40%.
e) The two steps in the controller output (70→55% and 55→40%) were both of the same size. Are
the model parameters the same for these two steps? How/why are they different?

Q-3.2 Click the Case Studies button on the LOOP-PRO main screen and from the pop-up list of
processes click on "Heat Exchanger" to start the simulation.

a) To generate dynamic process data, we will step the controller output signal. At the lower left
of the exchanger graphic, locate the white number box below the Controller Output label.
Click once on this box, type 49 and press Enter to change the controller output from its current
value of 39% up to the new value of 49%. This increases the flow of cooling liquid into the shell
side of the exchanger and causes the exit temperature of liquid on the tube side to fall.

35
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 b) Watch as the process responds. When the measured process variable (exit temperature) reaches
its new steady state, click on the “pause” icon on the toolbar above the graphic to stop the
simulation and then click on the "view and print plot" icon to create a fixed plot of the response.
Use the plot options as needed to refine your plot so it is well suited for graphical calculations
and then print it.

c) Using the methodology described in this chapter, use a graphical analysis to fit a first order plus
dead time (FOPDT) dynamic model to the process response curve. That is, compute from the
response plot the FOPDT model parameters: steady state process gain, KP, overall time constant,
τP, and apparent dead time, θP.

d) Repeat the above procedure for a second step in the controller output from 49% up to 59%.

e) The two steps in the controller output (39→49% and 49→59%) were both of the same size.
Are the model parameters the same for these two steps? How/why are they different?

Q-3.3 Start the jacketed reactor simulation (not the cascade case) by clicking on the Case Studies
button on the LOOP-PRO main screen and then clicking on "Jacketed Reactor."

a) To the right of the graphic, locate the white number box below the Controller Output label.
Click once on this box change the controller output from 42% up to a new value of 52%. This
increases the flow of cooling liquid through the cooling jacket side of the reactor and causes the
measured exit temperature of liquid from the reactor to fall.

b) Watch as the process responds. When the measured process variable (exit temperature) reaches
its new steady state, click on the “pause” icon on the toolbar above the graphic to stop the
simulation and then click on the "view and print plot" icon to create a fixed plot of the response.
Use the plot options as needed to refine your plot so it is well suited for graphical calculations
and then print it.

c) Using the methodology described in this chapter, use a graphical analysis to fit a first order plus
dead time (FOPDT) dynamic model to the process response curve. That is, compute from the
response plot the FOPDT model parameters: steady state process gain, KP, overall time constant,
τP, and apparent dead time, θP.

d) Repeat the above procedure for a second step in the controller output from 52% up to 62%.

e) The two steps in the controller output (42→52% and 52→62%) were both of the same size. Are
the model parameters the same for these two steps? How/why are they different?

36
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 4. Process Control Preliminaries

4.1 Redefining “Process” for Controller Design

As shown in Fig. 4.1, the feedback control loop consists of individual pieces: a controller, a final control
element, a process, and a sensor/transmitter. The sensor measures the process variable and feeds back the
signal to the controller. This measured value is subtracted from the set point to determine the controller
error. The controller uses this error in an algorithm to compute an adjustment signal to the final control
element. The final control element reacts to the controller output signal and changes the manipulated
variable in an effort to make the measured process variable equal the set point.
The final control element, process, sensor/transmitter and even controller all have individual
dynamic behaviors. That is, they all have their own gain, overall time constant and apparent dead time.
Awareness of this fact is important when specifying instrumentation for a new process. Also, if
undesirable dynamic behaviors of an existing control loop, such as a large dead time, can be traced to the
final control element or measurement sensor, then the purchase of new equipment or relocation of existing
equipment should be explored.
In general, a final control element and measurement sensor should be specified and installed to
respond immediately (add little dead time) and complete the response quickly (have a small time
constant). Note that qualifiers such as “little” and “quickly” are relative to the overall time constant of the
entire loop. A dead time of 9 minutes is large relative to a time constant of 10 minutes and is small relative
to a time constant of 1000 minutes.

manipulated measured
process process
controller controller
variable, m(t) variable, y(t)
error, e(t) output, u(t)

Final
Set Point Control
+- Controller Process
Element

Disturbance

feedback
signal, ym(t)

Measurement
Sensor/Transmitter

Figure 4.1 - The control loop from a designer's view point

While the final control element, process and sensor/transmitter all have individual dynamic
behaviors, from a controller’s viewpoint it is impossible to separate out these different behaviors.
Consider that a controller sends a signal out on one wire and sees the result of this action as a change in
the process variable when the measurement returns on another wire. From the controller’s viewpoint, the
individual gains, time constants and dead times all lump together into a single overall dynamic behavior.

37
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 manipulated variable is the controller’s
controller output, u(t)
controller “process”
error, e(t)

Set Point Final Control Element

+ Controller Process
-
Sensor/Transmitter

Disturbance

feedback signal is the

process variable y(t)

Figure 4.2 - The “process” of a control loop when tuning a controller

As we will learn, controller design and tuning proceeds in the same fashion regardless of whether
this lumped behavior exhibits a large or small gain, time constant or dead time. The procedure is
unaffected by whether a particular behavior is due to the process or sensor or final control element. No
matter which piece contributes a dominant influence, the combined or overall loop behavior must be
addressed.
Since controller design and tuning is the focus of the remainder of this manuscript, it is the
controller’s viewpoint that is taken. Consequently, for the remainder of this manuscript, “process
dynamics” refers to the combined behaviors of the final control element, process and sensor/transmitter as
illustrated in Fig. 4.2.

4.2 On/Off Control – The Simplest Control Algorithm

The simplest control law is on/off control. For on/off control, the final control element is either completely
open/on/maximum or completely closed/off/minimum. There are no intermediate values or positions for
the final control element in on/off control. One big disadvantage with pure on/off control is that
mechanical final control elements can experience significant wear as they continually and rapidly switch
from open to closed and back again.
To protect the final control element, a popular modification is the use of a dead band, which is a
zone bounded by an upper and a lower set point, or dead band limit. As long as the measured variable
remains between these limits, no changes in control action are made. Thus, if a valve is closed, it stays
closed until the measured process variable passes below the lower dead band limit. The valve will then
open completely and will remain open until the measured process variable eventually passes above the
upper dead band limit.
On/off with dead band is found in many places in our daily lives. A home heating system uses
this control law. So does an oven, a refrigerator and an air conditioner. All of these appliances cycle the
temperature under their control between an upper and low limit around a set point specified by the home
owner. Figure 4.3 illustrates the actions of a controller output and measured process variable in an on/off
with dead band control system.
As shown in the figure, the controller output is either completely on or completely off. The
change in the manipulated variable from one position to the other occurs whenever the measured process
variable crosses one of the dead band limits. Note that by using a dead band, the wear and tear on the final
control element is reduced but the size of the oscillations in the measured variable increases.

38
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Process: Custom Process Controller: Manual Mode
56 upper dead band limit

Process Variable
54
52
50
48
46
44
Controller Output
lower dead band limit
ON
55

50

OFF
45

250 300 350 400 450 500 550 600 650 700
Time (time units)

Figure 4.3 - Process under on/off control with dead band

Usefulness of On/Off Control

On/off controllers have the advantage of being easy to design. The biggest algorithm design issue is
specification of the dead band limits. Unfortunately, for many applications, such a control law is just too
limiting. Think about riding in a car that has on/off cruise control. This means the gas pedal can be either
fully depressed to the car floor or completely disengaged with no intermediate pedal positions possible. It
would be quite a ride indeed.
In general, on/off control is not sufficiently capable for most lab and production applications. The
remainder of this book explores the design and tuning of intermediate value control laws that offer this
necessary sophistication and consequently are widely used in the lab and plant.

4.3 Intermediate Value Control and the PID Algorithm

More powerful control algorithms that permit tighter control with less oscillation in the measured process
variable are required for a variety of process applications. These algorithms compute a complete range of
control actions between full on and full off. Not surprisingly, one requirement for implementation is a
final control element that, when signaled by the controller, can assume intermediate positions between full
on and full off. Example final control elements include process valves, variable speed pumps and
compressors, and electronic heating elements.
The most popular intermediate algorithm, the proportional-integral-derivative controller,
computes an intermediate value signal based on the current value of the control error (error equals set
point minus measurement). The basic PID algorithm is expressed as follows:

KC de(t )
u(t) = ubias + K C e(t ) +
1
424 3 τ ∫ e(t )dt + K Cτ D
14243 dt
(4.1)
proportional 1I42
4 43
4
integral derivative

where:
u(t) = controller output signal
ubias = controller bias or null value
e(t) = controller error; e(t) = ysetpoint – y(t)
y(t) = measured value of process variable

39
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 KC = controller gain (proportional tuning parameter)
τI = controller reset time (integral tuning parameter)
τD = controller derivative time (derivative tuning parameter)

The PID algorithm continually computes control actions u(t) in an attempt to drive error e(t) to zero.
As labeled in Eq. 4.1, each of the three terms works independently and with a slightly different
agenda. The proportional term computes a contribution to the control action based on the current size
of e(t) at the moment of measurement. No influence of past measurements or future trends is
considered in the proportional computation.
The integral term continually sums or accumulates e(t) over time. The integral continues to
grow as long as error is positive and begins shrinking when error becomes negative. Thus, the integral
term increases its influence when either positive or negative error persists for some time.
The derivative term looks at the slope or rate of change in error. Thus, its influence grows
when error is rapidly changing and seeks to slow down such movement. One result is that derivative
action tends to dampen oscillations in the measured process variable.
The PID algorithm can be implemented in P-Only, PI, PD and full PID forms. The design and
tuning of such controllers and the strengths and weaknesses of each P, I and D controller mode is the
focus of the next several chapters.

40
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 5. P-Only Control - The Simplest PID Controller
5.1 The P-Only Controller
The PID (proportional-integral-derivative) controllers are by far the most widely used family of
intermediate value controllers in the chemical process industry. The simplest in this family is called
proportional or P-Only control.
P-Only Control of Gravity Drained Tanks
Process: Gravity Drained Tank Controller: PID ( P= RA, I= off, D= off )

2.8
Level/Setpoint

2.6 offset
2.4
2.2 offset
2.0
1.8 design set point
60
Controller Output

58
56
54
52
50

10 15 20 25 30 35 40
Time (mins)
Tuning: Bias = 55.2, Gain = 11.5, Sample Time = 1.00

Figure 5.1 - Gravity drained tanks under P-Only control

Figure 5.1 shows the gravity drained tanks under P-Only control. The set point, the square wave in the
upper half of the plot, starts at a level of 2.4 m and steps up to 2.8 m followed by a step down to 2.0 m.
The measured level in the lower tank, the upper plot line distinguished by the noise or random error,
generally moves along with the changing set point though it displays some oscillation with the effort.
The controller output signal is the line in the lower half of the plot. As the term “intermediate
value control” implies, the controller output assumes many intermediate values between full open and full
closed as the controller works to cause the liquid level in the lower tank to track the changing set point.
The pumped disturbance flow (not shown) was constant at 2.0 L/min during the test.
This control system, typical of the family of PID controllers, repeats the measurement,
computation and action procedure at every loop sample time:
- a sensor measures the liquid level in the lower tank,
- this measurement is subtracted from the set point level to determine a control error,
- the control algorithm uses this error to compute a new controller output signal, which when
transmitted to the valve causes it to move to a new position,
- the change in valve position causes the flow rate of liquid into the top tank to change, which
ultimately causes a change in level in the lower tank.
The goal is to eliminate (or at least minimize) the controller error by making the measured level in the
lower tank equal the set point level.
One distinguishing feature of a P-Only controller, as shown in Fig. 5.1, is that it is only able to
make the measurement equal the set point when the set point is at the design value of 2.4 m. When set
point is not at the design value, offset occurs in most processes. This phenomenon is discussed later in this
chapter.

41
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 The P-Only control algorithm computes a controller output signal every loop sample time as:

u (t ) = u bias + K C e(t ) (5.1)

where u(t) is the controller output signal, ubias is the controller bias (also called the null value by some
manufacturers) and KC is the single tuning parameter called controller gain. Controller error, e(t), is
computed at every sample time as:
e(t) = ysetpoint − y(t) (5.2)

where ysetpoint is the set point and y(t) is the current value of the measured process variable.
Controller gain, KC, in Eq. 5.1 is the second "gain" we have discussed. Controller gain should be
distinguished from steady state process gain, KP. Rearrangement of Eq. 5.1 reveals that controller gain,
KC, describes how the controller output signal changes given a change in the controller error. A larger KC
means the controller output will change more for a given error. Similar to KP , controller gain also has a
size, sign and units.

5.2 The Design Level of Operation

Nonlinear processes have a dynamic process behavior that changes as operating level changes. Because
real processes are nonlinear, it is important to recognize that a controller should be designed for a specific
level of operation. When designing a cruise control system for a car, would it make sense to perform
dynamic modeling studies when the car is traveling twice the normal speed limit while going down hill in
a hurricane? Of course not.
For a proper controller design, first determine where you expect the set point to be set during
normal operation. Second, determine typical values for the important disturbance variables. This is
important because as disturbance variables change, so can the values for the FOPDT model parameters
KP, τP and θP that best describe the dynamic behavior of the process.
With this information in hand, dynamic testing should occur as close as practical to the design
value of the measured process variable when the disturbances are quiet and near their typical values. Thus,
the design level of operation for a cruise control system is when the car is traveling near the normal speed
limit on flat ground on a relatively calm day. The design level of operation for the P-Only control system
shown in Fig 5.1 is a level in the lower tank near 2.4 m when the pumped disturbance flow is 2.0 L/min.

5.3 Understanding Controller Bias, ubias

Suppose Eq. 5.1 is the algorithm used for cruise control in an automobile and u(t) is the flow of gas to the
engine computed by the controller. Further suppose that the velocity set point for the car is 70 kph and
the current measured velocity is also 70 kph. Since y(t) equals ysetpoint, then e(t) equals zero and Eq. 5.1
becomes:
u(t) = ubias (5.3)

If ubias is zero, then Eq. 5.3 says that when set point equals measurement, the flow of gas to the
engine, u(t), is also zero. This makes no sense. Clearly if the car is traveling 70 kph then some baseline
flow of gas is going to the engine. This baseline value of the controller output is called the bias or null
value. In this example, the bias is the flow of gas that, in open loop, causes the car to travel the design
velocity of 70 kph when the disturbance variables are at their normal or expected values.
Consider a second example of the gravity drained tanks. Here suppose the measured level equals
the set point value, so e(t) in Eq. 5.1 equals zero. If no liquid is flowing into the top tank, the tanks will
empty. Hence, to maintain the liquid level in the lower tank at set point, some baseline flow rate of liquid,
or ubias, must always be entering the top tank.

42
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Similarly for the heat exchanger, if no cooling water is flowing through the shell side, then the
hot liquid will exit the exchanger at the same temperature that it enters. To achieve any amount of cooling,
some baseline volume of cooling water must always be flowing through the exchanger. This requirement
of a controller output baseline value holds true for most all processes.
Thus, when a process is under P-Only control and the set point equals the measurement, some
baseline or bias value of the controller output must exist or the measured process variable will drift from
set point. This bias value of the controller output is determined from the design level of operation of the
process to be controlled. Specifically, ubias is the value of the controller output that, in open loop (manual
mode), causes the measured process variable to maintain steady state at the design level of operation
when the process disturbances are at their design or expected values.
A P-Only controller bias is assigned a value as part of the controller design and remains fixed
once the controller is put in automatic. Some commercial manufacturers call the bias the null value.

5.4 Controller Gain, KC , From Correlations

The P-Only controller has the advantage of having only one adjustable or tuning parameter, KC,, that
describes how aggressive the controller output will move in response to changes in controller error, e(t).
For a given value of e(t) in the P-Only algorithm of Eq. 5.1, if KC is small, then the amount added to ubias
is small and the controller responds sluggishly. If KC is large, then the amount added to ubias is large and
the controller responds aggressively. Thus, KC can be adjusted or tuned for each process to make the
controller more or less active in its actions when measurement does not equal set point.
Designing any controller from the family of PID algorithms entails the following steps:
- specifying the design level of operation,
- collecting dynamic process data as near as practical to this design level,
- fitting a FOPDT model to the process data to obtain model parameters KP, τP and θP,
- using these model parameters in a correlation to obtain initial controller tuning values.
Final tuning may require some trial and error once the controller is online so loop performance will match
that desired by the control designer.
The tuning correlations available in Design Tools for all controllers of the PID family are the
Internal Model Control (IMC) relations. The exception is the P-Only controller, as no IMC correlation can
be derived for this simple controller form.
As an alternative, Design Tools computes KC using the integral of time-weighted absolute error
(ITAE) tuning correlation for set point tracking (also called servo control) as:
0.20
KC = (τ p / θ p ) 1.22 (5.4)
Kp
Although not automatically computed by Design Tools, the FOPDT model parameters can be used in the
ITAE for disturbance rejection (also called regulatory control) correlation:
0.50
KC = (τ p /θ p ) 1.08 (5.5)
Kp
These correlations provide an initial guess or starting point for final controller tuning. Even with this
formal controller design procedure, final tuning requires online trial and error because:
- the control designer may desire performance different from that provided by the correlation,
- the FOPDT model used for tuning may not match the actual dynamic behavior of the plant,
- performance may need to be balanced over a range of operation for nonlinear processes,
- performance may need to be balanced for best set point tracking and disturbance rejection.
Ultimately, it is the designer who defines what “best” control performance is for an application.

43
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 When two correlations provide quite different initial controller gain estimates, the conservative
approach is to start with the smallest KC value. This will give the least aggressive (most sluggish)
performance. If the resulting performance is too sluggish in rejecting disturbances and tracking changes in
the set point, KC should be increased. Conversely, if the process responds quickly and oscillates to an
uncomfortable degree, KC is too large and should be reduced.

5.5 Reverse Acting, Direct Acting and Control Action

Controller gain, KC, computed from tuning correlations such as Eq. 5.4 and 5.5 will always have the same
sign as the process gain, KP. Time constant and dead time cannot affect the sign of the controller gain in
these correlations because parameters of time are always positive. Thus, a process with a positive KP will
have a controller with a positive KC.
For a process with a positive KP, the process variable increases when the controller output
increases. Thus, in closed loop, if the process variable is too high, the controller has to decrease the
controller output signal to correct the error. The controller action is the reverse of the problem. Thus,
when a process has a positive KP and thus KC, the controller must be reverse acting. Conversely, when KP
and thus KC are negative, the controller must be direct acting, or:

KP and KC positive → reverse acting

KP and KC negative → direct acting

In most commercial controllers, a positive value of the controller gain is always entered. The sign
(or action) of the controller is then assigned by specifying that the controller is either reverse or direct
acting to indicate a positive or negative KC respectively. If the wrong control action is entered, the
controller will quickly drive the final control element to full on/open or full off/closed and remain there
until a proper control action entry is made.

5.6 Set Point Tracking in Gravity Drained Tanks Using P-Only Control
P-Only controller design requires specifying a design level of operation, a controller bias and controller
gain. Here we explain the design of the P-Only controller used in Fig. 5.1. The design level of operation
for this example is a measured level in the lower tank of 2.4 m while the pumped flow disturbance during
normal operation is expected to be about 2.0 L/min.
The controller bias is determined by searching for the value of the controller output that, in
manual mode, causes the measured level to steady at 2.4 m when the disturbance is at its design value.
Through trial and error, the bias value of the controller output, or ubias, is determined to be 55.2% (please
note that in typical industrial operations, such a three significant digit accuracy for the controller output
far exceeds realistic expectations).
To verify that this bias is used, note that ysetpoint and y(t) both equal 2.4 m (the design value) for
the first few minutes in Fig 5.1. Since e(t) equals zero for these minutes, then Eq. 5.1 says u(t) should
equal the ubias value of 55.2%. The controller output trace in Fig. 5.1 reveals that this is true.
To compute KC for the P-Only controller, we need a FOPDT model fit of dynamic process data
collected around the design level of operation. We take advantage of the FOPDT fit detailed in Fig. 3.2,
3.4 and 3.6, which show a step test for the gravity drained tanks. In these figures, the process is initially
at steady state with the controller output at 50% while the pumped flow disturbance is constant at 2.0
L/min. The controller output is stepped from 50% up to 60%, causing the measured tank level to rise
from its initial steady state value of 1.93 m up to a new steady state level of 2.88 m.

44
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Because the gravity drained tanks is nonlinear, a best design would fit dynamic process data
collected from above and below the design level of operation so as to average the nonlinear effects.
The plots contain dynamic data equally distributed above and below the design level of 2.4 m,
making this FOPDT fit well-suited for our design. The analysis presented in Chapter 3 yields the
model parameters:
Process Gain, KP = 0.095 m/%
Time Constant, τP = 1.6 min
Dead Time, θP = 0.40 min
Using these parameters in the ITAE for set point tracking correlation of Eq. 5.4 produces an
initial KC estimate:
1.22
0.20 ⎛ 1.6 ⎞
KC = = 11.5 %/m
0.095 ⎜⎝ 0.40 ⎟⎠

Because KC is positive, the controller must be specified as reverse acting. As discussed above, the bias is
55.2%. Thus, the P-Only controller is as follows:

u(t) = 55.2% + 11.5 e(t) (5.6)

Again please recognize that there are more significant digits used in Fig 5.6 than can be realistically
obtained from typical plant data.
The performance of this controller in tracking set point changes is pictured in Fig. 5.1. The upper
right corner of the plot displays information about the controller: PID (P= RA, I= off, D= off). Since the
integral and derivative terms are off, this confirms a P-Only reverse acting controller. As shown in Fig.
5.1, the level in the lower tank does not track the first set point step all the way up to 2.8 m. This sustained
error between the measurement and set point is called offset and is discussed in the next section. The
measured level then responds well and exhibits no offset when the set point returns to the design level of
operation of 2.4 m. With the final set point step down to 2.0 m, the controller again exhibits offset.

5.7 Offset - The Big Disadvantage of P-Only Control

The biggest advantage of P-Only control is that there is only one tuning parameter to adjust, so it is
relatively easy to achieve a “best” final tuning. The disadvantage is that this control algorithm permits
offset. Offset occurs in most processes under P-Only control when the set point and/or disturbances are at
a value other than that used as the design level of operation (that used to determine ubias).
Consider the P-Only controller of Eq. 5.6. If y(t) is steady at ysetpoint, then e(t) is steady at zero. If
e(t) is steady at zero, then controller output u(t) is steady at the ubias value of 55.2%. And if u(t) is steady at
55.2% (and assuming the disturbance is constant at 2.0 L/min), then the lower tank will steady at the
design level of 2.4 m. We know this because the ubias value of 55.2% was determined by trial and error as
the controller output in manual mode that causes the lower tank to steady at 2.4 m. Thus, under P-Only
control, the measured process variable can be steady at the design set point and no sustained error or offset
will be present.
How can Eq. 5.6 produce a value for u(t) that is different from ubias at steady state? The only way
this can happen is if e(t) is not steady at zero. That is, whenever the desired level of operation is other than
the design value of 2.4 m, there must be a steady state error so u(t) can assume a value other than the ubias
value of 55.2%. This sustained error is called offset.

45
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 As controller gain increases, the offset will decrease. Unfortunately, as controller gain increases,
the oscillations in the measured process variable will increase and can even go unstable. The impact of KC
on offset and oscillatory behavior is illustrated in Fig. 5.2, which shows a step set point change for the
gravity drained tanks under P-Only control for two different controller gains.

Impact of KC on Offset and Oscillatory Behavior

Process: Gravity Drained Tank Controller: PID ( P= RA, I= off, D= off )
3.0

2.8
PV/Setpoint

2.6

2.4 offset
offset
2.2

KC = 11.5 KC = 40
75
Controller Output

70
65
60
55
50
45

10 20 30 40 50 60 70
Time (mins)
Tuning: Bias = 55.2, Gain = 40.0, Sample Time = 1.00

Figure 5.2 – P-Only offset decreases and oscillations increase as controller gain increases

5.8 Disturbance Rejection in Heat Exchanger Using P-Only Control

Whether the control objective is set point tracking or disturbance rejection, controller design always
begins by determining the design level of operation. Dynamic process data is then collected as near as
practical around this level. For this study, a constant measured exit temperature of 147°C is desired. The
control objective is to reject disturbances that occur when the warm oil flow rate changes, causing a
change the mixed stream temperature entering the tube side of the exchanger. The warm oil disturbance
flow is normally about 10 L/min but spikes as high as 20 L/min on occasion.
The bias is the value of the controller output that, in open loop, causes the measured exit
temperature to steady at 147°C when the disturbance flow is 10 L/min. Through trial and error, ubias is
found to be 29.2% (again, for industrial operations, this three significant digit accuracy far exceeds
realistic expectations). As shown for the first few minutes in Fig 5.3, when y(t) and ysetpoint both equal
147°C and the disturbance equals 10 L/min, then u(t) indeed equals the ubias value of 29.2%.
Even though disturbance rejection is the goal, it is still controller output to measured process
variable dynamics that must be fit with the FOPDT model. We take advantage of the FOPDT fit detailed
in Fig. 3.3, 3.5 and 3.7 for the heat exchanger. In these figures, the process is initially at steady state
with the controller output at 25% while the warm oil disturbance flow is constant at 10 L/min. The
controller output is stepped from 25% up to 35%, causing the measured exit temperature to fall from
its initial steady state value of 151.2°C down to a new steady state level of 142.6°C.

46
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Because the heat exchanger is nonlinear, a best design would fit dynamic process data
collected from above and below the design level of operation to average out the nonlinear effects.
The plots contain dynamic data equally distributed above and below the design level of 147°C,
making this FOPDT fit well-suited for our design. The analysis presented in Chapter 3 yields the
model parameters:
Process Gain, KP = −0.86 °C/%
Time Constant, τP = 1.0 min
Dead Time, θP = 0.3 min

The ITAE correlation of Eq. 5.5, intended for disturbance rejection applications, results in the
controller gain:
1.08
0.50 ⎛ 1.0 ⎞
KC = = − 2.1 %/°C
−0.86 ⎜⎝ 0.3 ⎟⎠

As expected, KC carries the same negative sign as KP. Thus, we choose direct acting as the
action of the P-Only controller. Using Eq. 5.1, the P-Only controller is shown below in Eq. 5.7:

u(t) = 29.2% − 2.1 e(t) (5.7)

Figure 5.3 shows the performance of this controller in rejecting step changes in the warm oil
disturbance flow rate. The set point is held constant throughout the experiment at the design operating
level of 147°C. As expected, offset equals zero whenever the set point and disturbance are at their design
values. When the disturbance flow rate steps from 10 L/min up to 20 L/min, however, offset results even
though the set point remains at the design value.

Impact of Kc on Disturbance Rejection

Process: Heat Exchanger Controller: PID ( P= DA, I= off, D= off )
150
offset offset
Disturbance Controller Output PV/Setpoint

148
146
144

KC = - 2.1
35
KC = - 1.0
30
25

20

10

5 10 15 20 25 30 35 40 45 50 55 60 65 70 75
Time (mins)
Tuning: Bias = 29.2, Gain = -1.00, Sample Time = 1.00

Figure 5.3 - Disturbance rejection performance of heat exchanger under P-Only control
for two different values of controller gain while set point is constant

47
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Half way through the experiment, the (absolute value of the) controller gain is decreased from the
ITAE design value of −2.1 %/°C to −1.0 %/°C. The disturbance flow rate step is then repeated. Similar to
the controller gain investigation shown in Fig. 5.2 for gravity drained tanks set point steps, the offset
increases and the oscillatory nature of the response decreases as the (absolute value) controller gain is
decreased.

5.9 Proportional Band

Some instrumentation manufacturers use different terminology for the controller gain tuning parameter.
The popular alternative found in the marketplace is proportional band, PB. If the measured and
manipulated variable of a process have units of percent and both can range from 0 to 100%, then the
conversion between gain and proportional band is:

100
PB = (5.8)
KC

While the examples in this book assign engineering units to the measured process variable, most
commercial installations use units of % for both controller output and the process variable.

5.10 Bumpless Transfer to Automatic

Commercial controllers with a “bumpless transfer" feature achieve a smooth transition to closed loop by
automatically setting values for the controller bias and set point when the controller is put in automatic.
Specifically, when the control loop is closed:
1) the bias is set equal to the current value of the controller output, and
2) the set point is set equal to the current value of the measured process variable.
Hence, when the controller is put in automatic, there is no controller error and the bias is properly set to
produce no offset. As a result, no immediate control action is necessary that would “bump” the measured
process variable.

5.11 Exercises
Unless otherwise directed, use default values for noise level and all other simulation parameters.

Q-5.1 For the gravity drained tanks, the level in the lower tank is to remain constant at 3.0 m. Your
process operator tells you that disturbances are a problem because the pumped flow
disturbance, normally constant at 2.0 L/min, occasionally spikes to 5.0 L/min. Your objective
is to explore a P-Only controller designed to reject these disturbances. As a first step, record
the design set point for your P-Only controller.
a) Adjust the controller output signal to move the process to the design level of operation (a
measured process variable of 3.0 m when the major disturbance is 2.0 L/min). Record the bias
value for your P-Only controller.
b) Perform an appropriate open loop step test and use the dynamic response data to estimate a
first order plus dead time (FOPDT) model for this process. (Hint: even though disturbance
rejection is the goal, it is still the controller output that is stepped. Because the process is
nonlinear, it is good practice to include data from above and below the design value of the
measured process variable).

48
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 c) Use these FOPDT model values in the ITAE for disturbance rejection correlation to compute
a P-Only controller gain, KC. Be sure to specify the sign, magnitude and units of KC.
d) Using your design set point, bias and controller gain, implement a P-Only controller.
Generate plots showing the performance of the controller in rejecting step increases in the
pumped disturbance flow rate from 2.0 L/min up to 5.0 L/min and back again, allowing the
response to settle after each step.
e) Double your KC computed in part d, step increases in the pumped disturbance flow rate from
2.0 L/min up to 5.0 L/min and back again, and generate a new set of plots. Halve your KC
from part d and repeat the experiment again. Using these plots, discuss how the magnitude of
KC impacts offset and the oscillatory nature of the response.

Q-5.2 Design a P-Only controller for the heat exchanger where the design level of operation is a
measured exit stream temperature of 133°C when the warm oil disturbance flow is at its
design value of 10 L/min. Several times a day, the exchanger must respond to requests to step
the set point from 133°C up to 138°C. Begin by recording the design set point.
a) Adjust the controller output signal to move the process to the design level of operation (a
measured process variable of 133°C when the major disturbance is 10 L/min). Record the
bias value for your P-Only controller.
b) Perform an appropriate open loop step test and use the dynamic response data to estimate a
first order plus dead time (FOPDT) model for this process. (Hint: the process is nonlinear so
include data from above and below the design value of the measured process variable).
c) Use these FOPDT model values in the ITAE set point tracking correlation to compute a P-
Only controller gain, KC. Be sure to specify the sign, magnitude and units of KC.
d) Using your design set point, bias and controller gain, implement a P-Only controller.
Generate a plot showing the performance of the controller in tracking steps in the set point
from 133°C up to 138°C and back again, allowing the response to settle after each step.
e) Fine tune your controller gain by trial and error and search for the “best” KC that balances
offset with oscillatory behavior when tracking this set point step. Remember that as the
designer, you define what constitutes “best” performance. As well as generating a plot of
your final tuning, be sure to explain what criteria you used to define “best.”
f) Using your “best” KC, generate a plot showing a step change in set point from 133 oC up to
138°C and back again, allowing the response to settle after each step. Follow that with a step
decrease from 133°C down to 128°C and back again. Using the plot to support your
argument, discuss how the nonlinear nature of this process impacts controller performance for
set point tracking.

Q-5.3 For the jacketed reactor (not the cascade case), the design level of operation is a measured
reactor exit temperature of 92°C when the cooling jacket inlet temperature (a disturbance) is
at its design value of 50 L/min. During product change over, which occurs once a shift, the
reactor temperature must be changed from 92°C up to 95°C. Your assignment is to design a
P-Only controller to track this set point change.

49
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
a) Determine and record the design set point and controller bias value for your controller.
b) Perform an appropriate open loop step test and use the dynamic response data to estimate a
first order plus dead time (FOPDT) model for this process. Be sure the response plot includes
data from above and below the design level of operation.

c) Use these FOPDT model values in the ITAE set point tracking correlation to compute a P-
Only controller gain, KC. Be sure to specify the sign, magnitude and units of KC.
d) Using your design set point, bias and controller gain, implement a P-Only controller.
Generate a plot showing the performance of the controller in tracking steps in the set point
from 92°C up to 95°C and back again, allowing the response to settle after each step.
e) Fine tune your controller gain by trial and error and search for the “best” KC that balances
offset with oscillatory behavior when tracking this set point step. Remember that as the
designer, you define what constitutes “best” performance. As well as generating a plot of
your final tuning, be sure to explain what criteria you used to define “best.”

50
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 6. Automated Controller Design Using Design Tools

6.1 Defining Good Process Test Data

Whether you are fitting a model to process test data for controller tuning, for process simulation using
Custom Process, or for developing models for advanced strategies such as feed forward or the Smith
predictor, the answer to all five of these questions about your data should be "yes," and ultimately, it is
your responsibility to consider all of these five steps to ensure success.

1. Was the process at steady state before data collection started?

Suppose a controller output change forces a dynamic response in a process, but the data file only
shows the tail end of the response without showing the actual controller output change that caused
the dynamics in the first place. Design Tools will indeed fit a model to this data, but it will skew the fit
in an attempt to account for an unseen "invisible force." This model will not be descriptive of your
actual process and hence of little value for control. To avoid this problem, it is important that data
collection begin only after the process has settled out. Design Tools can then properly account for all
process variations when fitting the model.

2. Is the first point in the test data equal to the steady state value of step 1?
Asking Design Tools to assume important facts about your data can lead to unfortunate results if the
assumptions are wrong. Knowledge of the initial steady state of the process prior to data collection is
fundamental to computing process and controller gain for any software tool. Thus, you must assume
responsibility for verifying the initial steady state. Design Tools helps you by using the first data point
in your data file as representative of the initial state of the process. But if noise or random error has
caused this first point to shift from a reasonable initial value, edit the data using a Design Tools utility
and adjust this data point by hand based on your knowledge of the process.

3. Did the test dynamics clearly dominate the process noise?

When generating dynamic process data, it is important that the change in controller output cause a
response in the process that clearly dominates the measurement noise. We suggest defining a noise
band as ±3 standard deviations of the random error around the process variable during steady
operation. Then, when during data collection, the change in controller output should force the process
variable to move at least ten times this noise band (the signal to noise ratio should be greater than ten).
If you meet or exceed this requirement, your data will be rich in the dynamic information needed for
controller design.

4. Were the disturbances quiet during the dynamic test?

It is essential that the test data contain process variable dynamics that have been clearly (and in the
ideal world exclusively) forced by changes in the controller output as discussed in step 3. Dynamics
caused by unmeasured disturbances can seriously degrade the accuracy of an analysis because Design
Tools will model those behaviors as if they were the result of changes in the controller output signal.
In fact, a model fit can look perfect, yet a disturbance that occurred during data collection can cause
the model fit to be nonsense. If you suspect that a disturbance event has corrupted test data, it is
conservative to rerun the test.

5. Did the model fit appear to visually approximate the data plot?
This is perhaps the easiest of the steps because Design Tools will display a plot that shows the model
fit on top of the data. If the two lines don't look similar, then the model fit is suspect. Of course, as
discussed in step 4, if the data has been corrupted by unmeasured disturbances, the model fit can look
great yet the usefulness of the analysis can be compromised.

51
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
6.2 Limitations of the Step Test
Chapter 3 explores generating dynamic process data using a step test, where the controller output is
stepped from one constant value to another, causing the measured process variable to move from one
steady state to a new steady state. As shown in that chapter, step tests are useful because FOPDT
(first order plus dead time) models can be fit to the response plot by direct graphical analysis.
Unfortunately, the step test is simply too limiting to be useful in many practical applications.
The drawback is that it takes the process away from the desired operating level for a relatively long
period of time. Recall that when collecting data for a nonlinear process using a step test, best practice
is to move the measured process variable to one side of the design level of operation and then step the
controller output so that response data is centered on average around the design level. As a
consequence, step tests in the plant can result in significant off-spec product that may need
reprocessing or even disposal.
A second problem common to all open loop tests, including those discussed in the next
section, is that in production situations, operating personnel may simply be unwilling to open a loop
(put the controller in manual) “just” to generate dynamic process data. In this surprisingly common
situation, closed loop testing must be performed as discussed toward the end of this chapter.

6.3 Pulse, Doublet and PRBS Test

Popular open loop experiments used to generate dynamic process data beyond the step test include
the pulse, doublet and pseudo-random binary sequence (PRBS) tests. Unlike the step test, model
parameters cannot be fit by a simple graphical analysis of the process response plots from these tests.
In fact, a computer program such as LOOP-PRO’s Design Tools is required for data analysis.

Pulse Testing
A pulse test, shown in Fig. 6.1, can be thought of as two step tests performed in rapid succession. The
controller output is stepped up, and as soon as the measured process variable shows a clear response,
the controller output is returned to its original value. Ordinarily, the process does not reach steady
state before the return step is made.

Open Loop Pulse Test

Process: Custom Process Controller: Manual Mode

60
Pulse Test
Process Variable

55

50

60
Controller Output

55

50

0 5 10 15 20
Time (mins)
Time

Figure 6.1 - Controller output pulse with measured process variable response

52
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Pulse tests have the desirable feature of starting from and returning to an initial steady state.
Unfortunately, they only generate data on one side of this steady state (which presumably is the
design level of operation) and this is not best practice when nonlinear processes are being studied.
The doublet test solves this problem.

Doublet Testing
A doublet test, shown in Fig. 6.2, is two pulse tests performed in rapid succession and in opposite
direction. The second pulse is implemented as soon as the process has shown a clear response to the
first pulse. The process does not ordinarily respond to steady state for either pulse.

Open Loop Doublet Test

Process: Custom Process Controller: Manual Mode
55
Doublet
Process Variable

50

45

55
Controller Output

50

45

0 5 10 15 20
Time (mins)
Time

Figure 6.2 - Controller output doublet with measured process variable response
(popular with practitioners)

The doublet test offers several benefits, including:

- starting from and returning reasonably quickly to the design level of operation,
- producing data both above and below the design level, and
- having a relatively small maximum deviation in the measured process variable from the initial
steady state, thus minimizing off-spec production.

For these reasons, many industrial practitioners find the doublet to be the preferred method of
generating open loop dynamic process data.

PRBS Testing
A pseudo-random binary sequence (PRBS) test, shown in Fig. 6.3 is characterized by a sequence of
controller output pulses that are uniform in amplitude, alternating in direction, and of random
duration. The "pseudo" is in the name because true random behavior is a theoretical concept that is
unattainable by computer algorithm. Thus, in practice, approximate or pseudo-randomness using a
random number generator must suffice for determining the duration of each pulse.
The PRBS test permits generation of useful dynamic process data while causing the smallest
maximum deviation in the measured process variable from the initial steady state. Since this implies a
minimum of off-spec production, PRBS, in theory anyway, is the most desirable of the open loop
methods. Unfortunately, proper PRBS experiment design presents practical difficulties that often
counterbalance this benefit.

53
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Open Loop PRBS Test
Process: Custom Process Controller: Manual Mode
55
PRBS

Process Variable
50

45

55
Controller Output
50

45

0 5 10 15 20
Time (mins)
Time

Figure 6.3 - Controller output pseudo-random binary sequence with process variable response

PRBS experiment design requires specification of several characteristics of the controller

output signal trace, including the controller output:
- initial value,
- pulse amplitude,
- average pulse duration, and
- standard deviation of the random change in pulse duration around this average.
The length of the experiment itself must also be specified. Since an optimum experiment design
requires knowledge of the process gain, time constant and dead time, the very values you are
performing the experiment to determine, the design logic becomes somewhat circular. If you
ultimately perform the experiment a number of times in a search of a “best” test, you may well have
been better off simply performing the quick and practical doublet test

6.4 Noise Band and Signal to Noise Ratio

When generating dynamic process data, it is important that the change in the controller output signal
causes a response in the measured process variable that clearly dominates the measurement noise.
One way to quantify the amount of noise in the measured process variable is with a noise band.
As illustrated in Fig. 6.4, a noise band is based on the standard deviation of the random error
in the measurement signal when the controller output is constant and the process is at steady state.
Here we define noise band as ±3 standard deviations of the measurement noise around the steady state
of the measured process variable (99.7% of the signal trace is contained within the noise band). While
other definitions of the noise band have been proposed, this definition is conservative when used for
controller design.
Employing this definition for generating dynamic process data, the change in controller
output should cause the measured process variable to move at least ten times the size of the noise
band. Expressed concisely, the signal to noise ratio should be greater than ten. In Fig. 6.4, the noise
band is 0.25°C. Hence, the controller output should be moved far and fast enough during a test to
cause the measured exit temperature to move at least 2.5°C. This is a minimum specification. In
practice it is conservative to exceed this value.

54
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Noise Band of Heat Exchanger PV
Process: Heat Exchanger Controller: Manual Mode

140.2

Exit Temp (deg C)

140.1
140.0
139.9
139.8
139.7
Controller Output
± 3σ of random error is 0.25 oC
39.5

39.0

38.5 controller output is constant

10 12 14 16 18 20 22 24 26 28 30
Time (mins)

Figure 6.4 - Noise band encompasses ± 3 standard deviations of the measurement noise

6.5 Automated Controller Design Using Design Tools

Design Tools offers two tools for controller design and analysis. One is the dynamic model fitting
tool that automatically fits dynamic models to process data. Models can be fit to data generated by
LOOP-PRO, other software packages, and perhaps most important, to data generated by real
processes in the lab or plant.
Figure 6.5 shows a portion of a data file in proper format for use in Design Tools. The file
must contain at least three columns, one for time stamp data, one for manipulated variable data
(which is the controller output for PID algorithm design), and one for the measured process variable
data. The data must be ASCII text with entries separated by tabs, commas or spaces. Every space in a
column must have an entry; there can be no blank spaces.

Time Controller Output Process Variable

0.00 70.0 4.00
0.15 70.0 4.01
0.30 80.0 3.99
0.45 80.0 4.03
0.60 80.0 4.09
0.75 80.0 4.17

process must be at steady state first PV value must equal

when data collection begins the true initial steady state

Figure 6.5 – Example process data file used by Design Tools

55
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
The linear dynamic models available in the Design Tools library include:
First Order Plus Dead Time (FOPDT)
First Order Plus Dead Time Integrating
Second Order Plus Dead Time (SOPDT) Overdamped
Second Order Plus Dead Time Overdamped with Lead Time
Second Order Plus Dead Time (SOPDT) Underdamped

The fit routine systematically searches for the model parameters that minimize the sum of
squared errors (SSE) between the response contained in the measured data and the response predicted
by the model being fit when it is forced by the manipulated variable data in the file. With i indicating
any one of the N total data points in the set, the SSE is expressed:
N
SSE = ∑ [Measured Data i − Model Data i ]2 (6.1)
i=1

In general, the smaller the SSE, the better the model describes the data. To obtain a meaningful fit, it
is essential to recognize that:

• the process must be at steady state before collection of dynamic data begins,
• the first data point in the file must equal this initial steady state value.

If these conditions are not met, the model fit will be incorrect and of little use.
When your goal is controller design and tuning, a FOPDT model should be fit so the popular
controller tuning correlations and design rules-of-thumb can be exploited. The other dynamic models
in the library are useful when constructing advanced controller architectures such as feed forward,
Smith predictor and model predictive control (MPC) algorithms. They are also useful when accurate
simulation of dynamic process behavior using Custom Process is the goal.
The second feature of Design Tools is the controller tuning tool. Using the results of a
successful FOPDT model fit, tuning values for P-Only, PI and PID controllers are computed. The
library of popular tuning correlations is based on the internal model control (IMC) relations. These are
an extension of the popular lambda tuning correlations and include the added sophistication of directly
accounting for dead time in the tuning computations.

Example: Heat Exchanger Doublet Test

Section 5.8 presented a P-Only controller design for the heat exchanger using a graphical fit of step
test data. The design level of operation for that study was a measured exit temperature of 147°C. The
control objective focused on disturbance rejection because the warm oil disturbance flow rate,
normally about 10 L/min, was said to occasionally spike as high as 20 L/min. We learned in that
study that in open loop, a controller output of 29.2% causes the measured exit temperature to steady
at 147°C when the disturbance is at its design value of 10 L/min. In fact, this is how we determined
the controller bias in that example.
In this example we fit a FOPDT model and tune a P-Only controller using a doublet test and
Design Tools, and compare the results with the graphical step test analysis of Section 5.8. As shown
in Fig.6.6, we start with the heat exchanger at steady state at the design level of operation.

56
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 While saving data to file, we pulse the controller output first up 3% and then down 3% from
the initial value of 29.2% (pulses of this size move the measured process variable across a range of
nonlinear operation similar to that of the previous step test). After each pulse, we wait for a clear
response in the measured process variable but do not wait for steady state.
The data file is then read into Design Tools and the columns containing time,
manipulated variable (controller output) and process variable data are correctly labeled. The
FOPDT model is selected from the library and the fit routine is executed. The model fit is also
shown in Fig.6.6.

FOPDT Fit of Heat Exchanger Doublet Test Data

Model: First Order Plus Dead Time (FOPDT) File Name: PROCESS.DAT
150
measured exit
Process Variable

149
148 temperature data
147
146
145
144
34
Controller Output

32 FOPDT model fit

30
28
26

0 5 10 15
Time

Gain (K) = -0.90, Time Constant (T1) = 1.14, Dead Time (TD) = 0.89
SSE: 3.20

Figure 6.6 - Design Tools fit of heat exchanger doublet test data using a FOPDT model

The model fit may appear to be in error because its trace lies below the data during much
of the dynamic portion of the experiment. In reality, however, the model slightly overshoots the
data on the downward trace and then undershoots the data on the upward trace. Thus, the Design
Tools fit of the linear FOPDT model effectively averages the nonlinear behavior of the heat
exchanger.

← Open Loop Data → Closed Loop Data

Graphical Analysis Design Tools Design Tools Fit
Of Step Test Doublet Fit of Set Point Doublet

Process Gain, KP (°C/%) −0.86 −0.90 −0.86

Time Constant, τP (min) 1.0 1.1 1.2
Dead Time, θP (min) 0.3 0.9 1.0
Sum of Squared Errors (SSE) 44.1 3.2 5.4
ITAE Controller Gain, KC (%/°C) −2.1 −0.7 −0.7

Table 6.7 – Comparing FOPDT models for heat exchanger

57
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 In Table 6.7, the two columns under the “Open Loop Data” label show the model fit from this
doublet test along side the FOPDT fit results from the step test graphical analysis. As shown, the
step test analysis yields (what we will learn is) essentially the same values for KP and τP when
compared to the Design Tools fit of open loop doublet test data. The θP for the two methods, on
the other hand, are quite different.
So which fit is “better?” And how do we determine “better?” One time-tested method of
comparing a model to data is by visual inspection. Figure 6.8 shows the same heat exchanger
doublet test data that was used in Fig. 6.6. Also shown is a FOPDT model trace generated using
the KP, τP and θP from the step test graphical analysis.

FOPDT Model From Step Test Graphical Analysis

Model: First Order Plus Dead Time (FOPDT) File Name: PROCESS.DAT
150
FOPDT
Process Variable

149
148 model
147
146
145
144
measured exit
34
temperature data
Controller Output

32
30
28
26

0 5 10 15
Time

Gain (K) = -0.86, Time Constant (T1) = 1.00, Dead Time (TD) = 0.30
SSE: 44.06

Figure 6.8 - Comparing FOPDT model from step test analysis to doublet test data

It is clear from visual inspection that the Design Tools model of Fig. 6.6 more accurately
describes the data then does the step test model shown in Fig. 6.8. An alternative to visual
inspection is to compare the SSE as defined in Eq. 6.1 for the two fits. As listed in Table 6.7, the
Design Tools doublet fit has an SSE of 3.2, while the step test graphical analysis model has an
SSE of 44.1. The dramatically lower SSE for the Design Tools fit confirms our visual conclusion.
Note that the poor model from the step test was a result of the graphical analysis
methodology, which makes many simplifying assumptions. The step test itself is capable of
producing useful dynamic data, and in general, Design Tools can accurately model data
regardless of whether a step, pulse, doublet or PRBS is used as the testing method (though to be
useful in controller design, the data must have been properly generated).

Õ Õ Õ

58
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
6.6 Controller Design Using Closed Loop Data
It is increasingly common to have operations personnel in a production facility reject the idea of
opening an existing loop so controller design data can be collected. In these situations, you must be
prepared to perform dynamic studies when the controller is in automatic. In theory, closed loop
testing can be problematic because the information contained in the data will reflect the character of
the controller as well as that of the process. In practice, however, this theoretical concern rarely
causes real world problems.
For closed loop studies, dynamic data is generated by stepping, pulsing or otherwise
perturbing the set point. To generate proper data, the controller must be tuned aggressively enough so
that, similar to open loop testing, the changing controller output forces the measured process variable
to move more than ten times the noise band. Also, the data set must begin at steady state.

Closed Loop Test of Heat Exchanger

Process: Heat Exchanger Controller: PID ( P= DA, I= off, D= off )

150
PV/Setpoint

148
146
144 exit temperature
142 under P-Only control
38
set point
Controller Output

36
34
32
30
28
26
24

0 5 10 15 20 25
Time (mins)
Tuning: Bias = 29.2, Gain = -1.00, Sample Time = 1.00

Figure 6.9 - Closed loop test of heat exchanger using a set point doublet

Figure 6.9 shows a closed loop test of the heat exchanger. The P-Only controller gain is KC =
−1 for this test (a wide range of controller gains will produce similar FOPDT models). The set point
doublet used in the experiment generates a measured process variable response that covers a range of
nonlinear operation similar to that evident in the previous open loop step and doublet tests.
As required for a Design Tools analysis, the process is at steady state before the first set point
step is made. While data is being saved to file, the set point is stepped up and then down 4 oC from
the initial design level of operation. After the experiment, the file is read into Design Tools, the data
columns are labeled, and a FOPDT model is fit to the data.
Figure 6.10 shows the FOPDT model fit of the heat exchanger data generated in the closed
loop test of Fig. 6.9. A visual inspection of Fig. 6.10 and a comparison of SSE’s and model
parameters as summarized in Table 6.7 establish that it is certainly possible to obtain an accurate
dynamic model from closed lop data.

59
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 FOPDT Fit of Closed Loop Data
Model: First Order Plus Dead Time (FOPDT) File Name: CLOSED.DAT

150

Process Variable
148
146
144 closed loop data and
142 FOPDT model fit
38
36
Controller Output

34
32
30
28
26
24

0 5 10 15 20 25
Time

Gain (K) = -0.86, Time Constant (T1) = 1.20, Dead Time (TD) = 0.96
SSE: 5.38

Figure 6.10 - Design Tools fit of closed loop heat exchanger test data

6.7 Do Not Model Disturbance Driven Data!

For a controller to take appropriate action in response to controller error, the algorithm must
specifically associate how the controller output affects the measured process variable. The way this
association is best developed is through a FOPDT dynamic model. The appropriateness of initial
tuning from tuning correlations is directly related to the accuracy of the model employed.
For this reason, it is essential that the process data used for modeling studies contains
measured process variable dynamics that have been clearly (and in the best of all worlds exclusively)
forced by changes in the controller output. Extraneous events such as dynamics caused by disturbance
variables will degrade the accuracy of the FOPDT model, and thus the initial tuning of the controller.
To illustrate, consider the extreme case shown in Fig. 6.11. Similar to Fig. 6.9, the heat
exchanger is under P-Only control with KC = −1. Here, however, the set point remains constant and
the dynamic event is forced by changes in the warm oil disturbance flow rate. With data being saved
to file, the disturbance flow, initially at 10 L/min, is stepped up to 12.5 L/min, down to 7.5 L/min,
and back to the initial value of 10 L/min. The experiment data is read into Design Tools and fit with a
FOPDT model. The resulting model and experiment data are shown in Fig. 6.12. Visual inspection of
Fig. 6.12 reveals that the model accurately represents the data.
Table 6.13 compares the results of this disturbance driven FOPDT model with a proper
model, assigned here as the Design Tools fit of open loop doublet data listed in Table 6.7. The fit of
disturbance driven data is quite disturbing. Not only are all of the FOPDT model parameters quite
wrong, but the steady state process gain even has the wrong sign!

60
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Disturbance Drives Dynamic Process Event
Process: Heat Exchanger Controller: PID ( P= DA, I= off, D= off )

Disturbance Controller Output PV/Setpoint

148
147
146
Set point is constant
31
30
29
28
27
disturbance changes
12.5 drive dynamic event
10.0
7.5

0 5 10 15 20 25 30
Time (mins)
Tuning: Bias = 29.2, Gain = -1.00, Sample Time = 1.00

Figure 6.11 – Dynamic process data forced by changes in the disturbance variable

FOPDT Model Fit of Disturbance Driven Data

Model: First Order Plus Dead Time (FOPDT) File Name: disturb.txt
Process Variable

148

147

146
disturbance driven data
and FOPDT model fit
Controller Output

32
31
30
29
28
27

0 5 10 15 20 25 30
Time

Gain (K) = 0.97, Time Constant (T1) = 0.015, Dead Time (TD) = 0.0SSE: 1.07

Figure 6.12 – Disturbance driven process data is well described by the FOPDT model

Although not shown by example, be aware that open loop data can also be corrupted by
disturbance events. Further, it is not required for the event to be solely disturbance driven for
problems to arise. To be conservative, if you suspect that a disturbance significant enough to
influence the measured process variable has occurred during a test, you should repeat the test.

61
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Controller Output Driven Data Disturbance Driven Data
(Proper FOPDT Model) (Nonsense Result)

Process Gain, KP (oC/%) −0.90 0.97

Time Constant, τP (min) 1.1 0.02
Dead Time, θP (min) 0.9 0.0

Table 6.13 – Comparing controller output driven model to disturbance driven model

6.8 FOPDT Fit of Underdamped and Inverse Behaviors

FOPDT models cannot accurately describe the dynamic behavior of some processes even in narrow
operating ranges. Underdamped and inverse processes display dynamic response behavior that is
quite different than the classic step response shape discussed to this point.

Underdamped Process
Figure 6.14 shows a FOPDT fit of an underdamped step response. When a process is underdamped,
the measured process variable oscillates as it responds to a controller output step.
Interestingly, even though the process response looks quite different from the FOPDT model
fit, the popular tuning correlations still yield initial controller tuning parameters that provide
reasonable closed loop performance. This is because, in spite of the mismatch between the
measurement and model, the FOPDT model still describes the direction, how far, how fast and with
how much delay the measured process variable responds to the change in controller output. This, after
all, is the information ultimately required for controller design and tuning.

FOPDT Fit of Underdamped Data

Model: First Order Plus Dead Time (FOPDT) File Name: process.txt
Controller Output Process Variable

58
56
54
52
50 FOPDT model fit

60

55

50

0 50 100 150
Time

Gain (K) = 0.51, Time Constant (T1) = 1.82, Dead Time (TD) = 3.60
SSE: 925.17

Figure 6.14 - FOPDT model fit of an underdamped process

62
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
Inverse Process
Figure 6.15 shows a FOPDT fit of an inverse (also called nonminimum phase) step response. When a
process exhibits inverse behavior, the measured process variable first moves in one direction before it
ultimately responds to steady state in the opposite direction. An example of this behavior can be
found on the heat exchanger, where a step change in the warm oil disturbance flow rate produces an
inverse response in the measured exit temperature.
As shown in Fig. 6.15, the FOPDT model approximates the inverse portion of the measured
process variable as dead time. By doing so, the tuning correlations will cause the controller to
essentially ignore the inverse portion of the response and rely on the final behavior in making control
action decisions. Otherwise, the controller will be “chasing its tail” in trying to compensate for the
temporary deviations. Thus the FOPDT model again provides appropriate information for the popular
tuning correlations.

FOPDT Fit of Inverse Response Data

Model: First Order Plus Dead Time (FOPDT) File Name: process.txt
Controller Output Process Variable

60 FOPDT model fit

55

50

45

60

55

50

0 50 100 150
Time

Gain (K) = 1.01, Time Constant (T1) = 21.1, Dead Time (TD) = 34.3
SSE: 82.74

Figure 6.15 - FOPDT model fit of an inverse process

63
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 7. Advanced Modeling of Dynamic Process Behavior
7.1 Dynamic Models Have an Important Role Beyond Controller Tuning
As discussed toward the end of Chapter 3, no real process has its dynamic behavior exactly described
by a FOPDT (first order plus dead time) model. Yet when forced by a change in the controller output,
a FOPDT model reasonably describes the direction, how far, how fast and with how much delay the
measured process variable will respond. This simple model thus provides the essential information
required for controller tuning.
Dynamic models play an important role in process control beyond tuning. Two uses
considered in this chapter are modeling for offline simulation, and modeling for the construction of
model based controller architectures.
For offline simulation, LOOP-PRO provides for the creation of stand alone process models
using Custom Process. Simulations are useful, for example, to investigate how different controller
algorithms, architectures and tuning options perform on a particular process. The varied behaviors of
processes can also be studied. The more studies that can substitute a simulation for the actual process
are, the greater will be the savings in both time and money.
Offline studies become especially important when operations makes a process “off limits” for
extended experimentation. It should come as no surprise to hear that the value of conclusions drawn
using an offline simulation depend on how well the simulation model describes the true dynamic
behavior of the process.
For model based control, LOOP-PRO offers a number of popular architectures including feed
forward, Smith predictor and multivariable decouplers. These advanced architectures employ a
dynamic model within the controller structure to predict the future behavior of a process. This enables
control actions to be taken in advance if the predicted future does not match a desired behavior. As
might be expected, the controller model must predict the true dynamics of a process with reasonable
accuracy for success with these advanced architectures.
For offline simulation and model based control, a model must describe more than the basic
“how far, how fast and with how much delay” features required for tuning. Improved accuracy is
achieved by fitting the best model form to an appropriate set of dynamic process data. For these uses,
an expanded list of models should be considered.
As summarized in Table 7.1, the two most popular dynamic models other than FOPDT
include second order plus dead time (SOPDT) and second order plus dead time with lead time
(SOPDT w/L). These models contain additional parameters that permit them to describe certain
process behaviors with greater accuracy than can be achieved with the FOPDT form.
In earlier chapters we discussed what constitutes an “appropriate” set of dynamic process
data. In particular, we noted that the manipulated variable must be moved far enough and fast enough
so that the measured process variable displays a clear response that dominates the noise band of the
measurement signal. Also, the data sample rate must be ten times the process time constant or faster,
the process must start at steady state before data collection begins, and the first point in the data file
must equal this steady state value.
In this chapter we explore the three dynamic models of Table 7.1 to better understand how to
select the “best” model form. Like controller tuning, there is not one answer. Ultimately, it is the
designer who decides when a dynamic process model is suitably descriptive for the task at hand.

64
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
7.2 Overdamped Process Model Forms
This book focuses on the control of process properties like temperature, pressure, level, flow, density,
concentration and the like where the process streams are comprised of gases, liquids, slurries and
melts. For this class of control challenges, the process variables very rarely display a natural tendency
to oscillate (unlike, say, a mass and spring process that will oscillate quite nicely when perturbed).
Processes that do not have an inherent (open loop) tendency to oscillate are called overdamped.
A large portion of overdamped processes are also self regulating. Self regulating processes
seek a steady state operating level if all manipulated and disturbance variables are held constant for a
sufficient period of time. LOOP-PRO’s gravity drained tanks, heat exchanger, jacketed reactor and
distillation column are examples of overdamped, self regulating processes. If you perform a step test
on the pumped tank, on the other hand, you will see the behavior of a non-self regulating process.

First Order Plus Dead Time Where:

y(t) = measured process variable signal
dy (t )
Time Domain τP + y (t ) = K P u (t − θ P ) u(t) = controller output signal
dt KP = process gain; units of y(t)/u(t)
Y ( s ) K P e −θ P s τP = process time constant; units of time
Laplace Domain G P (s) = = τL = process lead time; units of time
U (s) τ P s +1
θP = process dead time; units of time
Second Order Plus Dead Time
d 2 y (t ) dy (t )
Time Domain τ P1τ P 2 2
+ (τ P1 + τ P 2 ) + y (t ) = K P u (t − θ P )
dt dt

Y (s) K P e −θ P s
Laplace Domain G P ( s) = =
U ( s ) (τ P1 s + 1)(τ P 2 s + 1)
Second Order Plus Dead Time with Lead Time
d 2 y (t ) dy (t ) ⎡ du (t − θ P ) ⎤
Time Domain τ P1τ P 2 2
+ (τ P1 + τ P 2 ) + y (t ) = K P ⎢u (t − θ P ) + τ L ⎥
dt dt ⎣ dt ⎦

Y ( s ) K P (τ L s + 1) e −θ P s
Laplace Domain G P ( s) = =
U ( s ) (τ P1 s + 1)(τ P 2 s + 1)

Table 7.1 – Practical Overdamped Dynamic Model Forms

Table 7.1 lists the three most popular dynamic models used in process control to describe
overdamped, self regulating dynamic process behavior. Each model is of higher order than the
previous model on the list and this implies additional parameters that must be fit to the data.
An additional adjustable parameter lets a model better track the “bumps and curves”
commonly found in process data. Because a dynamic model must reliably interpolate the data it is
fit to, this added capability seems desirable.

65
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Yet in control applications a model is also often required to extrapolate the data. During
extrapolation beyond the limits of the original data, it is not uncommon for this “bumps and curves”
descriptive capability to become exaggerated. The result can be wholly unrealistic predictions of
process behavior. In this sense, a higher order model is a detriment.
In general, always choose the simplest model form that provides an acceptable fit of your
data. Simple models will likely provide the most reasonable extrapolation beyond the limits of that
data.

7.3 The Response Shape of First and Second Order Models

Table 7.1 shows that the FOPDT model has one time derivative and a single time constant. In
comparison, the SOPDT model has a first and second time derivative and two time constants. The
added derivative and associated time constant of the SOPDT model permits a rather subtle but
surprisingly valuable benefit in describing dynamic process behavior.

Process Response Shapes When Forced by Step Change

First Order Second Order

Process Variable
Process Variable

single arc ‘s’ shaped

response response

Figure 7.2 – Added time constant of second order model enables a gradual “s” shaped response

As shown in Fig. 7.2, the second time derivative and time constant of the SOPDT model
permits the computed process variable, y(t), to respond with a gradual “s” shape when forced by a
step change. The single time derivative and time constant of the FOPDT model restricts the computed
process variable response to a sudden or disjoint arc shape as shown in the figure.

Fit of Gravity Drained Tanks Process Data

5 FOPDT fit
Process Variable

data
SSE =0.71
4
model
3

5 SOPDT fit
Process Variable

SSE =0.02
4
model and data

Figure 7.3 – Second order model permits better description of gravity drained tanks data

66
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 The ability of the SOPDT model to respond in a more gradual fashion means that the model
can describe real process data with greater accuracy. Figure 7.3 shows doublet response data from the
gravity drained tanks. A Design Tools fit of a FOPDT (upper plot) and SOPDT (lower plot) model are
also shown. Not only is the fit better in the lower chart based on visual inspection, but the sum of
squared errors (SSE) is significantly lower (recall as discussed in Section 6.5, the smaller the SSE the
better the model describes the data).

7.4 The Impact of KP, τP and θP on Model Behavior

As the process gain, KP, time constant, τP, and dead time, θP, change, so does model behavior. The
following study of these parameters is presented to help you refine your intuition. LOOP-PRO’s
Custom Process module is used to implement the models.

Steady State Process Gain, KP

As shown in Fig 7.4, process gain is the “how far” variable. When the gain doubles in the center
portion of the plot, the model computes a response that travels twice as far for the same change in
controller output. When gain changes sign in the right portion of the plot, the response reverses
directions but the response shape is unaffected. Though the plot is from a FOPDT model with a time
constant of 10 time units and no dead time, the behaviors and observations hold for the other models
in Table 7.1.

Impact of Process Gain on Dynamic Model Behavior

Process: Custom Process Controller: Manual Mode
70 τP = 10
Process Variable

60 θP = 0
50

40

KP = 1 KP = 2 KP = –1
Controller Output

65

60

55

50

0 50 100 150 200 250 300 350 400

Time (time units)

Figure 7.4 – Larger gain means the process variable responds farther for a step in controller output

Time Constant, τP
Figure 7.5 illustrates that time constant is the “how fast” variable, describing how quickly a process
responds to a change in controller output. The right portion of Fig. 7.5 shows that when the first order
time constant increases from 10 to 25 time units while the FOPDT gain and dead time are held
constant, the response takes two and half times longer to complete.
Relating this time constant observation to a SOPDT model requires Eq. 7.1, which explains
how the two time constants of a second order model can be approximated with a single equivalent
first order time constant. To compute the first order equivalent, add the larger of the two second order
time constants to one half the smaller value, or:

67
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 τFO ≈ τSO,max + 0.5τSO,min (7.1)

where: τFO = equivalent first order time constant

τSO,max = larger of the two second order time constants
τSO,min = smaller of the two second order time constants

Impact of Time Constant on Dynamic Model Behavior

Process: Custom Process Controller: Manual Mode

60
KP = 1
Process Variable

55 θP = 0
slower
50
response

τP = 10 τP = 25
Controller Output

65

60

55

50

0 50 100 150 200 250 300 350 400 450 500

Time (time units)

Figure 7.5 – Larger time constant means the process variable responds slower

To demonstrate, consider a SOPDT model with time constants of 7.5 and 5.0 time units. The first
order approximation is computed as:
τFO ≈ 7.5 + 0.5(5.0) = 10 time units
That is, the behavior of the second order model can be approximated as a first order model with a
single time constant of 10 time units.
This result is demonstrated in Fig 7.7, which shows the response of a FOPDT model with a
single time constant of 10 time units to the left and a SOPDT model with time constants of 7.5 and
5.0 time units to the right. Except for the more gradual “s” shape at the beginning of the second order
response, the two plots are very similar as predicted by Eq. 7.1.
The implication of this result is that doubling the response time of a second order model is
not achieved by doubling the two individual time constants. On the contrary, it requires doubling the
equivalent first order time constant approximation. As suggested by Eq. 7.1, this can be achieved a
number of ways for a model. Two ways the response time of the previous second order process can be
doubled to 20 time units are shown in Table 7.6:

τP1 τP2 τFO

17.5 5.0 τFO ≈ 17.5 + 0.5(5.0) = 20 time units
15.0 10.0 τFO ≈ 15.0 + 0.5(10.0) = 20 time units

Table 7.6 – Two ways to make the response time of a second order model equal 20 time units

68
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Comparing a First Order and Second Order Response
Process: Custom Process Controller: Manual Mode

60 KP = 1

Process Variable
58
56
θP = 0
54
52
50
First Order Second Order
Controller Output

τP = 10
65
τP1 = 7.5 τP2 = 5.0
60

55

50

0 20 40 60 80 100 120 140 160 180 200 220 240 260 280
Time (time units)

Figure 7.7 – How the time constants of a SOPDT model relate to that of the FOPDT model

Dead Time, θP
Dead time is the “with how much delay” variable. A response shape is unaffected by a change in dead
time except that as dead time increases, the time delay before the process begins its first detectable
response to a change in controller output increases. This is true for all of the models in Table 7.1.

Impact of Dead Time on Dynamic Model Behavior

Process: Custom Process Controller: Manual Mode

60 delayed KP = 1
Process Variable

response τP = 10
55

50

65
θP = 0 θP = 25
Controller Output

60

55

50

0 50 100 150 200 250 300 350

Time (time units)

Figure 7.8 – Longer dead time means there is a longer delay before the process responds

Figure 7.8 shows the response of two models with identical gain and time constant when
forced by the controller output. The plot to the left is from a model with zero dead time while that to
the right is from a model with a dead time of 25 time units. Except for the delayed response indicated
in Fig 7.8 due to dead time, the response shapes are identical.

69
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
7.5 The Impact of Lead Element τL on Model Behavior
As indicated in Table 7.1, lead time, τL, is a parameter that weights the rate of change (derivative) of
the controller output signal, u(t). It has units of time, but unlike a process time constant, it can be
either positive or negative. This is possible because lead time does not describe a time frame in which
a process variable evolves (which must proceed forward in time). Rather, it describes the influence
that the controller output has on the measured process variable.
Suppose u(t) is rapidly increasing (has a large positive derivative). If lead time, τL, is positive,
the model says to impart a large positive movement to y(t) on top of that dictated by the process gain
and time constants (which describe the “natural” dynamics of the process). If τL is negative, the model
says to impart a large negative movement to y(t) when u(t) is rapidly increasing.

Influence of Lead Term on Process Response

Process: Custom Process Controller: Manual Mode
70
overshoot inverse KP = 1
Process Variable

60 τP1 = 10
τP2 = 5
50
θP = 0
40

τ L = + 20 τ L = – 20
Controller Output

60

55

50

0 50 100 150 200 250 300 350 400

Time (time units)

Figure 7.9 – Lead time can be positive or negative and has a profound impact on response shape

The influence of lead time is shown in Fig. 7.9 for a SOPDT w/L (second order plus dead
time with lead time) model. When lead time is large and positive as shown to the left in Fig. 7.9, the
additional large positive movement imparted to y(t) causes it to overshoot the steady state before it
returns to its proper final value. When lead time is large and negative as shown to the right in the
figure, the additional large negative movement imparted to y(t) actually causes it to go the wrong
direction (inverse response) before correcting itself and moving to its proper steady state.
In spite of the extra dynamic influence of the lead element, the final steady state is still
established by the value of the process gain, KP. As always, the process time constants influence the
speed of response, but their impact is well masked in the above example.
When the SOPDT w/L model is viewed in the Laplace domain (shown in Table 7.1), note
that the lead term is in the numerator while the time constant terms are in the denominator. Because
of this arrangement, it is possible for the lead term to cancel out a process time constant if they are the
same value.
Y (s) (10 s + 1) e − 2 s
= (7.2)
U ( s ) (5.0 s + 1)(10 s + 1)

70
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
Equation 7.2 shows a SOPDT w/L model where KP = 1, θP = 2, τP1 = 5, τP2 = 10, and τL = 10. Because
terms cancel as shown, Eq. 7.2 will behave exactly like a FOPDT model.

Lead Term Can Cancel a Time Constant

Process: Custom Process Controller: Manual Mode

Process Variable 60 KP = 1
58
56 θP = 2
54
52
50
48 FOPDT SOPDT w/ L
τP = 5 τP1= 5 τP2 = 10 τL = 10
Controller Output

60

55

50

0 20 40 60 80 100 120 140 160 180 200 220

Time (time units)

Figure 7.10 – FOPDT response is identical to SOPDT w/ L when terms cancel

Figure 7.10 demonstrates this. Two responses are shown, one for a true FOPDT model
response with KP = 1, θP = 2 and τP = 5 to the left and one for the SOPDT w/L model discussed above
to the right. Because terms cancel as shown in Eq. 7.2, the two responses are identical.

71
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 8. Integral Action and PI Control
8.1 Form of the PI Controller
Like the P-Only controller, the Proportional-Integral (PI) controller computes an output signal to the final
control element based on tuning parameters and the controller error, e(t). Called the ideal, continuous and
position form, the PI controller is expressed in Eq. 8.1:
KC
u (t ) = u bias + K C e(t ) +
τI ∫ e(t ) dt (8.1)

As before, u(t) is the controller output, ubias is the controller bias, and KC is controller gain (a tuning
parameter). The additional tuning parameter, τI, provides a separate weight to the integral term, has units
of time (and thus is always positive), and is commonly referred to as reset time. Because τI is in the
denominator, smaller values of reset time provide a larger weight to (increase the influence of) the
integral term.
The first two terms to the right of the equal sign in Eq. 8.1 are identical to the P-Only controller
of Eq. 5.1. The integral mode of the PI controller is an additional and separate term added to the
equation that integrates or continually sums the controller error over time.
Though not required to understand the actions of a PI controller, we present the Laplace
transfer function form for completeness. The Laplace form of the PI controller is:

U (s) ⎛ 1 ⎞
GC ( s ) = = K C ⎜⎜1 + ⎟⎟ (8.2)
E ( s) ⎝ τIs⎠

8.2 Function of the Proportional and Integral Terms

As with the P-Only controller, the proportional term of the PI controller, KC e(t), adds or subtracts from
ubias based on how far the measured process variable is from the set point at any instant in time. That is,
the contribution to u(t) from the proportional term is based on the size of e(t) at time t. As e(t) grows or
shrinks, the amount added to ubias grows or shrinks immediately and proportionately. The past history and
current trajectory of the controller error have no influence on the proportional term computation.
Proportional Computation
( , ,
64

62
Process Variable/Setpoint

60

58
e(48) = 60 − 61 = −1
56

54

52

50 e(20) = 60 − 53 = 7
48
0 10 20 30 40 50 60 70 80 90 100
Time (time units)

Figure 8.1 – Proportional calculation is based only on current error at time t

72
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Figure 8.1 illustrates this for a set point response with arrows that mark the size of e(t) used in the
proportional term computation at time t = 20 and at time t = 48. Figure 8.2 shows the identical data cast as
a plot of error itself, created by subtracting the measured process variable from set point at each point in
time. As shown, controller error continually changes size and sign as time passes. Controller error also has
units, which in commercial systems is often percent of span.

Error = Set Point − Measurement

10
e(20) = 7
8
Controller Error

4
e(48) = −1
2

e(t) = 0

-2

0 10 20 30 40 50 60 70 80 90 100
Time

Figure 8.2 – Error, e(t), continually changes size and sign as time passes

In contrast, the integral term of the PI computation considers the history of the error, addressing
how long and how far the measured process variable has been from the set point over time. The integral
term integrates or continually sums up the error history. Thus, even a small error, if it persists, will have a
sum total that grows over time and the integral term contribution added to ubias will similarly grow.

Integral Computation ( , , )
64
area = −28 at t = 56
62
Process Variable/Setpoint

60

58
area = 4
56

54
integral area = 120 at t = 30
52

50
each box has integral area = 2
48
0 10 20 30 40 50 60 70 80 90 100
Time (time units)

Figure 8.3 - Integral of error is the area between the set point and measurement

73
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 As indicated in Fig. 8.3, the result of the continual summing of integration, starting from the
moment the controller is put in automatic, is the computation of the area on the graph between the set
point and measured process variable. Thus, at time t = 30 min, when the measured process variable first
crosses the set point in Fig.8.3, the integral is:
30 min

∫ e(t ) dt = 120 (8.3)

0 min

The contribution this integral ultimately makes to ubias in the PI controller of Eq. 8.1, and thus its
impact on u(t), depends on the values of the tuning parameters KC and τI. Because KC is in the numerator
and τI is in the denominator, larger values of KC and smaller values of τI increase the influence of the
integral term.

Integral of Error

10

8
Controller Error

6
integral area = 120 at t = 30
4

2
area = 4
e(t) = 0

-2 area = −28 at t = 56

0 10 20 30 40 50 60 70 80 90 100
Time

Figure 8.4 - Integral of error continually grows and shrinks as time passes

Integration is continual, growing as long as e(t) is positive and shrinking when the error is
negative. At time t = 56 min, when the measured process variable crosses the set point the second time in
Fig. 8.3 (or when the error changes sign for the second time in Fig. 8.4), the total value of the integral is
(+120 − 28) = 92. When the dynamic event (transient) ends, the total integral is (+120 − 28 + 4) = 96.
Recognize that after the transient is over, the integral term can have a residual value even though
e(t) is constant at zero. In Fig. 8.4, the transient has essentially died out yet the integral of the complete
transient has a final or residual value of 96. As discussed in the next section, the consequence of this is
that integral action enables the PI controller to eliminate offset.

8.3 Advantages and Disadvantages to PI Control

As discussed in Section 5.7, offset occurs in most processes under P-Only control when the set point
and/or disturbance are at a value other than that used in the controller design, or more specifically, used to
determine ubias. The big advantage of a PI controller is that it eliminates offset.

74
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Offset is eliminated with the PI controller of Eq. 8.1 because as long as there is any error (as long
as e(t) is not zero), the integral term will grow or shrink in size, causing the controller output, u(t), to
change. Changes in u(t) will only cease when y(t) equals ysetpoint (e(t) = 0) for a sustained period of time. At
that point, the proportional term of Eq. 8.1 equals zero, and the integral term may have a residual value as
just discussed. This residual value of integration, when added to ubias, in effect creates a new overall bias
value that corresponds to the new level of operation.
The ability to eliminate offset is a tremendous advantage. In fact, PI controllers are the most
widely used of the PID family. There are disadvantages with the algorithm, however, including that:
- two tuning parameters that interact in their influence must be balanced by the designer,
- the integral term increases the oscillatory or rolling behavior of the closed loop system.
Because the two tuning parameters interact with each other, it can be challenging to arrive at “best” tuning
values once the controller is placed in automatic.

8.4 Controller Bias From Bumpless Transfer

As just discussed, if e(t) is not zero, the integral term will cause the controller output, u(t), to change.
Changes in u(t) will only cease when e(t) equals zero for a sustained period of time. At that point, the
integral term can have a residual value that, when added to ubias, in effect creates a new overall bias value
corresponding to the current level of operation.
Although integral action enables the controller to adjust to changes in operating level, the bias
should still be properly initialized when the controller is put in automatic. Achieving this is surprisingly
simple and requires no work on your part. When put in automatic with integral action enabled, most
commercial controllers (including LOOP-PRO) initialize the bias to the current value of the controller
output. This “bumpless transfer” feature means that when the control loop is closed, the bias is correct for
the current level of operation. From that point forward, integral action enables the controller to track any
changes in operating level.

8.5 Controller Tuning From Correlations

Designing any controller from the family of PID algorithms entails the following steps:
- specifying the design level of operation,
- collecting dynamic process data as near as practical to this design level,
- fitting a FOPDT model to the process data, and
- using the resulting model parameters in a correlation to obtain initial controller tuning values.
As always, final tuning is performed by trial and error once the controller is online so loop performance
matches that desired by the control designer.
The tuning correlations available in Design Tools for the PI controller are the Internal Model
Control (IMC) relations. These are an extension of the popular lambda tuning correlations and include the
added sophistication of directly accounting for dead time in the tuning computations.
The first step in using the IMC (lambda) tuning correlations is to compute,τC , the closed loop
time constant. All time constants describe the speed or quickness of a response. The closed loop time
constant describes the desired speed or quickness of a controller in responding to a set point change.
Hence, a small τC (a short response time) implies an aggressive or quickly responding controller.
LOOP-PRO's Design Tools automatically computes a τC for moderate tuning, which can be
changed to more aggressive or conservative values using software options.Moderate tuning provides a set
point response that, while reasonably energetic, shows little or no overshoot. The closed loop time
constants are computed using Eq. 8.4:

75
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Aggressive Tuning: τ C is the larger of 0.1 τ P or 0.8 θ P (8.4a)
Moderate Tuning: τC is the larger of 1.0 τ P or 8.0 θ P (8.4b)
Conservative Tuning: τC is the larger of 10 τ P or 80.0 θ P (8.4c)

With τC computed, the PI correlations for IMC (lambda) tuning are

1 τp
KC = τI = τ P (8.5)
KP (θ P + τ C )

Another popular set of tuning correlations are the integral of time-weighted absolute error (ITAE)
tuning correlations. We do not believe they are as dependable as the IMC correlations so they are not
computed by Design Tools. They can easily be computed directly from the FOPDT model parameters,
however. When set point tracking is the control objective, the ITAE tuning correlation is shown in Eq.
8.6:
0.586 τP
KC = (θ P /τ P )-0.916 τI = (8.6)
KP 1.03 - 0.165(θ P /τ P )

When disturbance rejection is the control objective, the ITAE tuning correlation is shown in Eq. 8.7:

0.859 τP
KC = (θ P /τ P )-0.977 τI = (θ P /τ P )0.680 (8.7)
KP 0.674

The different correlations will yield different tuning values and your judgment is required in selecting
which to use. If you are uncertain, remember that it is most conservative to start with the smallest gain and
largest reset time as this will give the least aggressive controller.
Final tuning is performed on-line and by trial and error until desired controller performance is
obtained. If the process is responding sluggishly to disturbances and changes in the set point, the
controller gain is too small and/or the reset time is too large. Conversely, if the process is responding
quickly and is oscillating to a degree that makes you uncomfortable, the controller gain is too large and/or
the reset time is too small.

8.6 Set Point Tracking in Gravity Drained Tanks Using PI Control

The design level of operation for this study is a measured level in the lower tank of 2.4 m while the
pumped flow disturbance is at its expected value of 2.0 L/min. The control objective is to track set point
steps in the range of 2.0 to 2.8 m. The process is currently under P-Only control and operations personnel
will not open the loop for controller design experiments. Hence, closed loop set point steps are used to
generate dynamic process data.
As shown in Fig. 8.5, the P-Only controller being used has a KC = 40 %/m and a bias value of
55.2% (determined as the value of the controller output that, in open loop, causes the measured level
in the lower tank to steady at the design value of 2.4 m when the pumped flow disturbance is at its
expected value of 2.0 L/min). With data being saved to file, the dynamic testing experiment begins.
Specifically, the set point is stepped up to 2.8 m, then down to 2.0 m, and finally back to the design
level of 2.4 m (set point sequences of other sizes and durations would be equally reasonable).

76
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 P-Only Set Point Step Test
Process: Gravity Drained Tank Controller: PID ( P= RA, I= off, D= off )
3.0
2.8 measured tank level

PV/Setpoint
2.6
2.4
2.2
2.0
1.8 set point steps
1.6

70
Controller Output

60
50
40
30

0 5 10 15 20 25
Time (mins)
Tuning: Bias = 55.2, Gain = 40.0, Sample Time = 1.00

Figure 8.5 – Set point step tests on gravity drained tanks under P-Only control

Visual inspection of Fig. 8.5 confirms that the dynamic event is set point driven (as opposed
to disturbance driven). Also, control action appears energetic enough such that the response of the
measured process variable clearly dominates the noise.

FOPDT of Closed Loop Data

Model: First Order Plus Dead Time (FOPDT) File Name: closed.txt
3.0
2.8
Tank Level

2.6
2.4
2.2
2.0 process data
1.8
1.6 and FOPDT fit
Manipulated Variable

70
60
50
40
30

0 5 10 15 20 25
Time

Gain (K) = 0.094, Time Constant (T1) = 1.61, Dead Time (TD) = 0.56
SSE: 0.3837
Gain (K) = 0.094, Time Constant (T1) = 1.61, Dead Time (TD) = 0.56
SSE: 0.3837

Figure 8.6 – FOPDT fit of closed loop dynamic data generated in Fig.8.5

The dynamic data of Fig. 8.5 is read into Design Tools and fit with a FOPDT model. A plot
of the model and closed loop process data is shown in Fig. 8.6. The model appears to be reasonable
and appropriate based on visual inspection, thus providing the design parameters:

77
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Process Gain, KP = 0.094 m/%
Time Constant, τP = 1.6 min
Dead Time, θP = 0.56 min

We first compute the closed loop time constant. Here we choose aggressive tuning, which in the
software is computed as:
τC = larger of 0.1τP or 0.8θP = larger of 0.1(1.6) or 0.8(0.56) = 0.45 min.
Substituting this closed loop time constant and the above FOPDT model parameters into the IMC
tuning correlations of Eq. 8.5 yields the following tuning values:

1 ⎛ 1.6 ⎞
KC = ⎜ ⎟ = 16.9 %/min τI = 1.6 min
0.094 ⎝ 0.56 + 0.45 ⎠
A reverse acting controller is required because KC is positive. Integral with anti-reset windup logic
(which is always desired, as discussed in Section 8.11) is enabled to complete the PI algorithm.
Because integral action is being used, the bias value is not entered but rather is automatically
initialized to the current value of the controller output at the moment the loop is closed.

PI Controller Design From P-Only Set Point Data

Process: Gravity Drained Tank Cont.: PID ( P= RA, I= ARW, D= off, F = off)

2.8
2.6
PV/Setpoint

2.4
2.2
2.0
1.8

65
Controller Output

60
55
50
45

5 10 15 20 25 30
Time (mins)
Tuning: Gain = 16.9, Reset Time = 1.60, Sample Time = 1.00

Figure 8.7 – Performance of PI controller in tracking set point steps

The performance of this controller in tracking set point changes is pictured in Fig. 8.7. The
upper right corner of the plot shows; PID (P= RA, I= ARW, D= off, F=off), confirming that a reverse
acting PI controller with anti-reset windup logic is being used. Although good or best performance is
decided based on the capabilities of the process, the goals of production, the impact on downstream
units and the desires of management, Fig. 8.7 exhibits generally desirable performance. That is, the
process responds quickly, shows modest overshoot, settles quickly, and displays no offset. Compare
this to Fig. 8.5, that shows P-Only performance for the same control challenge.

78
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
8.7 Disturbance Rejection in Heat Exchanger Using PI Control
For this study, a constant measured exit temperature of 147°C is desired. The control objective is to reject
disturbances that occur when the warm oil flow rate, normally about 10 L/min, occasionally spikes as
high as 20 L/min. Using the open loop doublet test shown in Fig.6.6, appropriate FOPDT model
parameters for this design are as follows:
Process Gain, KP = −0.90°C/%
Time Constant, τP = 1.1 min
Dead Time, θP = 0.9 min

The ITAE for disturbance rejection tuning correlations of Eq. 8.7 are demonstrated here, though they
are not our first choice for industrial applications. Substituting the above FOPDT model parameters
into these correlations yield the tuning values:
0.859 1 .1
KC = (0.9 / 1.1)-0.977 = − 1.2 %/°C τI = (0.9 / 1.1)0.680 = 1.4 min
− 0.90 0.674

A direct acting controller is required because KC is negative. Integral with anti-reset windup logic is
enabled to complete the PI algorithm. Because integral action is being used, the bias value is
automatically initialized to the current value of the controller output at the moment the loop is closed.

Disturbance Rejection Under PI Control

Process: Heat Exchanger Cont.: PID ( P= DA, I= ARW, D= off, F = off)
150
PV/Setpoint

148
146
144
142

32
Control Out

28
24
Disturbance

20
15
10

5 10 15 20 25 30 35
Time (mins)
Tuning: Gain = -1.20, Reset Time = 1.40, Sample Time = 1.00

Figure 8.8 - Disturbance rejection performance of heat exchanger under PI control

Figure 8.8 shows the performance of this controller in rejecting step changes in the warm oil
disturbance flow rate. The set point is held constant throughout the experiment at the design operating
level of 147°C. When the disturbance flow rate steps from 10 L/min up to 20 L/min, the PI controller
succeeds in returning the measured process variable to set point, thus eliminating offset.

79
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
8.8 Interaction of PI Tuning Parameters
As mentioned, one disadvantage of the PI controller is that there are two tuning parameters to adjust and
difficulties can arise because these parameters interact with each other. Fig.8.9 shows a tuning map that
illustrates how a typical set point response might vary as the two tuning parameters are changed.
The center of Fig. 8.9 shows a set point step response that is labeled as the base case
performance. It is important to recognize that this base case plot will not be considered by some to be the
"best" performance. What is best must be determined by the operator or engineer for each
implementation. Some require no overshoot while others will tolerate some overshoot in exchange for
a faster set point response. In any event, the grid shows how a set point step response changes as the two
tuning parameters are doubled and halved from a base (here defined as desired) tuning.
The plot in the upper left of the grid shows that when gain is doubled and reset time is halved, the
controller produces large, slowly damping oscillations. Conversely, the plot in the lower right of the grid
shows that when controller gain is halved and reset time is doubled, the response becomes sluggish. This
chart is called a tuning map because, in general, if a controller is behaving poorly, you can match the
performance you observe with the closest picture in Fig. 8.9 and obtain guidance as to the appropriate
tuning adjustments required to move toward your desired performance.

PI Controller Tuning Map

2.0 KC

KC

Base Case
Performance

0.5 KC

0.5 τI τI 2.0 τI

Figure 8.9 – How PI controller tuning parameters impact set point tracking performance

80
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
8.9 Reset Time Versus Reset Rate
Different manufacturer of controllers use different names for the tuning parameters. As discussed in
Chapter 5, some use proportional band rather than controller gain. Also, some use reset rate instead of
reset time, shown in Eq. 8.8 below:
1
Reset Rate = (8.8)
τI

Reset rate has units of 1/time or sometimes repeats/minute. In any case, it is important to know your
manufacturer before you start tuning any controllers!

8.10 Continuous (Position) Versus Discrete (Velocity) Form

As expressed in Eq. 8.1, the continuous form of the PI controller is:

KC
u(t) = ubias + K C e(t ) +
τI ∫ e(t ) dt (8.9)

This is sometimes called the position form because the computed u(t) is a specific value in the range
from 0-100%. When the final control element is a process valve, this controller output in essence is
specifying the actual position between opened and closed that the valve should take.
The first step in deriving the discrete form is to take the time derivative of the continuous
form (the time derivative, or rate of change, of a position is a velocity. Thus, the discrete form of the
PI controller is sometimes called the velocity form). Taking the derivative with respect to time yields
the following:

d u (t ) d u bias d e(t ) K C
= + KC + e (t ) (8.10)
dt dt dt τI

Since ubias is a constant, then d ubias/dt = 0. Assigning finite difference approximations for the continuous
derivatives, Eq. 8.10 becomes the following:

∆u ⎛ e −e ⎞ K
= K C ⎜⎜ i i −1 ⎟⎟ + C ei (8.11)
∆t ⎝ ∆t ⎠ τI

where ei is the current controller error and ei−1 is the controller error from the last sample. Assigning
loop sample time as T = ∆t, then the discrete or velocity PI controller form results in Eq, 8.12, shown
below:

⎛ T ⎞
∆ u = K C ⎜⎜1 + ⎟⎟ ei − K C ei −1 (8.12)
⎝ τI ⎠

This discrete form of the PI controller computes a change in valve position rather than the absolute
position itself. Wherever the valve (final control element) position happens to be, the ∆u instructs the
valve in what direction and by how much to change.
As long as the controller output never reaches the maximum (100%) or minimum (0%) value,
the continuous and discrete forms of the PI controller behave identically (see the next section on reset
windup to learn more).

81
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
8.11 Reset Windup
As long as an error persists, the integral term in Eq. 8.9 will continue to grow. If an error is large
enough and persists long enough, it is mathematically possible for the integral term to grow so large
and its contribution to ubias to become so great that the final control element saturates or reaches a
physical limit of fully open/on or fully closed/off. If this extreme position is still not sufficient to
eliminate the error, the mathematics of Eq. 8.9 permit the integral term to grow yet more.
When the computed u(t) exceeds the physical capabilities of the final control element because
the integral term has reached a huge value, the condition is known as windup. Because windup is
associated with the integral term, it is often referred to as reset windup. Once this condition has
occurred, the controller losses the ability to regulate the process. If and when the error eventually
changes sign and the integral term "unintegrates" or shrinks sufficiently so that the final control
element is no longer saturated, control action can then resume.
The discrete (velocity) form of Eq. 8.12 will not windup because the continuous integral is
eliminated with the introduction of the time derivative in Eq. 8.10. Thus, using the discrete form of
Eq. 8.12 not only eliminates the bias term, but also solves the windup problem.
Unfortunately, the usefulness of Eq. 8.12 in industrial practice is limited because the form
suffers problems when derivative action is included. When taking the derivative of the full PID
algorithm as shown in Eq. 8.13, a derivative of the derivative term yields a second derivative. A
numerical second derivative applied to data that contains even modest noise produces nonsense
results.

d u (t ) d u bias d e(t ) K C d 2 e(t )

= + KC + e (t ) + K C τ D (8.13)
dt dt dt τI dt 2

Hence, most industrial controllers employ the continuous PID algorithm and include jacketing logic
to halt integration when the controller output reaches a maximum or minimum value.
Modern controllers will not suffer from reset windup. It is an old problem that has long been
solved. Beware if you program your own controller, however. Reset windup is a trap that novices fall
into time and again.

82
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 9. Evaluating Controller Performance
9.1 Defining “Good” Controller Performance
Consider that a bioreactor might not be able to tolerate sudden changes in operating conditions
because the fragile living organisms could die. In such a process, it would be desirable to design the
controllers so that the measured process variables respond slowly when counteracting disturbances or
working to track changes in set point.
On the other hand, designers of a concentration control application where the stream then
flows into a mixing tank may seek a very rapid response to set point changes. They may be willing to
tolerate the slowly damping oscillations in the measured concentration that result from such
aggressive control action because the averaging effect of the mixing tank minimizes the ultimate
impact on product quality. In this situation, an aggressive controller that causes significant oscillation
in the measured process variable may be considered acceptable or even good performance.
As these examples illustrate, different applications can have quite different definitions of
“good” closed loop performance. Ultimately, it is the operator or engineer who defines what is good
or best for a particular application. To make this determination, he/she should consider the following:
- goals of production,
- capabilities of the process,
- impact on down stream units, and
- desires of management.

9.2 Popular Performance Criteria

There are popular measures or criteria used in industrial practice to describe closed loop performance.
This permits an orderly comparison of different process response shapes. As shown in Fig. 9.1,
certain features of the closed loop response are assigned the definitions:

A = Size of the set point step

B′ = Size of the first peak above the new set point
B = Size of the first peak above the new steady state
C = Size of the second peak above the new steady state
y (
4.0

3.8

3.6
Process Variable/Setpoint

3.4

3.2

3.0 B C
A
2.8

2.6

2.4

2.2

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Time (mins)

Figure 9.1 - Process response to a set point change with labels indicating response features

83
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
And as shown in Fig. 9.2, the time of occurrence of certain events, such as the time when the
measured process variable first crosses the new set point or reaches its first peak are also used to
describe controller performance.
y (
4.0

3.8

3.6 ± 5% of ∆y(t)
Process Variable/Setpoint

3.4

3.2

3.0
∆y(t)
2.8

2.6

2.4

2.2

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Time (mins)
tsettle
tpeak
trise

Figure 9.2 - Process response to a set point change with labels indicating response features

Another popular criteria is settling time, or the time required for the measured process variable to first
enter and then remain within a band whose width is computed as ± 5% (or 3 % or 10 %) of the total
change in y(t), labeled as ∆y(t) in Fig. 9.2. These popular performance criteria are summarized:
Peak Overshoot Ratio (POR) = B′/A
Decay Ratio = C/B
Rise Time = trise
Peak Time = tpeak
Settling Time = tsettle

Popular values for these criteria include a 10% peak overshoot ratio, a 25% decay ratio and a settling
time band of ± 5%. Also, these criteria are not independent. A process with a large decay ratio will
likely have a long settling time. A process with a long rise time will likely have a long peak time.

Example: Gravity Drained Tanks

As long as a closed loop response looks similar to the classical pattern depicted in Fig. 9.1 and
9.2, then computing the performance criteria is straightforward. Because of offset, the P-Only set
point response of the gravity drained tanks shown in Fig. 9.3 does not possess this classic pattern.
Here, the set point is stepped up from 2.4 m to 4.4 m. The P-Only controller gain is KC = 40
m/% and the controller bias is ubias = 55.2 %. Because of offset, the measured process variable
does not settle at the set point, but rather settles at a final steady state of about 4.0 m.

84
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Process: Gravity Drained Tank Controller: PID ( P= RA, I= off, D= off )

4.8
4.6
4.4

Process Variable/Setpoint
4.2
4.0
3.8
3.6
a b
3.4
∆y(t) c
3.2
3.0
2.8
final steady state
2.6
2.4
2.2
2.0
2 4 6 8 10 12 14 16 18 20
Time (mins)
Tuning: Bias = 55.2, Gain = 40.0, Sample Time = 1.00

Figure 9.3 - Performance of Gravity Drained Tanks under P-Only control

From Fig. 9.3, we read the following information:

- the set point step occurs at about time t = 4.0 min
- the measured process variable first crosses the new set point at about time t = 5.7 min
- the first process variable peak (“a” on plot) occurs at about time t = 6.2 min

- The steady state change in y(t), or ∆y(t), is (4.0 − 2.4 m) = 1.6 m

- The 5% settling time band is (2.4 + 0.95(1.6 m)) = 3.9 m, and (2.4 + 1.05(1.6 m)) = 4.1 m
- the process variable first enters and then remains within the settling time band (“c” on plot)
at about time t = 12.7 min

- Size of set point step is (4.4 − 2.4 m) = 2.0 m

- Size of first peak over new set point is (4.7 − 4.4 m) = 0.3 m
- Size of first peak (“a” on plot) over new steady state is (4.7 − 4.0 m) = 0.7 m
- Size of second peak (“b” on plot) over new steady state is (4.3 − 4.0) = 0.3 m

Using this data, the performance criteria are then computed as such:
- Peak Overshoot Ratio (POR) = (0.3/2.0)⋅100% = 15%
- Decay Ratio = (0.3/0.7)⋅100% = 43%
- Rise Time = (5.7 − 4.0 min) = 1.7 min
- Peak Time = (6.2 − 4.0 min) = 2.2 min
- Settling Time = (12.8 − 4.0 min) = 8.8 min

As discussed at the beginning of this chapter, whether or not these criteria qualify the controller
performance as “good” is subjective and is left to the opinion of the designer.
Õ Õ Õ

85
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 10. Derivative Action, Derivative Filtering and PID Control
10.1 Ideal and Non-interacting Forms of the PID Controller
Like the PI controller, PID computes an output signal to the final control element based on tuning
parameters and the controller error, e(t). As before, u(t) is the controller output and ubias is the controller
bias. Controller gain, KC, and reset time, τI, are tuning parameters. A new tuning parameter, τD, provides
weight to the derivative term, has units of time (so it is always positive), and is called the derivative time.
Larger values of τD provide a larger weighting to (increase the influence of) the derivative term.
Vendors offer different forms of the PID algorithm. We explore here the two most popular forms
and learn that they are identical in capability, but users must use caution because they require slightly
different correlations for tuning. Because there are three tuning parameters, these forms are commonly
called three mode controllers.

Ideal (Non-interacting) PID Form

One form is referred to under a variety of names including the ideal, non-interacting and ISA algorithm.
We will henceforth call this the ideal form. It is expressed in Eq. 10.1:

KC de(t )
u (t ) = u bias + K C e(t ) +
τI ∫ e(t )dt + K C τ D dt
(10.1)

The first three terms to the right of the equal sign in Eq. 10.1 are identical to the PI controller of Eq. 8.1.
The derivative mode of the PID controller is an additional and separate term added to the end of the
equation that considers the derivative, or rate of change, of the error as it varies over time.
We show the Laplace transfer function form because it helps us understand how the names
for the alogrithms evolved. The Laplace form of the ideal, non-interacting PID controller is expressed
in Eq. 10.2:
U (s) ⎛ 1 ⎞
GC ( s ) = = K C ⎜⎜1 + + τ D s ⎟⎟ (10.2)
E (s) ⎝ τI s ⎠

As is evident in Eq. 10.1 and Eq. 10.2, each term sits alone and does not interact with any of the other
terms in the algorithm.

Interacting PID Form

The other PID form is referred to interchangeably as the interacting, series and industrial form. We will
refer to this as the interacting PID form. It is expressed as follows:

⎛ K τ ⎞ K de(t )
u (t ) = u bias + ⎜⎜ K C + C D
⎝ τI
⎟⎟ e(t ) + C
⎠ τI ∫ e(t )dt + K C τ D dt
(10.3)

The Laplace form of the interacting PID algorithm reveals why it is so named. As Eq. 10.4 shows, the
derivative action is introduced as a separate term that is multiplied across the PI terms.

U ( s) ⎛ 1 ⎞
GC ( s ) = = K C ⎜⎜1 + ⎟ (τ D s + 1) (10.4)
E ( s) ⎝ τ I s ⎟⎠

86
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
The impact of this construction seems to fall on the proportional action when expressed in the time
domain of Eq. 10.3 because the proportional term is multiplied by a combination of all three tuning
parameters. As we show later in this chapter, however, all three tuning parameters, KC, τI, and τD,
require slightly different values to make the performance of the ideal and noninteracting forms
equivalent.

10.2 Function of the Derivative Term

As we have discussed, the proportional term considers how far y(t) is from ysetpoint at any instant in time. It
adds or subtracts from ubias based on the size of e(t) only at time t. As e(t) grows or shrinks, the amount
added to ubias grows or shrinks immediately and proportionately.
The integral term addresses how long and how far y(t) has been away from ysetpoint. The integral
term integrates or continually sums up e(t) over time. Thus, even a small error, if it persists, will have a
sum total that grows over time and the integral term contribution added to ubias will similarly grow.
The derivative term considers how fast e(t) is changing at an instant in time. The derivative
computation yields a rate of change or slope of the error curve. An error that is changing rapidly yields a
large derivative regardless of whether a dynamic event has just begun or if it has been underway for some
time. A large derivative (a rapidly changing error) causes the derivative term to have a large impact on the
value added to ubias, and thus the controller output, u(t). It is interesting to note that the derivative
computation does not consider whether e(t) is positive or negative, just whether it is changing rapidly.

10.3 Derivative on Measurement is Used in Practice

Before exploring the contribution of the derivative term to the action of a PID controller, consider that if
ysetpoint is constant, then:
de(t )
=
d y setpo int − y (t )
= −
[
dy (t )
(10.5)
]
dt dt dt

That is, except for a negative sign, Eq. 10.5 shows that the derivative (or slope or rate of change) of the
controller error is the same as the derivative (or slope or rate of change) of the measured process variable,
y(t).

Error Trace After a Set Point Change Measured Variable Trace After a Set (Point Change
, , )
64
10
62
8
Process Variable/Setpoint

de(t)/dt = when set point changes

8

60
6
Controller Error

58
4
56

2 54

e(t) = 0 52

-2 50

48
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
Time Time (time units)

Figure 10.1 – The trace of e(t) and y(t) are reflections except when set point changes

87
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Figure 10.1 provides a visual appreciation of this by showing a set point response curve two
ways. The left plot shows the error trace, e(t), after a set point step. The right plot shows the measured
process variable trace, y(t), for the same event. After the set point step, the changing slope of e(t)
identically tracks the slope of y(t) except that they are reflections of one another (the derivatives have
opposite signs).
The big difference is that the left plot in Fig. 10.1 shows a momentary vertical spike in e(t) at the
instant that ysetpoint changes. The derivative or slope of this spike in error, de(t)/dt, in theory approaches
infinity and in the real world will be at least a very large value. If τD is large enough to provide any
meaningful weight to the derivative term, this very large derivative value will cause a large and sudden
manipulation in u(t). This large manipulation, referred to as derivative kick, is almost always
undesirable.
From the y(t) trace to the right in Fig. 10.1, we see that y(t) itself does not go through this
dramatic vertical change in slope. Though set points can change instantly causing the computed error to
change in kind, physical processes change in a more gradual and continuous fashion. Thus, y(t) will not
display dramatic vertical changes in slope when ysetpoint changes , but will trace a gradual and continuous
dynamic (assuming that loop sample time is T ≤ 0.1τP to adequately track the process behavior).
Because derivative on error behaves identically to derivative on measurement at all times except
at those moments when ysetpoint changes, and when the set point does change, the derivative on error
provides information we don’t want our controller to use, we substitute Eq. 10.5 into Eq. 10.1 and Eq.
10.3 to obtain the PID with derivative on measurement controller.
Ideal (Non-interacting) PID with Derivative on Measurement:
KC dy (t )
u (t ) = u bias + K C e(t ) +
τI ∫ e(t )dt − K C τ D dt
(10.6)

Interacting PID with Derivative on Measurement:

⎛ K τ ⎞ K dy (t )
u (t ) = u bias + ⎜⎜ K C + C D
⎝ τI
⎟⎟ e(t ) + C
⎠ τI ∫ e(t )dt − K C τ D dt
(10.7)

Because the performance of Eq. 10.6 and Eq.10.7 is the same as Eq. 10.1 and Eq. 10.2 respectively
except they do not “kick” when the set point changes, PID with derivative on measurement is the
preferred algorithm in practical applications.

10.4 Understanding Derivative Action

The derivative dy(t)/dt, as illustrated in Fig. 10.2, is computed as the slope of y(t). During a set point
response transient, the slope is large and positive when the measured process variable trace is increasing.
When y(t) is decreasing, the derivative (slope) is negative. And when the measurement goes through a
peak or a trough, there is a moment in time when the derivative is zero.
For discussion purposes, assume that KC is positive and that τD (always positive) is large enough
to provide meaningful weight to the derivative term (after all, if τD is zero, the derivative term has no
influence regardless of the slope of the measured process variable). The negative sign in front of the
derivative term in Eq. 10.6 and Eq. 10.7 means that the control action contribution from the derivative
will be opposite to the sign of the slope. Thus, when dy(t)/dt is large and positive, the derivative term has a
large influence and seeks to decrease the amount added to ubias. Conversely, when dy(t)/dt is negative, the
derivative term seeks to increase the amount added to ubias.

88
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Slope of y(t) Changes During a Transient
( , , )
64
dy(t)/dt is zero dy(t)/dt is negative
62

Process Variable/Setpoint
60

58

56

54 dy(t)/dt is large and positive

52

50

48
0 10 20 30 40 50 60 70 80 90 100
Time (time units)

Figure 10.2 - A set point transient produces y(t) slopes that are positive, negative or zero

It is interesting to note that the derivative term does not consider whether the measurement is
heading toward or away from the set point (whether e(t) is positive or negative). The only consideration is
whether the measurement is heading up or down and how quickly. At the peak of the transient in Fig.
10.2, dy(t)/dt = 0 and the derivative term momentarily makes no contribution to control action. The
proportional and integral terms definitely influence u(t) at that point in time, however.

10.5 Advantages and Disadvantages of PID Control

As we have discussed in previous chapters, offset occurs in most processes under P-Only control when
the set point and/or disturbances are at values other than that used in the controller design, or more
specifically, used to determine ubias. The benefit of the PI controller is that integral action works to
eliminate offset because as long as there is any error, the integral term will grow or shrink in influence and
cause u(t) to change. The integration ceases only when y(t) equals ysetpoint for a sustained period of time. A
disadvantage of the integral term is that it increases the oscillatory or rolling behavior of y(t). Also, the
two tuning parameters of the PI controller interact in their influence and it is sometimes difficult to
determine what action to take if controller performance is not as desired.
The big benefit of the derivative term is that it works to decrease oscillations in y(t) because it has
its largest influence when y(t) is changing rapidly. Thus, the three terms of a properly tuned PID controller
can work together to provide rapid response to error (proportional term), eliminate offset (integral term),
and minimize oscillations in the measured process variable (derivative term).
Unfortunately, there are significant disadvantages to including derivative action in your
controller. We just mentioned the challenge of balancing two interacting tuning parameters for PI control.
A three mode PID algorithm has three tuning parameters that interact and must be balanced to achieve
desired controller performance. Indeed, tuning a PID controller can be quite challenging as it is often not
at all obvious which of the three parameters is most responsible for an undesirable performance.
A second disadvantage relates to the uncertainty in the derivative computation for processes that
have noise in the measured process variable. This problem and a solution using a derivative filter is
explored in more detail later in Section 10.11.

10.6 Three Mode PID Tuning From Correlations

Designing any controller from the family of PID algorithms entails the following steps:

89
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 - specifying the design level of operation,
- collecting dynamic process data as near as practical to this design level,
- fitting a FOPDT model to the process data, and
- using the resulting model parameters in a correlation to obtain initial controller tuning values.
As always, final tuning is performed by trial and error once the controller is online so loop performance
matches that desired by the control designer.
The tuning correlations available in Design Tools for the PID controller are the Internal Model
Control (IMC) relations. These are an extension of the popular lambda tuning correlations and include the
added sophistication of directly accounting for dead time in the tuning computations.
The first step in using the IMC (lambda) tuning correlations is to compute,τC , the closed loop
time constant. All time constants describe the speed or quickness of a response. The closed loop time
constant describes the desired speed or quickness of a controller in responding to a set point change.
Hence, a small τC (a short response time) implies an aggressive or quickly responding controller.
LOOP-PRO's Design Tools automatically computes a τC for moderate tuning, which can be
changed to more aggressive or conservative values using software options.Moderate tuning provides a set
point response that, while reasonably energetic, shows little or no overshoot. The closed loop time
constants are computed using Eq. 10.8:

Aggressive Tuning: τ C is the larger of 0.1 τ P or 0.8 θ P (10.8a)

Moderate Tuning: τ C is the larger of 1.0 τ P or 8.0 θ P (10.8b)
Conservative Tuning: τC is the larger of 10 τ P or 80 θ P (10.8c)
With τC computed, three mode PID correlations for IMC (lambda) tuning are as follows:

KC τI τD

1 ⎛ τ P + 0.5 θ P ⎞ τP θP
PID Ideal ⎜ ⎟ τ P + 0 .5 θ P (10.9)
KP ⎜ τ + 0.5 θ ⎟ 2τ P + θ P
⎝ C P ⎠

1 ⎛ τP ⎞
PID Interacting ⎜ ⎟ τP 0.5 θ P (10.10)
KP ⎜ τ + 0.5 θ ⎟
⎝ C P ⎠

As mentioned earlier, all three tuning parameters are different for the ideal (non-interacting) and
interacting PID forms. Thus, it is essential that you know the PID form used by your manufacturer when
computing tuning values for a real implementation.
Like with all tuning correlations, your judgment is required for final tuning. Remember that the
most conservative (least aggressive) controller uses a smaller gain, larger reset time and smaller derivative
time.

10.7 Converting From Interacting PID to Ideal PID

By setting the individual terms of the ideal PID form of Eq. 10.1 (or Eq. 10.6) to those of the interacting
PID form of Eq. 10.3 (or Eq. 10.7) and solving for the ideal tuning parameters yields conversion relations
between the two forms as such:

90
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 ⎛ τ D ,Interact ⎞
K C ,Ideal = K C ,Interact ⎜1 + ⎟ (10.11a)
⎜ τ I ,Interact ⎟
⎝ ⎠

⎛ τ D ,Interact ⎞
τ I ,Ideal = τ I ,Interact ⎜1 + ⎟ (10.11b)
⎜ τ I ,Interact ⎟
⎝ ⎠

τ D ,Interact
τ D ,Ideal = (10.11c)
⎛ τ ⎞
⎜1 + D ,Interact ⎟
⎜ τ I ,Interact ⎟
⎝ ⎠

These conversion relations indicate that the ideal and non-interacting PID forms will provide
identical performance if the values of the individual tuning parameters are properly specified. And thus it
is not surprising that by applying Eq. 10.11 to the interacting PID tuning correlations of Eq. 10.10 indeed
yields the ideal PID correlations of Eq. 10.9.

10.8 Exploring Set Point Tracking Using PID Control

Building on the PI control set point tracking study for the gravity drained tanks presented in Chapter 8, we
recall that the design level of operation is a measured level in the lower tank of 2.4 m while the pumped
flow disturbance is at its expected value of 2.0 L/min. The control objective is to track set point steps in
the range of 2.0 to 2.8 m. As detailed in Section 8.6, FOPDT dynamic model parameters fit to controller
output driven data that was collected at the design level of operation are as follows:
Process Gain, KP = 0.094 m/%
Time Constant, τP = 1.6 min
Dead Time, θP = 0.56 min

We first compute the closed loop time constant. Here we choose aggressive tuning:

τC = larger of 0.1τP or 0.8θP = larger of 0.1(1.6) or 0.8(0.56) = 0.45 min.

Substituting this closed loop time constant and the above FOPDT model parameters into the
correlations of Eq. 10.9 and Eq. 10.10 yields the following tuning values:
PID Ideal KC = 27.5 %/m τI = 1.9 min τD = 0.24 min
PID Interacting KC = 23.5 %/m τI = 1.6 min τD = 0.28 min

The performance of these PID controllers are compared in Fig. 10.3 to an IMC tuned PI controller as
presented in Chapter 8. Those tuning values were as follows:
PI control KC = 16.9 %/m τI = 1.6 min
The result of the set point tracking comparison is shown in Fig. 10.3 (note that the "Advanced" box
must be checked in LOOP-PRO's controller design menu to access the interacting PID form).

91
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 PI vs PID for Set Point Tracking Using IMC Tuning
Process: Gravity Drained Tank
Cont.: PID ( P= RA, I= ARW, D= Interacting (meas), F = off)

2.8
2.6

PV/Setpoint
2.4
2.2
2.0
1.8
PI ideal PID interacting PID
80
Controller Output

70
60
50
40
30

0 10 20 30 40 50 60 70 80
Time (mins)
Tuning: Gain = 23.5, Reset Time = 1.60, Deriv Time = 0.28, Sample Time = 1.00

Figure 10.3 – Comparison of PI and PID controller performance in tracking set point steps

The most obvious difference between PI and PID control is that derivative action causes the
noise in the measured process variable to be amplified and reflected in the controller output signal.
Such extreme control action will wear a mechanical final control element, requiring continued and
expensive maintenance.
Closer scrutiny of Fig. 10.3 reveals that derivative action impacts the oscillatory behavior of
the measured process variable. Specifically, the PID controller not only achieves a faster rise time
(because KC is bigger), but also a smaller peak overshoot ratio and faster settling time because of
derivative action.
Finally, Fig. 10.3 establishes that the ideal PID and interacting PID algorithms can produce
identical control performance. Recall that each form was tuned in this example using its own IMC
tuning correlations of Eq.10.9 and Eq. 10.10.

10.9 Derivative Action Dampens Oscillations

In Fig. 10.4 we further explore this observation that derivative action dampens oscillations in the
measured process variable. To facilitate this study, measurement noise is set to zero to eliminate its
influence on the control signal. This lets us isolate derivative time, τD, as a tuning parameter and study its
impact on set point tracking performance.
The middle plot in Fig. 10.4 shows the set point tracking performance of a PID controller tuned
using IMC tuning and the standard τC value of Eq. 10.8. For all three plots, KC and τI remain constant.
The plot to the left shows how the oscillating nature of the response increases as derivative action is
decreased by half. The plot to the right shows that when derivative action is too large, it inhibits rapid
movement in the measure process variable, causing the rise time and settling time to lengthen.

92
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 y
Impact of Derivative Action on Noise Free PID Set Point Response
( , , y ( , , y ( , ,

0.5 τD τD 2 τD

Figure 10.4 - How derivative time,τD, impacts PID set point tracking performance

10.10 Measurement Noise Hurts Derivative Action

We noted in Fig. 10.3 that derivative action causes measurement noise to be amplified and reflected
in the controller output signal. The reason for this extreme control action, as illustrated in Fig. 10.5, is
that a noisy signal produces conflicting derivatives as the slope appears to dramatically alternate direction
at every sample.

Noise in PV Impacts the Derivative Computation

Process: Gravity Drained Tank Controller: PID ( P= RA, I= ARW, D= meas )

set point step

3.9
slope rapidly
Process Variable/Setpoint

3.8 decreasing

3.7

3.6
slope rapidly
increasing
3.5
y(t)

43 44 45
Time (mins)

Figure 10.5 – Measured process variable noise causes uncertainty in the derivative computation

As noted in Fig. 10.5, the derivative can alternate between a large increasing slope followed by a large
decreasing slope, sample after sample. The derivative mode thus reflects this noise as it computes a series
of alternating and compensating controller actions. The degree to which the noise is amplified in the
controller output depends on the size of τD.

93
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Figure 10.6 illustrates an additional problem that measurement noise can cause with derivative
action when the level of operation is near a controller output constraint (either the maximum value, umax,
or the minimum value, umin). The PID tuning values in Fig. 10.6 are constant throughout the experiment.
As indicated on the plot, measurement noise is increased from small to medium to large in three set point
tracking tests. (The terms “small” and “large” are used to indicate relative levels of noise. Please do not
use this plot as a standard for determining if noise in your process is small or large.)
As long as measurement noise causes the derivative to alternate equally between suddenly
increasing and suddenly decreasing, and the controller output can reflect this “equality in randomness”
unimpeded, then controller performance is reasonably consistent in spite of significant noise. A
comparison of the small noise and medium noise cases shows the controller output is not hitting a
constraint and set point performance is consistent.
If a constraint inhibits the controller output from the “equality in randomness” symmetry, causing
it to become skewed or off center, then controller performance degrades. This is demonstrated in Fig. 10.6
in the right most set point steps for the case where measurement noise is large. For this case, the controller
output signal becomes so active that it hits the maximum controller output value, umax. By constraining
u(t), the controller output loses its symmetry and controller performance is affected as the measured
process variable temporarily deviates from set point.

Derivative-Noise Interaction Impacts Performance

Process: Gravity Drained Tank Controller: PID ( P= RA, I= ARW, D= meas )

4.5 small noise medium noise large noise

4.4
PV/Setpoint

4.3
4.2
controller output
4.1 hits umax constraint
4.0
3.9

90
Controller Output

80
70
60
50

0 10 20 30 40 50 60
Time (mins)
Tuning: Gain = 15.0, Reset Time = 2.50, Deriv Time = 0.40, Sample Time = 1.00

Figure 10.6 – When controller output becomes constrained, controller performance degrades

10.11 PID With Derivative Filter Reduces the Impact of Noise

To improve performance when there is noise or random error in the measured process variable, the
PID control algorithm can be modified with the addition of a derivative filter. The filter works to limit
the large controller output moves that derivative action causes as a result of measurement noise. Even
if noise is not causing performance problems, the derivative filter can help reduce the constant
controller output fluctuations that can lead to wear in the final control element.
As with three mode PID, vendors offer different forms of a four mode PID with derivative
filter controller. We explore the two most popular forms here. While not shown, the controllers are
typically implemented in a "derivative on measurement" form analogous to Eq. 10.6 and Eq. 10.7.

94
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
Ideal (Non-interacting) PID with Filter Form
The ideal or non-interacting PID with derivative filter algorithm is expressed as:
KC de(t ) du (t )
u (t ) = u bias + K C e(t ) +
τI ∫ e(t )dt + K Cτ D
dt
− ατ D
dt
(10.12)

The first four terms on the right side of Eq. 10.12 are identical to the PID controller of Eq. 10.1. The last
term on the right side of the equation is the filter term that subtracts a derivative or rate of change in the
controller output, u(t), from the four PID terms.
When the first four PID terms compute large and sudden changes in u(t) as might arise from
noise in the process variable (see Fig. 10.5), then the rate of change (derivative) of the controller
output will also become large. This controller output derivative is scaled by ατD and subtracted from
the computation as shown in Eq. 10.12. If the derivative of u(t) is large, then in effect, the final
computed controller output change is moderated (or filtered). The actual change in u(t) sent to the
final control element is smaller than it otherwise would be with the traditional three mode PID form.
The size of the filter constant, α, dictates how much of each change is filtered out of the final u(t)
signal.
The Laplace form of the ideal, non-interacting PID with derivative filter controller is shown
in Eq. 10.13:
U ( s) ⎛ 1 ⎞⎛ 1 ⎞
GC ( s ) = = K C ⎜⎜1 + + τ D s ⎟⎟⎜⎜ ⎟⎟ (10.13)
E (s) ⎝ τI s ⎠⎝ ατ D s + 1 ⎠

This form shows that the filter term is multiplied across, and thus affects the computations of, all
terms of the three mode PID algorithm.

Interacting PID with Filter Form

The interacting or series form of the four mode PID with derivative filter form is expressed as
follows:
⎛ K τ ⎞ K de(t ) du (t )
u (t ) = u bias + ⎜⎜ K C + C D ⎟⎟ e(t ) + C ∫ e(t )dt + K Cτ D − ατ D (10.14)
⎝ τI ⎠ τI dt dt

The same filtering computation just discussed for the ideal form applies here. The Laplace form of the
interacting PID with derivative filter algorithm is expressed as follows:

U (s) ⎛ 1 ⎞⎛ τ D s + 1 ⎞
GC ( s ) = = K C ⎜⎜1 + ⎟⎜ ⎟ (10.15)
E (s) ⎝ τ I s ⎟⎠⎜⎝ ατ D s + 1 ⎟⎠

The interacting form is also commonly referred to as the “lead/lag” form.

10.12 Four Mode PID Tuning From Correlations

The derivative filter constant, α, becomes a fourth tuning parameter, creating a yet more challenging
tuning problem. The tuning correlations available in Design Tools are the IMC (lambda) relations that
use the closed loop time constant,τC.. Again, as computed in Eq. 10.8, LOOP-PRO offers both "standard"
and "conservative" τC values. With τC computed, four mode PID correlations for IMC (lambda) tuning are
as follows:

95
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 KC τI τD α

1 ⎛ τ P + 0 .5 θ P ⎞ τP θP τ C (τ P + 0.5 θ P )
PID Ideal w/filter ⎜ ⎟ τ P + 0.5 θ P (10.16)
KP ⎜ τ +θ ⎟ 2τ P + θ P τ P (τ C + θ P )
⎝ C P ⎠

1 ⎛ τP ⎞ τC
PID Interacting w/filter ⎜⎜ ⎟⎟ τP 0.5 θ P (10.17)
KP ⎝ τC + θ P ⎠ τC + θ P

All four tuning parameters are different for the ideal (non-interacting) and interacting PID forms.
As stated before, know the PID form used by your manufacturer when computing tuning values for a real
implementation. And again, as with all tuning correlations, your judgment is required for final tuning.

10.13 Converting From Interacting PID with Filter to Ideal PID with Filter
Setting the individual terms of the ideal four mode PID form of Eq. 10.12 to those of the interacting form
of Eq. 10.14 and solving for the ideal tuning parameters yields conversion relations between the two
forms, shown in Eq. 10.18:
⎛ τ D ,Interact ⎞
K C ,Ideal = K C ,Interact ⎜1 + ⎟ (10.18a)
⎜ τ ⎟
⎝ I , Interact ⎠

⎛ τ D ,Interact ⎞
τ I ,Ideal = τ I ,Interact ⎜1 + ⎟ (10.18b)
⎜ τ I ,Interact ⎟
⎝ ⎠

τ D ,Interact
τ D ,Ideal = (10.18c)
⎛ τ ⎞
⎜1 + D ,Interact ⎟
⎜ τ I ,Interact ⎟
⎝ ⎠

⎛ τ D ,Interact ⎞
α Ideal = α Interact ⎜1 + ⎟ (10.18d)
⎜ τ I ,Interact ⎟
⎝ ⎠

Based on these conversion relations, we conclude as we did with the three mode forms that the four mode
ideal and non-interacting PID controllers will provide identical performance if the values of the individual
tuning parameters are properly specified. And applying Eq. 10.18 to the interacting PID tuning
correlations of Eq. 10.17 indeed yields the ideal PID correlations of Eq. 10.16.

10.14 Exploring Set Point Tracking Using PID with Derivative Filter Control
We repeat the set point tracking study for the gravity drained tanks presented in section 10.8. Again,

Process Gain, KP = 0.094 m/%

Time Constant, τP = 1.6 min
Dead Time, θP = 0.56 min

96
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
and the closed loop time constant using aggressive tuning:

τC = larger of 0.1τP or 0.8θP = 0.45 min.

Substituting this closed loop time constant and the above FOPDT model parameters into the
correlations of Eq. 10.16 and Eq. 10.17 yields the following tuning values:

PID Ideal KC = 19.8 %/m τI = 1.9 min τD = 0.24 min α = 0.52

PID Interacting KC = 16.9 %/m τI = 1.6 min τD = 0.28 min α = 0.44

The performance of these PID controllers are compared in Fig. 10.3 to an IMC tuned PI controller as
presented in Chapter 8. Those tuning values were as follows:

PI control KC = 16.9 %/m τI = 1.6 min

The result of the set point tracking comparison is shown in Fig. 10.7 (note that the
"Advanced" box must be checked in LOOP-PRO's controller design menu to access the interacting
PID form and the derivative filter option).
Comparing the filtered PID of Fig. 10.7 to the unfiltered results of Fig. 10.3 reveals the
dramatic capability of the filter to temper the actions of the controller output signal. The ability of
derivative action to achieve a faster rise time, smaller peak overshoot ratio and faster settling time is
similar to that shown in Fig 10.3, implying that the filter does not necessarily impact performance.

PI vs PID with Filter for Set Point Tracking Using IMC Tuning
Process: Gravity Drained Tank
Cont.: PID ( P= RA, I= ARW, D= Interacting (meas), F= on)

2.8
2.6
PV/Setpoint

2.4
2.2
2.0
1.8

PI ideal PID interacting PID

80
with filter with filter
Controller Output

70
60
50
40
30

0 10 2040 30 50 60 70
Time (mins)
uning: Gain = 16.9, Reset Time = 1.60, Deriv Time = 0.28, Filter Con = 0.44, Sample Time =

Figure 10.7 – Comparing PI and PID with derivative filter in tracking set point steps

97
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 On the other hand, because the derivative filter works to limit sudden changes in the
controller output signal, a large filter can reduce the benefits of derivative action. Ideally, the filter
constant should be just large enough to contain the erratic fluctuations in the controller output signal,
but not so much as to degrade the overall performance of the controller.
Finally, Fig. 10.7 reinforces the observation of Fig. 10.3 that the ideal PID with filter and
interacting PID with filter algorithms can produce identical control performance. Recall that each
form was tuned in this example using its own IMC tuning correlations of Eq.10.16 and Eq. 10.17.

98
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Practical Theory

11. First Principles Modeling of Process Dynamics

11.1 Empirical and Theoretical Dynamic Models
Dynamic models describe how the behavior of a process changes with time. Empirical dynamic
models are essentially curve fits of observed process behavior, and as such, they can be developed
relatively quickly. To create an empirical model, a general dynamic model form is fit to process data
collected during an experimental study. We did this frequently in previous chapters when we fit the
first order plus dead time (FOPDT) model to data to compute controller tuning values.
Empirical models can describe dynamic process data quite accurately. The disadvantage of
empirical modeling is that because it is a data fit and not a process description, extrapolation beyond
the bounds of the data used to fit the model is speculative. The mismatch between model predictions
and actual process behavior will likely increase as the extrapolation increases.
Theoretical dynamic models are derived from first principles. Because they are based on
theory, they often can be extrapolated beyond the bounds of available data with some confidence.
This makes them useful for exploring a wide range of operating policies and control strategies.
Studies using theoretical models can be performed quickly, safely and inexpensively. They can also
form the basis of a training simulator. For example, LOOP-PRO’s Case Studies simulations are based
on theoretical models. The disadvantage of theoretical models is that they can be substantially more
challenging and time consuming to develop compared to empirical models.

11.2 Conserved Variables and Conservation Equations

First-principles (theoretical) dynamic models result from conservation equations. Conserved variables
include:
- mass
- mass of component i (a species balance)
- energy
- momentum

Balances are created by defining a boundary around the process and then computing the following:
Accumulation = In − Out + Generation − Consumption

Note that level, temperature and process variables other than those listed above are not conserved.
As developed below, a model describing the dynamic (or time-dependent) behavior of liquid
level in a tank begins with a mass balance. The specific assumptions of constant liquid density and
constant tank cross-sectional area enable the mass balance to yield a dynamic equation describing
liquid level. Similarly, a later example shows that a dynamic temperature model begins with an
energy balance.
When deriving a process model, regardless of the final application, good engineering practice
dictates that we include the following in the derivation:
- a picture of the process with appropriate labels,
- units used in the model,
- assumptions used in the derivation,
- step-by-step details of the derivation,
- a final dynamic model differential equation including boundary conditions.

These organizational details help prevent mistakes while also documenting the intended use and
application of your work. Also, others reviewing your work will benefit from the documentation, and
you will also if you revisit your derivation at a later date.

99
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
11.3 Mass Balance on a Draining Tank
The process pictured in Fig. 11.1 is a tank that has liquid flowing in the top and freely draining out of
a hole in the bottom. As the variable labels indicate, the flow rate in and out can change with time.
The cross section of the tank is constant with height, so the volume of the tank at any time t can be
related to the height of the tank.
Following good engineering practice, we begin with a picture, units and assumptions used,
followed by step-by-step details of the derivation.

Picture:
F0(t), ρ0(t)

h(t) V(t), ρ(t)

F1(t), ρ1(t)
AC

Figure 11.1 - Picture of a draining tank with process variables clearly labeled

Variables with units:

- Liquid flow rate F(t) [=] m3/s
- Liquid density ρ(t) [=] Kg/m3
- Liquid level height h(t) [=] m
- Area of tank cross section AC [=] m2
- Liquid volume V(t) [=] m3
- Time t [=] s

Assumptions:
- the process model is restricted to the liquid volumes and flows
- liquid can enter or leave the tank only through the flow streams shown (i.e. no evaporation)
- tank cross sectional area is constant with height, or V(t) = AC h(t)
- liquids are incompressible and thus density is constant, or ρ0(t) =ρ1(t) =ρ(t) =ρ

Step by step details:

Perform a mass balance on a draining tank experiment that takes time ∆t to complete, as follows:
- Mass in by flow: ρ 0 (t ) F0 (t ) ∆t
- Mass out by flow: ρ1 (t ) F1 (t ) ∆t
- Mass generation: 0
- Mass consumption: 0
- Mass accumulation in tank: ρ (t )V (t ) t + ∆t − ρ (t )V (t ) t

100
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
The bars over the density and flow rate variables, ρ i (t ) and Fi (t ) , indicate that these values are
averaged over finite time ∆t .

Sum the mass balance: ρ (t )V (t ) t +∆t − ρ (t )V (t ) t = ρ 0 (t ) F0 (t ) ∆t − ρ1 (t ) F1 (t ) ∆t

ρ (t )V (t ) t +∆t − ρ (t )V (t ) t
Divide by ∆t : = ρ 0 (t ) F0 (t ) − ρ1 (t ) F1 (t )
∆t

Take the limit as ∆t → 0 (as ∆t approaches zero) and recognize that the left side becomes the
definition of a derivative. We further recognize that the average process variables values ρ i (t ) and
Fi (t ) then become point values ρ i (t ) and Fi (t ) , or:

dρ (t )V (t )
= ρ 0 (t ) F0 (t ) − ρ1 (t ) F1 (t )
dt

We employ the assumption that liquids are incompressible and divide by density across the equation
to obtain:
dV (t )
= F0 (t ) − F1 (t )
dt

Employing the assumption that the cross-sectional area of the tank is constant with respect to liquid
level (liquid height), or V(t) = AC h(t), yields:
dh (t )
AC = F0 (t ) − F1 (t )
dt

An ordinary differential equation (ODE) is incomplete without boundary conditions. For dynamic
models used in process control, the most useful conditions are initial conditions, or those that define
the condition of the process at time t = 0. Suppose we know that the initial height of the liquid is
steady at hS, or:
h(0) = hS

Then the dynamic process model with initial conditions becomes

dh (t )
AC = F0 (t ) − F1 (t ) where h(0) = hS (11.1)
dt

Case 1: Assume Drain Flow Proportional to Hydrostatic Head

If we assume the drain flow out of the bottom of the tank is proportional to the height of the liquid in
the tank (the hydrostatic head), then we can write

F1 (t ) = α1 h(t ) with units α 1 [=] m2/s

101
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
Substituting this flow rate into the dynamic process model yields

dh(t )
AC + α1h(t ) = F0 (t )
dt
Divide by α 1 to obtain the following:

AC dh(t ) 1
+ h(t ) = F0 (t ) where h(0) = hS (11.2)
α1 dt α1

Compare the above to the general first order process model:

dy (t )
τP + y (t ) = K P u (t )
dt
(We assume no dead time for this model and will explore dead time as a separate issue when we
introduce Laplace transforms in Chapter 14.)

We see our payoff for this modeling exercise. Specifically, we now know the steady state process
gain, K P , and the process time constant, τ P , of the draining tank model:

1 AC
KP = [=] s/m2 and τP = [=] s
α1 α1

As expected, the process gain has units of y(t)/u(t) and more specifically h(t)/F0(t) [=] m/(m3/s), and
the time constant has units of time.

Case 2: Assume Drain Flow Proportional to Square Root of Hydrostatic Head

Unfortunately, while Case 1 had a pleasing outcome, drain flow out of a freely draining tank follows
much more closely a square root relationship with liquid height rather than the direct proportional
relationship presented previously. The square root relationship is expressed as such:

F1 (t ) = α 1 h(t )

Substituting yields the following process model:

dh(t )
AC + α1 h(t ) = F0 (t ) where h(0) = hS (11.3)
dt

Comparing the above equation to the general first order process model reveals a problem – there is
not a square root relationship, y (t ) , in the general first order form. Hence, we cannot directly
determine the model process gain and time constant for this case by a simple comparison.
Because the dependent variable, h(t), has a square root relation, this model is a nonlinear
ODE. Before we apply linear control theory to this process model, we must linearize it. We discuss
the method for linearizing nonlinear equations in Chapter 12.

102
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
11.4 Mass Balance on Two Draining Tanks
Consider two freely draining tanks stacked one above the other similar to the Gravity Drained Tanks
case study in LOOP-PRO and as shown below in Fig. 11.2. Although the lower tank receives its feed
from the upper tank, the dynamics and behavior of the lower tank have no impact on the upper tank.
This arrangement is traditionally called non-interacting tanks in series.

Picture:
F0(t)

h1(t) V1(t)

AC 1 F1(t)

h2(t) V2(t)

AC 2 F2(t)

Figure 11.2 - Two non-interacting draining tanks in series with process variables labeled

Variables with units:

- Liquid flow rate F(t) [=] m3/s
- Liquid density ρ(t) [=] Kg/m3
- Liquid level height h(t) [=] m
- Area of cross section AC [=] m2
- Liquid volume V(t) [=] m3
- Time t [=] s

Assumptions:
- the process model is restricted to the liquid volumes and flows
- liquid can enter or leave the tank only through the flow streams shown (i.e. no evaporation)
- tank cross sectional area is constant with height, or V(t) = AC h(t)
- liquids are incompressible and thus density is constant, or ρ0(t) =ρ1(t) =ρ(t) =ρ

103
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
Step by step details:
Applying the mass balance derived in Section 11.3 to each tank, we obtain the following:
dh1 (t )
Tank 1: AC1 = F0 (t ) − F1 (t ) where h1(0) = h1,S
dt

dh2 (t )
Tank 2: AC 2 = F1 (t ) − F2 (t ) where h2(0) = h2,S
dt

Applying the assumption that tank drain flow rate is proportional to the square root of the hydrostatic
head, or
F1 (t ) = α 1 h1 (t ) and F2 (t ) = α 2 h2 (t )

yields the following model equations:

dh1 (t )
Tank 1: AC1 + α 1 h1 (t ) = F0 (t ) where h1(0) = h1,S
dt

dh2 (t )
Tank 2: AC 2 + α 2 h2 (t ) = α 1 h1 (t ) where h2(0) = h2,S
dt

Both of these equations are nonlinear because of the square root relation for the dependent variables,
h1(t) and h2(t). Also note that h2(t) does not appear in the first of these coupled equations, which
supports the earlier statement that the dynamics of the lower tank have no effect on the upper tank.

11.5 Energy Balance on a Stirred Tank with Heater

Here we perform an energy balance on a well-stirred tank of liquid that is being heated with a steam
coil. As indicated in the picture, the flow rates, temperatures and other variables can all change with
time.

Picture:
F0(t) Τ0 (t)

V(t) T(t)
Q(t) h(t)
steam

F1(t) Τ1 (t)
AC
Figure 11.3 - Steam heated mixing tank with process variables labeled

104
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
Variable with units:
- Liquid flow rate F(t) [=] cm3/s
- Liquid density ρ(t) [=] g/cm3
- Liquid level height h(t) [=] cm
- Area of cross section AC [=] cm2
- Liquid volume V(t) [=] cm3
- Energy in from steam Q(t) [=] cal/s
- Liquid temperature T(t) [=] °C
- Liquid heat capacity CP [=] cal/g·°C
- Time t [=] s

Assumptions:
- AC constant with height of liquid, i.e. V(t) = AC h(t)
- ρ and CP constant
- No evaporation
- No other streams enter or leave the vessel than those shown
- Vessel is perfectly insulated
- Liquid is perfectly mixed, i.e. T1(t) = T(t)
- No shaft work (no WS) from mixer
- Only energy of liquid is considered
- Enthalpy balance = Energy balance (or ∆PE = ∆KE = 0)
- Reference Temperature is zero, i.e. Tref = 0
- No reaction occurs
- No frictional losses; no pressure drop losses

Step by step details:

Perform an energy balance on an experiment that takes time ∆t to complete, as follows:

- Energy in by flow: ρC P F0 (t )T (t )∆t

- Energy out by flow: ρC P F1 (t )T (t )∆t
- Energy in from steam: Q(t )∆t

- Energy accumulation in tank:

E |t + ∆t − E |t = H |t + ∆t − H | t = ρAC C P [ h(t )T (t ) |t + ∆t − h(t )T (t ) |t ]

Sum the energy balance (conserve the variable):

ρ AC CP [h(t )T (t ) |t +∆t −h(t )T (t ) |t ] = ρ CP F0 (t )T0 (t )∆t − ρ CP F1 (t )T (t )∆t + Q(t )∆t

105
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
Divide by ∆t and take the limit as ∆t → 0 . Recognize that the average values of the process variables
become point values, or F (t ) → F (t ), T (t ) → T (t ), Q(t ) → Q(t ) , yielding the following process
model:

d [h(t )T (t )] Q (t )
AC = F0 (t )T0 (t ) − F1 (t )T (t ) +
dt ρC P

We can then apply the product rule to the derivative:

dT (t ) dh(t ) Q(t )
AC h(t ) + AC T (t ) = F0 (t )T0 (t ) − F1 (t )T (t ) +
dt dt ρC P

Recall the mass balance for a freely draining tank:

dh(t )
AC = F0 (t ) − F1 (t )
dt

Substituting the mass balance into the energy balance yields

dT (t ) Q (t )
AC h(t ) + T (t )[ F0 (t ) − F1 (t )] = F0 (t )T0 (t ) − F1 (t )T (t ) +
dt ρC P

We can then eliminate like terms to arrive at the energy balance process model for liquid in the tank:

dT (t ) Q(t )
AC h(t ) + T (t ) F0 (t ) = F0 (t )T0 (t ) + where T(0) = TS
dt ρC P

To calculate the tank height used in the energy balance, we must include the mass balance as part of
the overall model:

dh(t )
AC = F0 (t ) − F1 (t ) where h(0) = hS
dt

Note that the energy balance has non-constant coefficients, so we cannot draw any conclusions by
comparing it directly to the general first order process model with constant coefficients:

dy
τP + y (t ) = K P u (t )
dt

The next chapter shows how to approximate non-constant coefficients with constant coefficients
using linearization techniques.

11.6 Species (Component) Balance on a Stirred Tank with Reaction

Here we perform a species (component) balance on a well-stirred tank of liquid where reaction A→B
takes place. As indicated in Fig. 11.4, the flow rates, temperatures and other variables can all change
with time.

106
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
Picture:
F0(t), T0(t), CA0(t)

V(t)
h(t) T(t)
CA(t)

F1(t), Τ(t), CA(t)

AC

Figure 11.4 - Mixing tank with chemical reaction and process variables labeled

Variables with units ( in addition to those in previous example):

- Concentration of A CA [=] gmol/cm3
- Moles of A nA [=] gmol
- Activation energy E [=] cal/gmol
- Gas constant R [=] cal/gmol·K
- Reaction rate constant k [=] 1/s
- Liquid flow rate F(t) [=] cm3/s
- Liquid density ρ(t) [=] g/cm3
- Liquid level height h(t) [=] cm
- Area of cross section AC [=] cm2
- Liquid volume V(t) [=] cm3
- Liquid temperature T(t) [=] K
- Liquid heat capacity CP [=] cal/g·K
- Rate of reaction r [=] gmol/cm3·s
- Time t [=] s

Assumptions (in addition to the previous example):

- There is no heat of reaction (reaction is neither exothermic nor endothermic)
- A → B is a first order reaction, i.e. r(t) = k(t)CA(t)
− E / RT (t )
- Arrhenius temperature dependence on reaction rate constant, i.e. k (t ) = k 0 e

107
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
Details:
Perform a species (component) balance on an experiment that took time ∆t to complete, as follows:
- Moles of A in by flow: F0 (t ) C A0 (t )∆t
- Moles of A out by flow: F1 (t ) C A (t )∆t
- A out due to reaction: r (t )V (t ) ∆t
- Accumulation of A in tank:

n A (t ) | t + ∆t − n A (t ) | t = V (t )C A (t ) | t + ∆t −V (t )C A (t ) | t = AC [h(t )C A (t ) | t + ∆t − h(t )C A (t ) | t ]

Sum up the species balance and take the limit as ∆t → 0. Recognize that the average values of the
process variables become point values, or C A (t ) → C A (t ), F (t ) → F (t ), T (t ) → T (t ) , yielding the
following process model:

d [h(t )C A (t )]
AC = C A0 (t ) F0 (t ) − C A (t ) F1 (t ) − AC h(t )r (t )
dt

Apply the product rule to the derivative and assume an Arrhenius temperature dependence:

dC A (t ) dh(t )
AC h(t ) + AC C A (t ) = C A0 (t ) F0 (t ) − C A (t ) F1 (t ) − AC h(t )C A (t ) k 0 e − E / RT (t )
dt dt

Substituting the mass balance into the species balance and eliminating like terms, we arrive at the
following process model:

dC A (t )
AC h(t ) + C A (t )[ F0 (t ) + AC h(t ) k 0 e − E / RT (t ) ] = C A0 (t ) F0 (t ) where CA(0) = CA,S
dt

An energy balance on this process yields

dT (t )
AC h(t ) + F0 (t )T (t ) = F0 (t )T0 (t ) where T(0)=TS
dt

A mass balance yields the following:

dh(t )
AC = F0 (t ) − F1 (t ) where h(0) = hS
dt

11.7 Exercises

Q-11.1 Consider the stirred tank reactor below. Following good engineering practice, derive the
steady state process gain and process time constant for this system. You should assume first
order reaction kinetics, rA(t) = kCA(t).

108
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 F
CAO (t)

V
CA (t)
F
CA (t)

Q-11.2 A constant density fluid flows into a perfect cone tank as pictured below. Showing all steps,
derive the dynamic ordinary differential equation (ODE) describing liquid height in
dh(t)
the tank. Express your result in the form =?
dt

F0(t)
R

H r(t)
h(t)

FI (t)
Exit flow, F1(t), should be assumed to be proportional to the square root of the hydrostatic
r(t) h(t)
head, and the geometry of the tank indicates that = .
R H

Q-11.3 Consider the stirred reactor below. Following good engineering practice, derive the steady
state process gain and process time constant for this system. You should assume first-order
reaction kinetics, rA(t) = kCA(t). Note that parameters r and h are not time dependent.

F
CA0(t)

CA (t)
h

F
CA (t)

109
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 12. Linearization of Nonlinear Equations and Deviation Variables
12.1 The Linear Approximation
Popular process control theory requires that all equations used in analysis be linear and have constant
coefficients. Recall that all of the process models we derived in the Chapter 11 were nonlinear
ordinary differential equations (ODEs) or linear ODEs with nonconstant coefficients.
The distinction between linear and nonlinear ODEs lies in the treatment of the dependent
variable, y(t). Specifically, if the dependent variable is raised to a power, inversed, or otherwise
manipulated, the equation is nonlinear. Interestingly, the complexity of the forcing function (the right
hand side of the ODE that is not a function of the dependent variable) does not influence whether a
process displays a linear or nonlinear character. For example:

dy (t )
Linear ODE with constant coefficients: A + B y (t ) = t 2 sin( t ) (12.1)
dt

dy (t )
Linear ODE with nonconstant coefficients: A(t ) + B (t ) y (t ) = u (t ) (12.2)
dt

dx (t ) B
Nonlinear ODE with constant coefficients: A + = C u (t ) (12.3)
dt x (t )

2
⎛ dx (t ) ⎞
Nonlinear ODE with nonconstant coefficients: A(t )⎜ ⎟ + B (t ) x (t ) = C ( t ) u ( t ) (12.4)
⎝ dt ⎠

Linearization is a method used to approximate nonlinear equations with nonconstant coefficients as

linear equations with constant coefficients.

linear approximation

nonlinear function

f[x(t)]
good agreement
in this range

xS
x(t)
point of linearization

Figure 12.1 - Linear approximation shows good agreement in a narrow operating range

110
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 As shown in Fig. 12.1, linearization is a procedure for approximating a nonlinear function
with a simple linear function. The linear approximation is exact at one point and has good agreement
in a range around that point. The approximation degrades as we move away from the point of
linearization.
Fortunately, we get to choose the point around which we will linearize. We choose the point
where the process will spend most of its time, the controller design level of operation. This should
correspond to the normal or expected set point, which in turn should correspond to the normal or
expected value of the process variable.

12.2 Linearization for Functions of One Variable

The method for linearizing a nonlinear equation f[x(t)] is to approximate with a Taylor series
expansion around the point xS:

f [ x (t )] = f [ x S ] + [ x (t ) − x S ]
d
f [ x (t )] +
[x(t ) − x S ]2 d2
f [ x (t )] x = xS + K (12.5)
x = xS 2
dx 2! dx
144444424444443
HOT

We consider only the linear terms of the expansion and ignore the higher order terms (HOT) of the
Taylor series. As shown in the following examples, isolating and linearizing the individual nonlinear
terms of a complicated equation and then substituting them back into the original form is often more
convenient than linearizing the entire original equation at once.

Example: Tank Liquid Level Model

A mass balance on a gravity drained tank as detailed in Section 11.3 produces the following tank
liquid level model:

dh (t )
AC = F0 (t ) − F1 (t ) where h(0)=hS (12.6)
dt

Assume that drain flow is proportional to the square root of hydrostatic head, or F1(t) = α h1/2(t),
and the model becomes the following:

dh(t)
AC + α h1 2 (t) = F0 (t) where h(0)=hS (12.7)
dt

The nonlinear term in the ODE is h1/2(t). We isolate this nonlinear term and linearize h1/2(t)
around the design level hS (a tank height design value) using the Taylor series approximation:

d 12
h1 2 (t ) ≅ hS1 2 + [h(t ) − hS ]{ [h (t )]} h = hS
dh

⎛ ⎞
⎜ 1 ⎟
≅ h1S 2 + [ h(t ) − hS ] ⎜ 1 ⎟
⎜ 2h 2 ⎟
⎝ S ⎠

111
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 1 12 1
≅ hS1 2 − hS + 1 2 h(t )
2 2hS
The linear approximation then becomes:

1 12 1
h1 2 (t ) ≅ hS + 1 2 h(t ). (12.8)
2 2 hS

Substitute the linear approximation into the original ODE to obtain our linear process model:

dh(t ) α α
AC + 1 2 h(t ) = F0 (t ) − h1S 2 where h(0)=hS (12.9)
dt 2h S 2

The general form of the dynamic model above is a linear ODE with constant coefficients:

dy (t )
A + B y (t ) = C u ( t ) + E (12.10)
dt

We compare this to the general first-order dynamic model (without dead time):

dy (t )
τP + y (t ) = K P u (t ) where y(0)= 0
dt

Although the tank height model ODE is linear with constant coefficients, we still cannot
determine KP andτP by comparison. This is because of the extra constant term, E, on the right-
hand side of the equation and the non-zero initial condition, hS.
Õ Õ Õ

12.3 Linearization for Functions of Two Variables

Linearize f[x(t),y(t)] and g[x(t),y(t)] around the point (xS, yS) using Taylor series expansion:

dx (t ) dy (t )
= f [x (t ), y (t )] = g [x(t ), y (t )]
dt dt

As with the single variable expansion, we ignore the higher order Taylor series terms, so the
approximations are as follows:

∂f ∂f
f [x(t ), y (t )] ≈ f ( x S , y S ) + [x(t ) − x S ] | x S , y S + [ y (t ) − y S ] | x S , y S
∂x ∂y
and
∂g ∂g
g [x(t ), y (t )] ≈ g ( x S , y S ) + [x(t ) − x S ] | xS , yS + [ y (t ) − y S ] | xS , yS
∂x ∂y

112
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
Again, individually isolating and linearizing the nonlinear terms and then substituting them back into
the original equation is often more convenient than linearizing the entire original equation at once.

Example: Consider the Tank Species Balance Model

A species (component) balance on a mixing tank in Section 11.6 produced the following tank
concentration model:

dC A (t ) F
= [C A0 − C A (t )] − k 0 e − E / RT (t ) C A (t ) where CA(0) = CAS
dt V

The nonlinear term, e-E/RT(t)CA(t), is a function of two variables, T(t) and CA(t). We isolate the
nonlinear term and linearize around CAS ,TS (design concentration and temperature values):

d u ( x) du ( x)
Using the Taylor series expansion and recalling that e = e u ( x) , then
dx dx

( )
e− E / RT (t ) C A (t ) ≈ e− E / RTS C AS + [T (t ) − TS ] e− E / RTS C AS E / RTS2 + [C A (t ) − C AS ] e− E / RTS

Substituting the linearized term into the original ODE yields the following:

dC A (t ) F
= [C A 0 − C A (t )]
dt V
{ ( )
− k 0 e − E / RTS C AS + [T (t ) − TS ]e − E / RTS C AS E / RTS2 + [C A (t ) − C AS ]e − E / RTS }

The resulting dynamic model is a linear ODE with constant coefficients of general form:

dy (t )
A + B y ( t ) = C u ( t ) + D d (t ) + E
dt

The variable d(t) might be a disturbance variable that impacts the measured process variable
y(t). As in the single variable case, we have an extra constant term, E, on the right hand side of
the equation.
Õ Õ Õ

12.4 Defining Deviation Variables

Deviation variables (also called perturbation variables) are used to recast an equation into a more
convenient form for control system analysis. Specifically, the time-dependent variables in an equation
are recast in terms of their deviation from a fixed value. For controller design, the value we choose is
not only the design level of operation, but also the point of linearization, and the point around which
we generated dynamic data. In a proper design, all of these have the same value.
For the time-dependent variable x(t), the deviation variable xP(t) is defined relative to fixed
point xS as such:
xP(t) = x(t) - xS

113
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
This is shown graphically in Fig. 12.2:

x(t) x(t) xP(t)

xs

time

Figure 12.2 - Deviation variable xP(t) is the deviation of x(t) from constant value xS

12.5 Deviation Variables Simplify the Equation Form

Deviation variables are useful in preparing equations for control system analysis for two important
reasons:
- using them eliminates constant terms, and
- the initial conditions for the ODEs become zero.

To explore this, consider the linear(ized) ODE:

dx (t )
= Ax (t ) + B where x(0) = xS
dt

At the point x(t) = xS, the ODE becomes

dx S
= Ax S + B where x(0) = xS
dt

Subtracting the second equation from the first, we obtain

d
[ x(t ) − x s ] = A[ x(t ) − x s ] + [ B − B ] where x(0) − x(0) = xS − xS = 0
dt

We then apply the definition of the deviation variable to obtain the following:

dx p (t )
= Ax p (t ) where xP(0) = 0
dt

Note that this form, expressed as a difference from xS, has no extra constant terms and has an initial
condition of zero.

114
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
Example: Tank Liquid Level Model
Recall the linearized tank liquid level model:

dh(t ) α α
AC + 1 2 h(t ) = F0 (t ) − h1S 2 where h(0) = hS
dt 2hS 2
We choose the design variables as h(t) = hS and F0(t) = F0,S. We will use these later to define
the deviation variables. Note that these values are not independent from each other because if you
maintain the inlet flow rate at F0,S for a sufficient length of time, the liquid level should ultimately
steady at hS.
Substituting these design variables into the original ODE, we obtain

dhS α α
AC + 1 2 hS = F0, S − h1S 2 where h(0) = hS
dt 2hS 2

We then subtract the second ODE from the first:

AC
d
[h(t ) − hS ] + α1 2 [h(t ) − hS ] = F0 (t ) − F0,S − ⎡⎢α h1S 2 − α h1S 2 ⎤⎥ where h(0) − h(0) = 0
[ ]
dt 2hS ⎣2 2 ⎦

and define the deviation variables:

h P (t ) = h(t ) − hS
F0P (t ) = F0 (t ) − F0 S

Substituting these, we obtain the linear ODE with constant coefficients and a zero initial condition:

dh P (t ) α
AC + 1 2 h P (t ) = F0P (t ) where hP(0) = 0
dt 2hS

We can now compare this to the general first order ODE without dead time. We have not
previously mentioned it, but the initial condition is zero for this general form:

dy (t )
τP + y (t ) = K P u (t ) where y(0) = 0
dt

We now have a payoff for this effort. By comparing the model forms, we can determine the
steady state process gain and overall time constant for this process:

2h1S 2 2 AC h1S 2
KP = τP =
α α
These KP and τP values are only exact at point (h(t)=hS, F0(t)=F0,S). They become
approximations that decrease in accuracy as h(t) moves away from liquid height hS or F0(t) moves
away from F0,S.

Õ Õ Õ

115
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
12.6 Exercises

Q-12.1 Consider the process below. A control system is under development where FO(t) is the
manipulated variable and CA(t) is be the measured variable.

FO(t)
CAO

h(t) V(t)
CA (t)
F(t)
CA (t)

AC

a) Assuming a third order reaction for A → B, or r(t) = k[CA(t)]3, and a constant cross section, or
V(t) = ACh(t), show all steps to show that the ODE describing this system is

dC A (t ) F (t ) C
AC + O C A (t ) + AC k [C A (t )]3 = AO FO (t )
dt h(t ) h(t )

with the initial condition CA(0) = CA,S.

b) Linearize this ODE and put it in deviation variable form. Assume that when FO(t) = FO,S, then
at steady state CA(0) = CA,S and h(0) = hS.

c) Near the expected point of operation, what are the (approximate) process gain and time
constant values?

Q-12.2 The density of an ideal gas, ρ (t ) , can be expressed as follows:

MP(t )
ρ (t ) =
RT (t )

where M is the molecular weight and R is the ideal gas constant. Showing all steps, linearize
this expression around ρ S to obtain a linear approximation for density as a function of
temperature and pressure.

116
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 13. Time Domain ODEs and System Behavior
13.1 Linear ODEs
After deriving process models, linearizing them, and simplifying them with deviation variables, we
are left with linear ODEs (ordinary differential equations) with constant coefficients:

dy (t )
1st order: a1 + a 0 y (t ) = Q(t )
dt

d 2 y (t ) dy (t )
2nd order: a2 2
+ a1 + a 0 y (t ) = Q (t )
dt dt

d 3 y (t ) d 2 y (t ) dy (t )
rd
3 order: a3 3
+ a2 2
+ a1 + a 0 y (t ) = Q(t )
dt dt dt

The coefficients of these equations signal specific dynamic behaviors. By solving these ODEs in the
time domain, we learn the relationship between an equation form and the dynamic behavior it implies.

13.2 Solving First Order ODEs

For the general first order ODE of the form

dy (t )
+ P(t ) y (t) = Q (t ) where y(0) = yS
dt

the solution method involves defining an integrating factor, solving a general solution form based on
this integrating factor, and then solving for the constant of integration using the initial conditions.

An integrating factor, µ, is defined as follows:

P (t )dt
µ = e∫

1⎡
The solution to the ODE is then y (t ) = ∫ µ Q(t )dt +C1 ⎤⎥
µ ⎢⎣ ⎦

where the constant C1 is computed from the initial condition y(0) = yS.

Example 1: Solve this ODE in the time domain:

d y (t )
+ 2 y (t ) = 4t where y(0) = 1
dt

Solution: Using the steps outlined, we first compute the integrating factor, µ = e ∫
2 dt
= e 2t

117
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 1 ⎡ ⎤ ⎡ ⎤
so the solution becomes y (t ) =
e 2t ⎢
⎣ ∫ ⎦ ⎣ ∫
e 2t (4t )dt + c1 ⎥ = e − 2t ⎢ 4 te 2t dt + c1 ⎥
⎦

∫
We must then integrate by parts: udv = uv − vdu ∫
1 2t
u = t; du = dt; v= e ; dv = e 2t dt
2

Substituting the result, the solution becomes

⎡4 4 2t ⎤
y (t ) = e − 2t ⎢ te 2t −
⎣ 2 2 ∫
e dt + c1 ⎥
⎦

⎡ ⎤
= e −2t ⎢ 2te 2t e 2t +c1 ⎥
⎣ ⎦

= 2t − 1 + c1e −2t

Next, we apply initial conditions: @ t = 0, y = 1, so 1 = −1 + c1, and thus c1 = 2,

which yields the total solution: y (t ) = 2e −2t + 2t − 1

Õ Õ Õ

Example 2: Solve this ODE in the time domain:

dx (t )
= x(t ) + sin t where x(0) = 0
dt

Solution: We first rewrite the ODE as such:

dx(t )
− x(t ) = sint
dt

Then, we compute the integrating factor:

µ = e∫
−1d t
= e −t

⎡ ⎤
The solution then becomes
⎣ ∫
x(t ) = e t ⎢ e −t sint dt + c1 ⎥
⎦
From integral tables, we can see that
1 −t ⎡ ⎤
∫e
−t
sint dt = e ⎢ sint − cost ⎥
2 ⎣ ⎦

118
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Substituting this, the solution becomes
⎡ 1 ⎤
x(t ) = e t ⎢− e −t (sint + cost ) + c1 ⎥
⎣ 2 ⎦
1
= − (sint + cost ) + c1e t
2

1 1
We then apply initial conditions: @ t = 0, x = 0, so 0 = − [sin(0) + cos(0)] + c1 , thus c1 = ,
2 2

1⎡ t
which yields the total solution: x(t ) = e − sin t − cos t ⎤
2⎣ ⎦

Õ Õ Õ

13.3 Deriving "τp = 63.2% of Process Step Response" Rule

Our ability to solve linear first order ODEs now allows us to derive the “τp = 63.2% of Process Step
Response” rule used in model fitting, as referred to in Section 3.4.
Process Variable, y(t)

AKP

AKP
0.632 AKP
0
Controller Output, u(t)

A
A
0

0 τ
Time, (t)

Figure 13.1 – Response of true first order process to a step change in controller output

The figure above shows the open loop step response of a true first order process model:

dy (t )
τp + y (t ) = K p u (t ) where y(0) = 0
dt

As shown in Fig 13.1, the measured process variable, y(t), and controller output signal, u(t), are
initially at steady state with y(t) = u(t) = 0 for t < 0. At time t = 0, the controller output is stepped to
u(t) = A, where it remains for the duration of the experiment. Hence, the first order model becomes

dy (t )
τp + y (t ) = AK p
dt

119
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
To solve this ODE, rearrange as:

dy (t ) 1 AK p
+ y (t ) =
dt τp τp

First compute the integrating factor, µ = e ∫

(1 / τ p ) dt t /τ p
=e . Solving the ODE using this integrating
factor yields
1 ⎡ t / τ ⎛ AK p ⎞ ⎤
y (t ) =
e
t /τ p ∫
⎢ e p⎜
⎢⎣ ⎜ τp
⎝
⎟d t + c1 ⎥
⎟
⎠ ⎥⎦

−t / τ p ⎡ AK p t /τ p ⎤
=e ⎢
⎢⎣ τ p
∫e d t + c1 ⎥
⎥⎦

1 ax
∫e
ax
Recall that dx = e
a

∫e
t /τ P
and thus dt = τ P e t / τ P

so y (t ) = e
−t / τ p
[AK e p
t /τ p
+ c1 ]
−t / τ p
= AK p + c1e

Next, apply the initial condition: @ t = 0, y = 0, 0 = AKp + c1, and thus c1 = −AKp.

Substituting and rearranging, we obtain the solution to the ODE:

⎡ −t/τ ⎤
y (t ) = AK p ⎢1 − e P ⎥
⎣ ⎦

After the passage of one time constant, time t = τp, the solution becomes

⎡
⎣
−τ / τ ⎤
y (τ p ) = AK p ⎢1 − e P P ⎥ = AK p 1 − e −1
⎦
[ ]
Therefore, the measured process variable step response at time t = τp is

y (τ p ) = 0.632 AK p

As we set out to show, the measured process variable will have traveled to 63.2% of the total change
that it will ultimately experience at time t equals one time constant,τP.

120
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
13.4 Solving Second Order ODEs
For the general second-order ODE with constant coefficients,

d 2 y (t ) dy (t )
a2 2
+ a1 + a 0 y (t ) = Q (t )
dt dt

the total solution, y(t), is the sum of the particular solution, y p (t ) , and the complementary solution,
y c (t ) , or
y (t ) = y p (t ) + y c (t )

The complementary solution, y c (t ) , is the solution of the ODE when Q(t ) = 0 , or

d 2 y (t ) dy (t )
a2 2
+ a1 + a0 y (t ) = 0
dt dt

To solve the complementary solution, form the characteristic equation using operator notation:

a 2 s 2 + a1 s + a0 = ( s − p1 )( s − p 2 ) = 0

The roots of the characteristic equation, p1 and p 2 , are computed using the quadratic equation:

− a1 ± a12 − 4a 2 a 0
p1 , p 2 =
2a 2

The coefficients a 2 , a1 and a 0 define the roots of the characteristic equation, and the roots define the
form of the complementary solution.

Case 1: If a12 − 4a 2 a 0 > 0 , then the characteristic equation of the complementary solution will have
two distinct real roots, p1 and p2, and the solution will have the form

y (t ) = y p (t ) + c1e p1t + c 2 e p2t

Case 2: If a12 − 4a 2 a 0 = 0 , then the characteristic equation of the complementary solution will have
two equal or repeated roots, p3 = p4 , and the solution will have the form

y (t ) = y p (t ) + (c1 + c 2 t ) e p3t

Case 3: If a1 2 − 4a 2 a 0 < 0 , then the characteristic equation of the complementary solution will have
two complex conjugate roots, p 5 ± p 6 i , and the solution will have the form

y (t ) = y p (t ) + e p5t ⎡⎣ c 1 cos ( p6 t ) − c2 sin ( p6 t ) ⎤⎦

121
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
Note that the roots of the characteristic equation provide important clues about the dynamic behavior
+t -t
described by an ODE. Stability relates to the sign of a root because e grows without bound and e
approaches zero as time increases. And as results in Case 3, a solution with sines and cosines will
have a natural tendency to oscillate.

Example 1: Solve this ODE in the time domain:

d 2 y (t ) dy (t ) dy (t )
2
+4 + 3 y (t ) = t where y (0) = =0
dt dt dt t =0

d 2 y (t ) dy (t )
Complementary Solution: +4 + 3 y (t ) = 0
dt 2 dt

− 4 ± 16 − 12
The characteristic equation is s 2 + 4s + 3 = 0 with roots p1 , p 2 = = −1,− 3
2

so yc(t) solution has the form y c (t ) = c1e −t + c 2 e −3t

Particular Solution: The right hand side of the original ODE indicates that y p (t ) = c3 + c 4 t

dy p (t ) d 2 y p (t )
so = c4 and =0
dt dt 2

Substituting this into the original ODE yields

(0) + 4(c 4 ) + 3(c3 + c 4 t ) = t

We then equate like terms and solve for constants:

1
3c 4 t = t so c4 =
3
⎛1⎞ 4
and 4c 4 + 3c3 = 0 so 4⎜ ⎟ + 3c3 = 0 and c3 = −
⎝3⎠ 9

so y p (t ) = − 4 + 1 t
9 3

Total Solution: y (t )= y p (t ) + y c (t ) = − 4 + 1 t + c1e −t + c 2 e −3t

9 3

We apply initial conditions to determine the constants:

1. @ t = 0, y = 0 0 = − 4 + 0 + c1 + c 2
9

dy (t ) dy (t )
2. @ t = 0, =0 = 0 = 1 − c1e 0 − 3c 2 e 0
dt dt 3

122
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 c1 + c 2 = 4 ⎫⎪ 1
9 1
Solving simultaneously yields
1 3 ⎬ c1 = , c 2 = −
c1 + 3c 2 = = ⎪ 2 18
3 9⎭

and thus the total solution is y (t ) = − 4 + 1 t + 1 e −t − 1 e −3t

9 3 2 18

Example 2: Solve this ODE in the time domain:

d 2 y (t ) dy (t ) dy (t )
2
+2 + y (t ) = t 2 where y (0) = 0 and =1
dt dt dt t =0

d 2 y (t ) dy (t )
Complementary Solution: 2
+2 + y (t ) = 0
dt dt

−2± 4−4
The characteristic equation is s 2 + 2s + 1 = 0 with roots p1 , p 2 = = −1,− 1
2

so yc(t) solution has the form y c (t ) = (c1 + c 2 t )e −t

Particular Solution: The right hand side of the ODE indicates that y p (t ) = c3 + c 4 t + c5 t 2

dy p (t ) d 2 y p (t )
so = c 4 + 2c 5 t and = 2c5
dt dt 2

Substituting this into the original ODE yields

(2c5 ) + 2(c 4 + 2c5 t ) + (c3 + c 4 t + c5 t 2 ) = t 2

equate like terms, 2c5 + 2c 4 + c3 = 0 4c5 t + c 4 t = 0 c5 t 2 = t 2

and solve for constants: c5 = 1 c 4 = −4 c3 = 6

so y p (t ) = 6 − 4t + t 2 .

Total Solution: y (t ) = y p (t ) + y c (t )=6 − 4t + t 2 + (c1 + c 2 t )e −t

Apply initial conditions to determine the constants:

1. @ t = 0, y = 0 0 = 6 − 0 + 0 + [c1 + c 2 (0)]e 0

dy
2. @ t = 0, =1 1 = −4 + 2(0)+[−6 + c 2 (0)]e 0 (−1) + c 2 e 0
dt

123
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
Solving for constants yields: c1 = −6 and c 2 = −1

and thus the total solution is y (t ) = 6 − 4t + t 2 − (6 + t )e −t

Õ Õ Õ

Example 3: Solve this ODE in the time domain:

d 2 y (t ) dy (t ) dy (t )
2 2
+2 + y (t ) = 2 where y (0) = =0
dt dt dt t =0

d 2 y (t ) dy (t )
Complementary Solution: 2 2
+2 + y (t ) = 0
dt dt

−2± 4−8 1 1
The characteristic equation is 2 s 2 + 2 s + 1 = 0 with roots p1 , p 2 = =− ± i
4 2 2

( − 1 + 1 i )t ( − 1 − 1 i )t
so yc(t) solution has the form y c (t ) = c1e 2 2 + c2 e 2 2

−t
⎡ it −i t ⎤
= e 2 ⎢ c1e 2 + c2 e 2 ⎥
⎣ ⎦

Recall the identities e iat = cosat + isin at , cos(− at ) = cos( at ) , and sin (− at ) = −sin (at )

Applying these, the complementary solution becomes

−t ⎧ ⎡ t t⎤ ⎡ t t ⎤⎫
y c (t ) = e 2
⎨c1 ⎢cos + isin ⎥ + c 2 ⎢cos − isin ⎥ ⎬
⎩ ⎣ 2 2⎦ ⎣ 2 2 ⎦⎭

1 1
If we let c1 = (c R + ic I ) and c2 = (c R − ic I )
2 2

where cR and cI are real constants, then

−t ⎧ t t⎫
y c (t ) = e 2
⎨c R cos − c I sin ⎬
⎩ 2 2⎭

Particular Solution: The right hand side of the original ODE indicates that y p (t ) = c3

dy p (t ) d 2 y p (t )
so =0 and =0.
dt dt 2

Substituting this into the original ODE yields

2(0) + 2(0) + c3 = 2 , and thus c3 = 2 ,

124
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
so y p (t ) =2

−t ⎧ t t⎫
Total Solution: y (t )= y p (t ) + y c (t ) = 2 + e 2
⎨c R cos c I sin ⎬
⎩ 2 2⎭

We apply initial conditions to determine the constants:

1. @ t=0, y=0 0 = 2 + e 0 {c R cos0 − c I sin 0} so c R = −2

d y (t ) d y (t ) 1 −t ⎡ t t⎤
2. @ t=0, =0 = 0 =− e 2 ⎢ −2cos 2 − cI sin 2 ⎥
dt dt 2 ⎣ ⎦

−t ⎡ ⎛ t ⎞⎛ 1 ⎞ ⎛ t ⎞⎛ 1 ⎞⎤
+e 2
⎢− 2⎜ − sin 2 ⎟⎜ 2 ⎟ − c I ⎜ cos 2 ⎟⎜ 2 ⎟⎥
⎣ ⎝ ⎠⎝ ⎠ ⎝ ⎠⎝ ⎠⎦

so at t = 0, c I = 2 .

−t ⎧ t t⎫
The total solution becomes y (t ) = 2 − 2e 2
⎨cos + sin ⎬
⎩ 2 2 ⎭

Õ Õ Õ

Example 4: Solve this ODE in the time domain:

d 2 x(t ) dx(t )
2
+ x(t ) = sin (2t ) where x(0) = 0 and =1
dt dt t =0

d 2 x(t ) dx(t )
Complementary Solution: 2
+0 + x(t ) = 0
dt dt

The characteristic equation is s 2 + 0s + 1 = 0 with roots p1 , p 2 = ±i

so xc(t) solution has form x c (t ) = c1 e it + c 2 e −it

We recall the identity e iat = cosat + isin at

1 1
We then let c1 = (c R + ic I ) and c2 = (c R − ic I )
2 2

so x c (t ) = c R cos(t ) − c I sin(t )

125
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Particular Solution: The right hand side of the ODE indicates that x p (t ) = c3 cos2t + c 4 sin 2t

dx p (t )
so = −2c3 sin 2t + 2c 4 cos2t
dt
d 2 x p (t )
and = −4c3 cos2t − 4c 4 sin 2t
dt 2

Substituting into the original ODE yields

− 4c3 cos2t − 4c 4 sin 2t + c3 cos2t + c 4 sin 2t = sin 2t

We then equate like terms: − 4c 4 sin 2t + c 4 sin 2t = sin 2t − 4c3 cos2t + c3 cos2t = 0

1
Solving for constants yields c4 = − and c3 = 0
3
1
so the particular solution is x p (t ) = − sin 2t
3

1
Total Solution: x(t ) = x p (t ) + x c (t ) = − sin 2t + c R cost − c I sint
3

We apply initial conditions to determine the constants:

1
1. @ t = 0, x = 0 0 = − sin 0 + c R cos0 − c I sin 0 so c R = 0
3

dx(t ) 2 5
2. @ t = 0, =1 1 = − cos0 − c I cos0 so c I = −
dt t =0 3 3

1 5
and thus the total solution is x(t ) = − sin 2t + sint
3 3

Õ Õ Õ

13.5 The Second Order Underdamped Form

A general and more useful form of the second order ODE is written

d 2 y (t ) dy (t ) dy (t )
τ n2 2
+ 2τ nξ + y (t ) = K P u (t ) where y (0) = =0
dt dt dt t =0

and y(t) = measured process variable

u(t) = forcing function (controller output)
KP = steady state process gain
τn = natural period of oscillation
ξ = damping factor

126
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
The total solution, y(t), is the sum of the particular solution, yP(t) , and the complementary solution,
yC(t):

y (t ) = y p (t ) + y c (t )

The complementary solution form is

d 2 y (t ) dy (t )
τ n2 2
+ 2τ n ξ + y (t ) = 0
dt dt

The characteristic equation is thus

τ n2 s 2 + 2τ n ξ s + 1 = ( s − p1 )( s − p 2 ) = 0
with roots
− 2τ nξ ± 4τ n2ξ 2 − 4τ n2 − ξ ± ξ 2 −1
p1 , p 2 = =
2τ n2 τn

As mentioned in Section 13.4, the roots of the characteristic equation provide information
about the behavior of the system. If the roots yield sines and cosines, there will be a natural tendency
for the system to oscillate. If the roots yield even one positive real part, then e + t for that term grows
without bound as time increases, leading to an unstable system. If all real parts are negative, then
e −t approaches zero for all terms as time increases, indicating that dynamics die out and the system
remains stable.
The second order underdamped form written above is useful for a quick evaluation of the
nature of the system dynamics based solely on the value of the damping factor,ξ. Since τn must be
positive (it is a period of time), complex roots and positive or negative real parts arise based solely on
the value of the damping factor.

13.6 Roots of the Characteristic Equation Indicate System Behavior

As shown in Fig. 13.2, yp is a constant for time t > 0. This would result from a step change in the
forcing function. While not a necessary assumption for the derivations that follow, it does
permit the characteristic behaviors of the different cases to be visually apparent.

− ξ ± ξ 2 −1
Case 1: ξ > 1 (overdamped) ( s − p1 )( s − p 2 ) = 0 ; − p1 , − p 2 =
τn
The overdamped case, ξ > 1, yields real negative distinct roots, −p1, −p2 (note that the negative signs
for p1, p2 are shown explicitly).

The total solution is thus y (t ) = y p + c1e − p1t + c2 e− p2t .

As time passes, the real exponentials e − p1t and e − p2t approach zero as time grows toward infinity. The
lack of imaginary roots leads to a lack of sine and cosine terms and their characteristic oscillatory
effect. Thus, y(t) approaches yp slowly, exponentially and without oscillations (a stable system).

127
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 ξ=0
case 4
y(t)

yp

0<ξ<1
case 3
ξ >1
ξ =1 case 1
case 2

Time, t

Figure 13.2 - Stable behaviors of underdamped system based on value of ξ

− ξ ± ξ 2 −1 1
Case 2: ξ = 1 (critically damped) ( s − p1 )( s − p 2 ) = 0 ; − p1 ,− p 2 = = −
τn τn
The critically damped case, ξ = 1, yields real negative repeated roots (note that the negative signs for
p1, p2 are shown explicitly).

The total solution is thus y (t ) = y p + (c1 + c2t )e−t τ n .

As time passes, the real exponential e −t / τ n approach zero as time grows toward infinity. The lack of
imaginary roots leads to a lack of sine and cosine terms and their characteristic oscillatory effect.
Thus, y(t) approaches yp exponentially and without oscillations (a stable system).

− ξ ± ξ 2 −1 − ξ ± i 1− ξ 2
Case 3: 0 < ξ < 1 (underdamped) − p1 , − p 2 = =
τn τn

The underdamped case, 0 < ξ < 1, yields distinct roots with a negative real part and an imaginary part.

⎛ i 1−ξ 2 i 1−ξ 2 ⎞
⎜ t − t ⎟
The total solution is thus y (t ) = y p + e −ξ t / τ n ⎜ c1e τn
+ c2 e τn
⎟.
⎜ ⎟
⎝ ⎠

We then apply the Euler identities, e iθ = cos θ + i sin θ and e −iθ = cos θ − i sin θ

128
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
and define c1 = 0.5(cR + icI ) and c2 = 0.5(cR − icI ) . The total solution thus becomes

⎛ 1−ξ 2 1−ξ 2 ⎞
y (t ) = y p + e −ξ t / τ n ⎜ cR cos( t ) − cI sin( t) ⎟ .
⎜ τn τn ⎟
⎝ ⎠

As time passes and for all values of the damping factor in the range 0 < ξ < 1, the sines and cosines in
the solution indicate that the system has a natural tendency to oscillate. The real exponential
e −ξt / τ n approaches zero as time grows toward infinity. Because the term that causes oscillations is
multiplied by a term that is decreasing to zero, the result is a damping oscillation. Hence, as time
passes, y(t) approaches yp exponentially and with oscillations (a stable system).

− ξ ± ξ 2 −1 −1 i
Case 4: ξ = 0 (undamped) − p1 , − p2 = = ± = ±
τn τn τn

The undamped case, ξ = 0, yields repeated roots with an imaginary part and no real part.

⎛ t t ⎞
The total solution is thus y (t ) = y p + ⎜ cR cos( ) − cI sin( ) ⎟
⎝ τn τn ⎠

As time passes, y(t) oscillates around yp with constant amplitude. This is due to the lack of a real
exponential term to dampen the oscillations as time increases. Oscillations that neither grow nor die
out indicate a system at the limit of stability. Any increase in ξ produces a damping oscillation as
shown in Case 3. Any decrease in ξ produces an unstable system as shown in Case 5 below.

Case 5: ξ < 0 (unstable)

All values of the damping factor, ξ, that are less than zero yield a solution that has a positive real part.
As time passes, e + t will grow without bound and the process will display diverging (unstable)
behavior. The behaviors of the solutions below are shown in Fig. 13.3.

⎛ 1−ξ 2 1−ξ 2 ⎞
+ ξ t /τ n ⎜ cR cos(
−1 < ξ < 0: y (t ) = y p (t ) + e t ) − cI sin( t) ⎟
⎜ τn τn ⎟
⎝ ⎠

ξ = −1: y (t ) = y p (t ) + (c1 + c2t )e+ t τ n

ξ < −1: y (t ) = y p (t ) + c1e+ p1t + c2 e+ p2t

129
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 y(t) ξ ≤ –1

yp(t)

–1<ξ<0

Time, t

Figure 13.3 - Unstable behaviors of underdamped system based on value of ξ in Case 5

13.7 Exercises

Q-13.1 For each ODE below, determine the natural period of oscillation,τn, and damping factor, ξ.
Based on the damping factor, describe the inherent nature of the system, e.g.
- is it naturally stable or unstable?
- does it have a natural tendency to oscillate or not?

d 2 y (t ) dy (t ) dy (t )
a) 2
+4 + 3 y (t ) = t where y (0) = =0
dt dt dt t =0

d 2 y (t ) dy (t ) dy (t )
b) 2
+2 + y (t ) = t 2 where y (0) = 0 and =1
dt dt dt t =0

d 2 y (t ) dy (t ) dy (t )
c) 2 2
+2 + y (t ) = 2 where y (0) = =0
dt dt dt t =0

d 2 x(t ) dx(t )
d) 2
+ x(t ) = sin (2t ) where x (0) = 0 and =1
dt dt t =0

130
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
Q-13.2 Use Custom Process to explore the second order linear model form:

d 2 y (t ) dy (t )
τ n2 + 2τ nξ + y (t ) = K P u (t )
dt 2 dt

Perform this study in open loop to understand how the parameters impact system behavior.

a) Start by clicking on the Custom Process button on LOOP-PRO’s main screen; then choose
Single Loop Process. When the simulation starts, notice that the graphic to the right of the
scrolling plots is comprised of a Process button, Disturb(ance) button and Controller button
(the C in the white circle).

Click the Process button on the graphic. This opens a “Construct Process and Disturbance
Models” form. On the Process Model tab, select Underdamped Linear Model and Self
Regulating (Stable) Processes.

Custom Process permits the construction of sophisticated model forms. Here we explore the
simple second order linear model form shown above.

For the underdamped self regulating linear process model, enter the following parameters:

Process Gain, KP 1
Natural Period, τPn 10
Damping Factor, ξ 0
Time Constant, τP 0
Lead Time, τPL 0
Dead Time, θP 0

We will not be studying the disturbance model, so when you are finished entering the process
model parameters, click Done at the bottom of the form to start the simulation.

b) When the damping factor, ξ, equals zero (as is the case above), the process is said to be
undamped. When forced by a step in the controller output signal, an undamped process will
oscillate with a period correlated to the natural period of oscillation, τn.

Specifically, the period of oscillation as measured on a strip chart, T, is related to τn as

T
τn =
2π

Step the controller output from 50 to 51 and let the measured process variable oscillate
through a few complete cycles. Does the process display an undamped character?

Pause the process and view a fixed plot. Measure on the plot the amount of time, T, for the
process to complete one cycle (from peak to peak or from trough to trough). Use the above
relation to confirm that the natural period specified in your Custom Process model matches
the behavior observed on the plot.

131
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
c) Return to your simulation and click the Process button on the graphic. On the “Construct
Process and Disturbance Models” menu, change the natural period of oscillation to 15. As
you did in step (b) above, step the controller output from 50 to 51. After the measured
process variable completes a few complete cycles, pause the process and view a fixed plot.

From the plot, determine T, the time for the process to complete one cycle. Use the above
relation of T to τn to confirm that the natural period specified in your Custom Process model
matches the behavior observed on the plot.

d) Explore how the damping factor, ξ, impacts system behavior. Click the Process button on the
graphic. On the “Construct Process and Disturbance Models” menu, change the natural
period of oscillation, τn, to 10 and the damping factor, ξ, to 0.3. Step the controller output
from 50 to 51, and when the dynamics die out, back to 50.

Repeat for a damping factor of 0.5, 0.7, 1.0 and 2.0. Describe how the system behavior you
observe relates to the chapter discussion about the damping factor.

Now set the damping factor to -0.1 and step the controller output from 50 to 51. Describe
how the system behavior you observe relates to the chapter discussion about the damping
factor.

132
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 14. Laplace Transforms
14.1 Laplace Transform Basics
Transforming functions from the time domain into the Laplace domain provides a convenient means
for manipulating and solving linear ODEs with constant coefficients. In particular, Laplace
transforms enable us to solve ODEs using algebra instead of calculus. They also provide a
straightforward method for handling the mathematical time shift associated with dead time equations.
Thus, complicated analysis can be performed in a straightforward manner. As we learned in previous
chapters, few processes are accurately described by linear models with constant coefficients, but
linearization and deviation variable techniques enable us to recast ODEs into a linear, constant
coefficient form.
Laplace transforms map an equation from the time domain (t) into the Laplace domain (s).
The definition of the Laplace Transform is

∞
L[ f (t )] ≡ ∫ f (t )e− st dt ≡ F (s)
0

The Laplace independent variable, s, is defined in the complex plane as s = a + bi, where a is the real
part and b is the imaginary part. To help visualize the complex s plane, plotted below in Fig. 14.1 are
the points A = 2 + i and B = – 2 – 2i.

Im
3

A
Re
–3 3

B
–3

s plane

Figure 14.1 – The complex plane (s = a + bi)

While the mapping of a function f(t) → F(s) can be derived using the above definition, we are
fortunate that tables have been created that list the transforms for common functions (see Appendix
B). We detail the derivation of a few entries to show that the table has a basis in the theoretical
definition of the Laplace Transform.

133
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 a
Example 1: Show that f (t ) = at → F ( s) =
s2
∞ ∞
First apply the definition of the Laplace transform: L[at ] = ∫ ate− st dt = a ∫ te− st dt
0 0

− e − st
Next, integrate by parts: ∫ ∫
u dv = uv − v du , so u = t ; du = dt ; v =
s
; dv = e − st dt

⎡ − st ∞ ∞ ⎤ ⎡ ∞⎤
−e− st ⎥ − e − st
Substituting, we obtain L[at ] = a ⎢ −te − ∫ dt = a ⎢(0 − 0) − ⎥
⎢ s s ⎥ ⎢ s2 ⎥
⎣ 0 0 ⎦ ⎣ 0 ⎦

a
The solution is thus L[at ] =
s2

Õ Õ Õ

1
Example 2: Show that f (t ) = e − at → F (s) =
s+a

Applying the definition of a Laplace transform and simplifying, we obtain

∞ ∞ ∞
1 − ( s + a )t
L[e− at ] = ∫ e− at e− st dt = ∫ e−( s + a )t dt = − e
s+a 0
0 0

1
The solution is thus L[e− at ] =
s+a

Õ Õ Õ

ω
Example 3: Show that f (t ) = sin ω t → F (s) =
s +ω2
2

First apply the definition of a Laplace transform:

∞
L[sin ω t ] = ∫ sin(ω t )e− st dt
0

e iωt − e −iωt
Next, recall that sin(ωt ) =
2i

134
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 ∞
Simplifying, we obtain L[sin ωt ] = ∫ 1 ⎡⎣e−( s −iω )t − e−( s+iω )t ⎤⎦ dt
2i
0

∞
1 ⎡ e − ( s −iω )t e− ( s +iω )t ⎤ 1⎡ 1 1 ⎤
= ⎢− + ⎥ = −
2i ⎢⎣ s − iω s + iω ⎥⎦ 2i ⎣ s − iω s + iω ⎥⎦
⎢
0

1 ⎡ s + iω − s + iω ⎤
=
2i ⎢⎣ s 2 + iω s − iω s + ω 2 ⎥⎦

ω
The solution is thus L[sin ω t ] =
s + ω2
2

Õ Õ Õ

14.2 Laplace Transform Properties

In the previous section, we derived the Laplace transforms for three specific time domain functions.
In this section, we explore Laplace transform properties and theorems, including the derivative and
integral functions.

Translation Property: Show that L[ f (t − θ )] = e−θ s F (s) .

We first apply the definition of a Laplace transform:

∞
L[ f (t − θ )] = ∫ f (t − θ )e− st dt
0

Note that e − st = e −θ s e − s (t −θ ) , and assume that f(t) = 0 for t < 0. We can then recast the above as

∞
L[ f (t − θ )] = e−θ s ∫ f (t − θ )e− s (t −θ ) d (t − θ )
θ
Next, we define τ = t − θ and substitute:
∞
L[ f (t − θ )] = L[ f (τ )] = e ∫ f (τ )e− sτ dτ
−θ s

We can then see that this is the definition of the Laplace transform, so we conclude that

L[ f (t − θ )] = e−θ s F ( s)

The translation property is useful for modeling the time shift associated with dead time.

135
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
Linearity Property: Show that L [ a1 f1 (t ) + a2 f 2 (t )] = a1L [ f1 (t )] + a2L [ f2 (t )]

Applying the Laplace transform definition, we obtain

∞
L [ a1 f1 (t ) + a2 f 2 (t )] = ∫ [ a1 f1 (t ) + a2 f 2 (t )] e− st dt
0

∞ ∞

∫ ∫
= a1 f1 (t )e dt + a2 f 2 (t )e− st dt
− st

0 0

= a1L [ f1 (t ) ] + a2L [ f 2 (t )]

We can now conclude that L [ a1 f1 (t ) + a2 f 2 (t )] = a1L [ f1 (t )] + a2L [ f2 (t )]

The linearity property states that the Laplace of a sum of functions equals the sum of the Laplace of
the individual functions:

First Derivative: Show that L ⎡⎢ df (t ) ⎤⎥ = sF ( s) − f (0) where f (0) is in the time domain.
⎣ dt ⎦

First, we apply the Laplace transform definition:

∞
L ⎡⎢ df (t ) ⎤⎥ = df (t ) e− st dt
∫
⎣ dt ⎦ 0 dt
df (t )
Next, we integrate by parts: ∫ u dv = uv −∫ v du , so u = e − st ; du = − se− st dt ; v = f (t ); dv =
dt
dt

∞
L ⎡⎢ df (t ) ⎤⎥ = f (t )e− st
∞
Substituting, we obtain
⎣ dt ⎦ 0 ∫
+ sf (t )e− st dt
0

∫
= [0 − f (0)] + s f (t )e − st dt
0

Recognizing that the last term is the definition of the Laplace Transform, we find that

L ⎡⎢ df (t ) ⎤⎥ = sF (s) − f (0)
⎣ dt ⎦

136
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 ⎡t ⎤ 1
∫
Integration: Show that L ⎢ f (t )dt ⎥ = F ( s ) .
⎣⎢ 0 ⎦⎥ s

Applying the Laplace transform definition, we obtain

⎡t ⎤ ∞
⎡t ⎤
L ⎢ ∫ f (t )dt ⎥ = ∫ ⎢ ∫ f (t )dt ⎥ e− st dt
⎢⎣ 0 ⎥⎦ 0 ⎢⎣ 0 ⎥⎦
t
1
We must then integrate by parts: u= ∫ f (t )dt; du = f (t )dt ; v = − e − st ;
s
dv = e − st dt
0

∞
⎡t ⎤ − st t ∞
Substituting yields L ⎢ f (t )dt ⎥ = − e
∫ ∫ f (t )dt +
1
∫ f (t )e
− st
dt
⎣⎢ 0 ⎦⎥ s 0
s 0
0

Recognizing that the last term is the definition of the Laplace Transform, then we obtain

⎡t ⎤ 1
L ⎢ ∫ f (t )dt ⎥ = F (s)
⎢⎣ 0 ⎥⎦ s

Final value theorem: lim f (t ) = lim sF ( s )

t →∞ s→0

Initial value theorem: lim f (t ) = lim sF ( s )

t →0 s→∞

s+4
Example: What is the final value (limit as t → ∞) of F ( s ) = ?
s ( s + 1)( s + 2)( s + 3)

Applying the final value theorem, we obtain

lim f (t ) = lim sF ( s )
t →∞ s→0

= lim ⎡ s s+4 ⎤
s → 0 ⎣ s ( s + 1)( s + 2)( s + 3) ⎥⎦
⎢

4
=
6

Õ Õ Õ

137
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
14.3 Moving Time Domain ODEs into the Laplace Domain
For the process control analyses we explore later, the procedure we follow will be to move our time
domain equations into the Laplace domain, combine and manipulate them using algebra (which is
why we bother changing them into the Laplace domain), and then move the result back into the time
domain for implementation. Hence, we must become comfortable with moving time domain
equations into and out of the Laplace domain.
We solved the ODEs below in the time domain in chapter 13. While these examples all
demonstrate moving from the time domain into the Laplace domain, following the steps in reverse
will achieve the opposite. When working through these examples, note that the boundary (initial)
conditions are applied early in the transformation process. Recall that applying the boundary
conditions is one of the last steps when solving ODEs in the time domain.

Example 1: Move the following time domain equation to the Laplace domain:

d y (t )
+ 2 y (t ) = 4t where y(0) = 1
dt

Consulting the Laplace transform table in Appendix B, we see that

L ⎡⎢ dy(t ) ⎤⎥ + 2L[y(t )] = L[4t ]

⎣ dt ⎦

4
Applying initial conditions, [sY ( s ) - y (0)]+ 2Y (s ) =
s2

4 4 + s2
and simplifying, we obtain Y ( s )[s + 2] = +1=
s2 s2

The equation can now be written in the Laplace domain as

4 + s2
Y (s) =
s 2 ( s + 2)

Õ Õ Õ

Example 2: Move the following time domain equation into the Laplace domain:

dx(t )
= x(t ) + sin t where x(0) = 0
dt

Consulting the Laplace transform table in Appendix B, we find that

L ⎡⎢ dx(t ) ⎤⎥ − L[x(t )] = L[sin t ]

⎣ dt ⎦

138
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 1
Applying initial conditions, [ sX ( s ) − x(0)] − X ( s ) = 2
s +1
we can now write the equation in the Laplace domain as

1
X ( s) = 2
( s + 1)( s − 1)

Õ Õ Õ

Example 3: Move the following time domain equation into the Laplace domain:

d 2 y (t ) dy (t ) dy (t )
2
+4 + 3 y (t ) = t where y (0) = =0
dt dt dt t =0

Consulting the Laplace transform table, we see that

⎡ y (t ) ⎤
2
⎡ dy (t ) ⎤
L ⎢d 2 ⎥
+ 4L ⎢ ⎥ +3L[y (t )] = L[t ]
⎣⎢ dt ⎦⎥ ⎣ dt ⎦

Applying initial conditions yields

⎡ 2 dy (t ) ⎤ 1
⎢ s Y ( s ) sy (0) ⎥ + [4 sY ( s ) - 4 y (0)] + 3Y ( s ) = 2
⎣ dt t =0 ⎦ s

1
Simplifying, we obtain Y ( s )[ s 2 + 4 s + 3] =
s2

The equation can now be written in the Laplace domain as

1 1
Y (s) = 2 2
= 2
s ( s + 4s + 3) s ( s + 3)( s + 1)

Õ Õ Õ

Example 4: Move the following time domain equation into the Laplace domain:

d 2 y (t ) dy (t ) dy (t )
2
+2 + y (t ) = t 2 where y (0) = 0 and =1
dt dt dt t =0

Consulting the Laplace transform table in Appendix B, we see that

⎡ 2 y (t ) ⎤ ⎡ dy (t ) ⎤
L ⎢d 2
⎥ + 2L ⎢ ⎥ + L [ y(t)] = L[t 2 ]
⎢⎣ dt ⎥⎦ ⎣ dt ⎦

139
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
Applying initial conditions yields

⎡ 2 dy (t ) ⎤ 2
⎢ s Y ( s ) − sy (0) − ⎥ + [2 sY ( s ) - 2 y (0)] + Y ( s ) = 3
⎣ dt t =0 ⎦ s

2 2 + s3
Simplifying, we obtain Y ( s )[ s 2 + 2s + 1] = +1=
s3 s3
The equation can now be written in the Laplace domain as

2 + s3 2 + s3
Y ( s )= =
s 3 ( s 2 + 2s + 1) s 3 ( s + 1) 2

Õ Õ Õ

Example 5: Move the following time domain equation into the Laplace domain:

d 2 y (t ) dy (t ) dy (t )
2 2
+2 + y (t ) = 2 where y (0) = =0
dt dt dt t =0

Consulting the Laplace transform table in Appendix B, we can see that

⎡ d 2 y (t ) ⎤ ⎡ dy (t ) ⎤
2L ⎢ 2 ⎥
+ 2L ⎢ ⎥ + L[y (t )] = L[2]
⎣⎢ dt ⎦⎥ ⎣ dt ⎦

Applying initial conditions yields

⎡ 2 dy (t ) ⎤ 2
⎢2 s Y ( s ) − 2 sy (0) − 2 ⎥ + [2 sY ( s ) − 2 y (0)] + Y ( s ) =
⎣ dt t =0 ⎦ s

2 1
Simplifying, we obtain Y (s ) = =
2
s (2s + 2 s + 1) ⎛ 1⎞
s ⎜ s2 + s + ⎟
⎝ 2⎠

The equation can now be written in the Laplace domain as

2 1
Y (s ) = =
s (2s + 2 s + 1) s ⎡⎣ s − ( − 2 + 2 i ) ⎤⎦ ⎡⎣ s − ( − 12 − 12 i ) ⎤⎦
2 1 1

Õ Õ Õ

Example 6: Move the following time domain equation into the Laplace domain:

d 2 x(t ) dx(t )
+ x(t ) = sin ( 2t ) where x(0) = 0 and =1
dt 2 dt t =0

140
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Consulting the Laplace transform table in Appendix B, we see that

⎡ x(t ) ⎤
2
L ⎢d 2 ⎥
+ L[x(t )] = L[sin(2t )]
⎢⎣ dt ⎦⎥

Applying initial conditions yields

⎡ 2 dx(t ) ⎤ 2
⎢ s X ( s )−sx(0)− ⎥ + X (s) = 2
⎣ dt t =0 ⎦ s +4

Simplifying, we obtain ( )
X ( s) s 2 + 1 − 1 = 2
2
s +4

The equation can now be written in the Laplace domain as

s2 + 6
X ( s) =
( s 2 + 4)( s 2 + 1)

Õ Õ Õ

14.4 Moving Laplace Domain ODEs into the Time Domain

Now that we have some understanding of Laplace transforms, let’s consider process oriented
challenges.

1
Example 1: A process response to a unit step forcing function is Y ( s ) = . What is the
s (3s + 1)
original process ODE in the time domain?

Solution: From the Laplace table (Appendix B), we know that

L[unit step] = 1
s
Applying this, we can see that our original equation is actually

1 ⎡1 ⎤
Y (s) =
(3s + 1) ⎢⎣ s ⎥⎦

1
Simplifying, we find that Y ( s ) [3s + 1] =
s

1
so 3sY ( s ) + Y ( s ) =
s

1
Assuming y(0) = 0, we obtain 3[sY ( s ) − y (0)] + Y ( s ) =
s

141
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
From Appendix B, we can see that the inverse Laplace, L -1 , is
dy (t )
3 + y (t ) = 1 where y(0) = 0
dt

Õ Õ Õ

KP
Example 2: A process response to a forcing function U(s) is Y ( s ) = U ( s ) . What is the
τ Ps +1
original process ODE in the time domain?

Solution: First, we rearrange the equation as such:

Y ( s )[τ P s + 1] = K PU ( s)

Y ( s )[τ P s ] + Y ( s) = K PU ( s)

Next, we assume that y(0) = 0: τ P [sY ( s ) − y (0)] + Y ( s ) = K PU ( s )

From the table (Appendix B), we can see that the inverse Laplace transform, L -1 , can be written
dy (t )
τP + y (t ) = K P u (t ) where y(0) = 0
dt

Y ( s) KP
With this result, we note that = is called the transfer function for the first order
U ( s) τ P s + 1
without dead time ODE.
Õ Õ Õ

Kp
Example 3: A process response to a forcing function U(s) is Y ( s ) = U (s) .
τ n2 s 2 + 2τ nξ s + 1
What is the original process ODE in the time domain?

Solution: We rearrange the equation as follows:

( )
Y ( s ) τ n2 s 2 + 2τ nξ s + 1 = K pU ( s )

Y ( s )(τ s ) + Y ( s )(2τ ξ s ) + Y ( s ) = K
2 2
n n pU ( s )

dy (t )
Next, we assume that y(0) = 0 and = 0:
dt t =0

142
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 ⎛ dy (t ) ⎞
Y ( s )⎜⎜τ n2 s 2 − sy (0) − ⎟ + 2τ nξ (sY ( s ) − y (0) ) + Y ( s ) = K pU ( s )
dt ⎟
⎝ t =0 ⎠

From Appendix B, we can see that the inverse Laplace transform, L -1 , can be written as
d 2 y (t ) dy (t ) dy (t )
τ n2 + 2τ nξ + y (t ) = K p u (t ) where y (0) = 0 and =0
dt dt dt t =0

Y ( s) Kp
Similar to the previous example, we note that = 2 2 is the transfer function for
U ( s ) τ n s + 2τ nξ s + 1
the second order without dead time ODE.
Õ Õ Õ

Y (s) 3
Example 4: Consider the following system: = 2 . Is this system stable? Does it have a
U (s) s + s + 1
natural tendency to oscillate?

Solution: Comparing to the general form of Example 3, we observe that KP = 3; τn2 = 1; 2τnξ =
1; thus, τn = 1; and ξ = 0.5. Since 0 < ξ < 1, the process is stable and underdamped (which
means it does have a natural tendency to oscillate).

Õ Õ Õ

14.5 Exercises

Q-14.1 Showing all steps, derive the Laplace transform of cos(ωt).

Q-14.2 Showing all steps, derive the Laplace transform of eat.

143
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 15. Transfer Functions
15.1 Process Transfer Functions
A process transfer function, GP(s), is an equation in the Laplace domain that describes the dynamic
response of the measured process variable to changes in the manipulated process variable (controller
output signal).

Time Domain: Consider the general time domain ODE describing a second order process:

d 2 y (t ) dy (t ) dy (t )
τ n2 2
+ 2τ nξ + y (t ) = K P u (t − θ P ) where y(0) = 0, =0 (15.1)
dt dt dt t =0

As we learned in Chapter 13, the characteristic equation is

τ n2 s 2 + 2τ nξ s + 1 = ( s − p1 )( s − p2 ) = 0

Also, the total solution is y (t ) = C1e p1t + C2 e p2 t + yP (t )

p1 t
or y (t ) = (C1 + C2t )e + yP (t )

We can express this as y (t ) = e (real part) t [C R cos(imag part t ) − C I sin(imag part t )] + y P (t )

Laplace Domain: As detailed in example 2 below, the transfer function for the ODE of Eq. 15.1 is

Y (s) K e−θ P s K P e−θ P s

GP ( s ) = = 2 2 P =
U ( s ) τ n s + 2τ nξ s + 1 ( s − p1 )( s − p2 )

Note that the denominator of a transfer function is the characteristic equation of the time domain
complementary solution. This is true for all transfer functions, including process transfer functions
as shown below.

Recall that the roots of the characteristic equation indicate a system’s stability and natural
tendency to oscillate. An unstable system results if any root has a positive real part, e + p1t , because
that term will grow without bound as time t grows to infinity. A stable system results if all roots have
negative real parts, e − p1t , as these terms all die out (go to zero) as t grows to infinity. The tendency to
oscillate is a consequence of sine and cosine terms in the solution that results from imaginary roots.
Because the denominator of a transfer function is the characteristic equation of the time
domain complementary solution, then the roots of the denominator of a transfer function (called
poles) similarly indicate system stability and tendency to oscillate. This knowledge will prove useful
in control system analysis and design studies in the following chapters.

144
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
Example 1: Derive the following FOPDT (first order plus dead time) transfer function:

dy (t )
τp + y (t ) = K p u (t − θ P ) where y(0) = 0 (15.2)
dt

Solution: Consulting the Laplace table (Appendix B), we see that

τ p [ sY ( s) − y (0)] + Y ( s ) = K p e−θ P U ( s )

Applying the initial condition yields

(τ p s + 1)Y ( s) = K P e−θ P U ( s )

We can then rearrange to form the FOPDT transfer function:

−θ s
Y (s) K P e P
GP ( s ) = =
U (s) τ Ps +1

Recall that the complementary equation for Eq. 15.2 is

dy (t )
τp + y (t ) = 0
dt

so the characteristic equation is τ p s + 1 = 0

This confirms the observation above that the denominator of a transfer function is the
characteristic equation of the time domain complementary solution.
Õ Õ Õ

Example 2: Derive the SOPDT (second order plus dead time) transfer function:

d 2 y (t ) dy (t ) dy (t )
τ n2 2
+ 2τ nξ + y (t ) = K P u (t − θ P ) where y(0) = 0, = 0 (15.3)
dt dt dt t =0

Answer: Consulting the Laplace table, we can see that

⎡ dy (0) ⎤
τ n2 ⎢ s 2Y ( s ) − sy (0) − ⎥ + 2τ nξ [ sY ( s ) − y (0)] + Y ( s ) = K P e −θ P sU ( s )
⎣ dt ⎦

Applying initial conditions yields τ n2 s 2Y ( s) + 2τ nξ sY ( s ) + Y ( s ) = K P e−θ P sU ( s)

We can then rearrange to form the SOPDT transfer function:

145
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Y (s) K e−θ P s
GP ( s ) = = 2 2 P
U ( s ) τ n s + 2τ nξ s + 1

Recall that the complementary equation for Eq. 15.3 is

d 2 y (t ) dy (t )
τ n2 2
+ 2τ nξ + y (t ) = 0
dt dt

so the characteristic equation is τ n2 s 2 + 2τ nξ s + 1 = 0

This again confirms the observation that the denominator of a transfer function is the
characteristic equation of the complementary solution.
Õ Õ Õ

15.2 Controller Transfer Functions

Analogous to a process transfer functions, control algorithm transfer functions, GC(t), are Laplace
domain equations that describe the dynamic behavior of the controller output signal as it responds to
changes in the controller error. The general form of a controller transfer function is

U (s)
GC ( s ) = (15.4)
E (s)
For all derivations, we observe that at the design level of operation, the control error should
be zero. Hence, when u(t) = ubias, then e(t) = 0. Consistent with this observation, we use the following
perturbation variable definitions:

u P (t ) = u (t ) − ubias (15.5a)

e P (t ) = e(t ) − 0 (15.5b)

Example: P-Only Control

The P-Only control algorithm is u (t ) = ubias + KC e(t )

Rearranging to group like terms gives us

u (t ) − ubias = KC [e(t ) − 0]

Employing Eq. 15.5 yields u P (t ) = KC e P (t )

Using the Laplace table, we can move into the Laplace domain:

U ( s ) = KC E ( s)

146
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
This gives us the P-Only transfer function:

U (s)
GC ( s ) = = KC (15.6)
E (s)

Õ Õ Õ

Example: PI Control
t
KC
The PI control algorithm is u (t ) = ubias + KC e(t ) +
τI ∫ e(t )dt
0

Rearranging to group like terms gives us

t
KC
u (t ) − ubias = KC [e(t ) − 0] +
τI ∫ [e(t ) − 0]dt
0
t
KC
u P (t ) = KC e P (t ) + ∫e
P
Employing Eq. 15.5 yields (t ) dt
τI 0

We can then use the Laplace tables and assume the process starts at steady state, to give us

KC ⎛ 1 ⎞
U ( s ) = KC E ( s ) + ⎜ E ( s) ⎟
τI ⎝ s ⎠

This results in the PI transfer function:

U ( s) ⎛ 1 ⎞
GC ( s ) = = KC ⎜1 + ⎟ (15.7)
E (s) ⎝ τIs ⎠

Õ Õ Õ

Example: PID Control

The PID w/ derivative on error control algorithm is

t
KC de(t )
u (t ) = ubias + KC e(t ) +
τI ∫ e(t )dt + K τ
0
C D
dt

Rearranging to group like terms gives us

t
KC d [e(t ) − 0]
u (t ) − ubias = KC [e(t ) − 0] +
τI ∫ [e(t ) − 0]dt + K τ
0
C D
dt

147
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 t
KC de P (t )
∫ e p (t ) dt + KCτ p
p p
Employing Eq. 15.5 yields u (t ) = KC e (t ) +
τI 0
dt

We then consult the Laplace tables and assume the process starts at steady state to give us

KC 1
U ( s ) = KC E ( s ) + E ( s ) + KCτ D sE ( s )
τI s
This results in the PID transfer function:

U (s) ⎛ 1 ⎞
GC ( s ) = = KC ⎜1 + +τ Ds ⎟ (15.8)
E ( s) ⎝ τIs ⎠

Õ Õ Õ

15.3 Poles of a Transfer Function and Root Locus

The roots of the denominator of a transfer function are traditionally called the poles of the transfer
function. Because the Laplace domain is defined in the complex plane (i.e. s = a + bi), a pole has a
real part and an imaginary part. A root locus plot shows the location of a system’s roots/poles on the
complex plane.

Example 1: Locate these points on the s plane: A = 1 + i; B = −2i; C = −2 −i

Solution:
Im
3

A
Re
–3 3
C
B
–3

s plane
A negative real part of the root means the term e− p1t goes to zero as time increases to infinity.
Thus, the dynamics will die out and the system will be stable. All of the roots must have negative
real parts for this to hold true. If even one root has a positive real part, it will eventually dominate
and push the system to instability.
Real parts are negative on the left-hand side of the s plane and they are positive on the right-
hand side. Consequently, all roots/poles describing a stable system lie in the left-hand side of the
complex plane.

148
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Im

Re

stable unstable

s plane

Figure 15.1 – Stable and unstable sides of the complex plane

Õ Õ Õ

Example 2: What do the poles of the following transfer function indicate about the system
behavior?

1
G sys ( s ) =
( s + 1)( s + 2.5)

Solution: The roots/poles are p1 = −1.0 and p2 = −2.5 as plotted on the s plane below:

Im
3

p2 p1
Re
–3 3

–3

The plot indicates that the system is stable and does not oscillate because the poles are real and
negative.
Õ Õ Õ

Example 3: What do the poles of the following transfer function indicate about the system
behavior?

1
G sys ( s ) =
( s − 2)( s + 2.5)

149
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Solution: The roots/poles are p1 = +2.0 and p2 = −2.5 as plotted on the s plane below:

Im
3

p2 p1
Re
–3 3

–3

The plot indicates that the system is unstable because one of the poles is positive and hence will
grow without bound as time passes. The system will not oscillate as it goes unstable because
neither pole has imaginary parts.
Õ Õ Õ

Example 4: What do the poles of the following transfer function indicate about the system
behavior?

1
G sys ( s ) = 2
( s + 4s + 5)

Solution: The roots/poles are p1 = −2 + i and p2 = −2 − i as plotted on the s plane below:

Im
3

p1

Re
–3 3
p2

–3

Because the real part of both poles is negative, the system is stable. The system will have a
natural tendency to oscillate due to the imaginary part of the roots/poles.

Õ Õ Õ

15.4 Poles as Complex Conjugates

If the roots/poles have an imaginary component, they will always be present as complex conjugate
pairs of the form a ± bi. That is, the pair will have the same real part and reflecting imaginary parts.

150
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 The property most appealing about the complex conjugate form is that when the pairs are
added together or multiplied together, they yield real numbers with no imaginary parts. This is
important because the processes we work on are real, and the solutions we devise must also be real
(would you want your boss to give you an imaginary raise?).

Example 5: Show that 1 + 2i and 1 − 2i are complex conjugates.

Answer: Beyond the fact that they have the proper a ± bi form, when we add and multiply them
together, they yield real numbers with no imaginary parts:

(1 + 2i) + (1 − 2i) = 2 and (1 + 2i)(1 − 2i) = (1 + 2i − 2i − 4i2) = 5

Õ Õ Õ

15.5 Poles of the Transfer Function Indicate System Behavior

As illustrated in Fig 15.2, as the roots/poles move farther out on the negative real axis, the dynamics
become faster. As the roots/poles move farther out on the imaginary axis, the oscillatory nature of the
response increases.

Im
increasing
oscillation

faster Re
dynamics

stable unstable

Figure 15.2 –Pole location indicates system behavior

To show that the roots/poles are related to system behavior, we can perform an analysis similar to the
one done in the time domain in Chapter 13. Here we consider the same four cases presented earlier in
the time domain.

151
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Im
case 4

case 3 + 1/τn
ξ=0
case 4
y(t)
case 2
yp

Re
0<ξ<1 case 1
case 3
ξ>1
ξ=1 case 1
case 2 – 1/τn

Time, t
s plane

Time Domain Laplace Domain

Figure 15.3 - Stable behaviors of underdamped system based on value of ξ

−ξ ± ξ 2 − 1
Case 1: ξ > 1 (overdamped) p1 , p2 = ; y (t ) = y p (t ) + C1e− p1t + C2 e− p2t
τn
KP
Transfer function: G ( s) =
( s + p1 )( s + p 2 )

The overdamped response has distinct roots/poles on the negative real axis. The time domain
response becomes slower and slower as ξ gets larger, but it is always stable and never oscillates.

1
Case 2: ξ = 1 (critically damped) p1 , p 2 = − ; y (t ) = y p (t ) + (C1 + C2t )e−t τ n
τn

Kp
Transfer function: G(s) =
1 1
(s + )( s + )
τn τn

When critically damped, the roots/poles are repeated on the negative real axis. There is no oscillation
in the response because there is no imaginary part leading to sine and cosine terms in the roots/poles.

152
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 −ξ ± i 1−ξ 2
Case 3: 0 < ξ < 1 (underdamped) p1 , p 2 = ;
τn
⎛ 1−ξ 2 1−ξ 2 ⎞
−ξ t / τ n ⎜
y (t ) = y p (t ) + e ⎜⎜ C R cos( t ) − C I sin( t ) ⎟⎟
τn τn ⎟
⎝ ⎠

KP
Transfer function: G ( s) =
⎡ ⎛ 2 ⎞⎤ ⎡ ⎛ ξ 1−ξ 2 ⎞⎤
⎢s − ⎜ − ξ + 1 − ξ ⎟ ⎜
i ⎟⎥ ⎢ s − ⎜ − −
⎟
i ⎟⎥
⎢ ⎜⎜ τ n τn ⎟⎥ ⎢ ⎜ τ n τn ⎟⎥
⎣ ⎝ ⎠⎦ ⎣ ⎝ ⎠⎦

As ξ varies in the continuum from zero to one, the roots/poles move from the imaginary axis to the
real axis. The roots/poles are always in complex conjugate pairs of the form a ± bi. The time domain
response shows oscillation that dampens. The closer ξ gets to 1, the greater the damping.

i ⎛ t t ⎞
Case 4: ξ = 0 (undamped) p1 , p 2 = ± ; y (t ) = y p (t ) + ⎜⎜ C R cos( ) − C I sin( ) ⎟⎟
τn ⎝ τn τn ⎠

Y (s) Kp
Transfer function: G ( s) = =
U (s) ⎛ i ⎞⎛ i ⎞
⎜ s − ⎟⎜ s + ⎟
⎜ τ ⎟⎜ τ ⎟
⎝ n ⎠⎝ n ⎠

At the limit of stability, the roots/poles are located on the imaginary axis and have no real part. The
time domain response shows oscillation that neither grows nor dies out. Thus, the oscillations
maintain constant form and continue forever.

Case 5: ξ < 0 (unstable)

All values of the damping factor, ξ, that are less than zero yield a solution that has a positive real part.
Hence, as time passes, e + t will grow without bound, an unstable result. All roots/poles will fall on the
right-hand side of the s-plane.

15.6 Exercises

Q-15.1 What are the steady state gain and time constants for a process described by the following
transfer function (pay careful attention to the form of the equation)?
7
G P (s) =
( s + 10)( s + 2)

153
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
Q-15.2 Starting with the Laplace domain transfer function, show all steps and derive the time domain
ODE for the Second Order Plus Dead Time Integrating form:

Y ( s ) K p e −θ P s
Laplace: GP ( s ) = =
U ( s ) s (τ p s + 1)

d 2 y (t ) d y (t )
Time: τp + = K p u (t − θ P )
d t2 dt

Q-15.3 Starting with the time domain ODE, show all steps and derive the Laplace domain transfer
function for the Second Order Plus Dead Time Overdamped with Lead Time form:

d 2 y (t ) d y (t ) d u (t − θ P )
Time: τ 1τ 2 2
+ (τ 1 +τ 2)
dt
[
+ y (t ) = K p u (t − θ p ) + τ L
dt
]
dt

Y ( s ) K p (τ L s +1) e − θ P s
Laplace: G P (s) = =
U (s) (τ 1s +1)(τ 2 s +1)

Q-15.4 Starting with the continuous position form of the Integral Only controller, show all steps to
derive the I-Only transfer function.

Q-15.5 The dynamics of a process are described by the following ODE, which is in perturbation
variable form:

d 2 y (t ) dy (t )
6 +5 + 4 y (t ) = 3 u (t )
2 dt
dt

a) Showing all steps, determine the transfer function, GP(s), for this process.

b) Determine the poles of the transfer function.

c) Plot the poles on the complex plane.

d) Based on the pole locations, describe the general dynamic behavior of this process.
Be sure to briefly explain your reasoning.

154
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 16. Block Diagrams

16.1 Combining Transfer Functions Using Block Diagrams

Laplace domain transfer functions enable us to manipulate complex equations using simple algebra.
For example, consider two non-interacting draining tanks. If we assume that drain flow rate is
proportional to hydrostatic head, we can write the equations using perturbation variables:

dh1P (t )
The tank 1 ODE is AC1 + α 1 h1P (t ) = F0P (t ) where h1P (0) = 0 (16.1)
dt

In the Laplace domain, Eq. 16.1 becomes

AC1sH1 ( s ) + α1 H1 ( s ) = F0 ( s ) (16.2)

H 1 ( s) 1 / α1
The tank 1 transfer function is thus G P1 ( s ) = = (16.3)
F0 ( s ) ( AC1 / α 1 )s + 1

dh2P (t )
The tank 2 ODE is AC 2 + α 2 h2P (t ) = α 1 h1P (t ) where h2P (0) = 0 (16.4)
dt

In the Laplace domain, Eq. 16.4 becomes

AC 2 sH 2 ( s ) + α 2 H 2 ( s ) = H1 ( s ) (16.5)

H 2 ( s) α1 / α 2
The tank 2 transfer function is thus G P 2 ( s ) = = (16.6)
H 1 ( s ) ( AC 2 / α 2 )s + 1

With the above as a basis, we write general coupled process ODEs as

dy1 (t ) Y1 ( s ) K P1
Process 1: τ P1 + y1 (t ) = K P1u (t ) ; y1 (0) = 0 ⇒ G P1 ( s ) = = (16.7)
dt U ( s) τ P1 s + 1

dy 2 (t ) Y2 ( s ) K P2
Process 2: τ P 2 + y 2 (t ) = K P 2 y1 (t ) ; y 2 (0) = 0 ⇒ G P 2 ( s) = = (16.8)
dt Y1 ( s ) τ P 2 s + 1

Combining the time domain ODE’s of Eqs. 16.7 and 16.8 into a single second-order differential
equation describing how y2(t) responds to changes in u(t) requires manipulation of ODEs as follows:

dy2 (t )
τ P2 + y2 (t )
Solve Eq. 16.8 for y1(t): y1 (t ) = dt (16.9)
KP2

155
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 d 2 y2 (t ) dy2 (t )
τ P2 +
dy1 (t ) dt 2 dt
Take the derivative of Eq. 16.9: = (16.10)
dt KP2

Substitute Eqs. 16.9 and 16.10 into Eq. 16.7:

d 2 y2 (t ) dy2 (t ) dy2 (t )
τ P2 + τ + y2 (t )
τ P1 dt 2 dt + P 2 dt = K P1 u (t ) (16.11)
KP2 KP2

Multiply both sides by KP2 and combine like terms:

d 2 y 2 (t ) dy 2 (t )
τ P1τ P 2 2
+ (τ P1 + τ P 2 ) + y 2 (t ) = K P1 K P 2 u (t ) (16.12)
dt dt

In the Laplace domain, we combine the transfer functions using simple algebra, which is the reason
for converting from the time domain into the Laplace domain and back:

Y1 ( s ) Y2 ( s ) Y2 ( s ) ⎛ K P1 ⎞⎛ K P 2 ⎞
Gsystem ( s ) = GP1 ( s )GP 2 ( s ) = = =⎜ ⎟⎜ ⎟ (16.13)
U ( s ) Y1 ( s ) U ( s ) ⎝ τ P1s + 1 ⎠⎝ τ P 2 s + 1 ⎠

and thus
Y2 ( s ) K P1 K P 2
= (16.14)
U ( s ) τ P1τ P 2 s 2 + (τ P1 + τ P 2 ) s + 1

As we expect, converting Eq. 16.14 back to the time domain yields Eq. 16.12. This comparison of
time domain versus Laplace domain equation manipulation helps demonstrate the benefit of using the
Laplace domain in our subsequent analyses.
A block diagram is a convenient way to visualize the combination and manipulation of
Laplace domain equations. As shown in Fig. 16.1, we use a summer block (a circle) to add block
inputs and a multiplier block (a square) to multiply block inputs:

Summer:

Multiplier:

Figure 16.1 – Blocks used to create a Laplace domain block diagram

156
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
Example 1: Show this manipulation using block diagrams: Y(s) = A(s) − B(s) − C(s)

Solution: Below are two of several possibilities:

A(s)

Y(s) +
A(s) Y(s)
B(s)
+ + –
– – –

B(s) C(s) C(s)

[A(s) − B(s)] − C(s) = Y(s) A(s) − B(s) − C(s) = Y(s)

Õ Õ Õ

Example 2: Show this manipulation using block diagrams: Y(s) = A(s)G1(s)G2(s)

Solution: Below are two of several possibilities:

A(s) B(s) Y(s) A(s) C(s) Y(s)

G1(s) G2(s) G2(s) G1(s)

Y(s) = B(s)G2(s) Y(s) = C(s)G1(s)

B(s) = A(s)G1(s) C(s) = A(s)G2(s)
Y(s) = A(s)G1(s)G2(s) Y(s) = A(s)G1(s)G2(s)

Õ Õ Õ

Example 3: Show using block diagrams: Y(s) = [A(s) − B(s)]G(s)

Solution: Below are two of several possibilities:

A(s)
G(s)
A(s) Y(s) +
Y(s)
G(s)
+
– –
B(s)
G(s)
B(s)

Y(s) = [A(s) − B(s)]G(s) Y(s) = A(s)G(s) − B(s)G(s) = [A(s) − B(s)]G(s)

Õ Õ Õ

157
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 KP 1
Example 4: Show using block diagrams: Y ( s ) = U (s) + M (s)
τ Ps +1 τ Ps +1

Solution: Below is one of several possibilities:

U(s)
KP
+ Y(s)
1
τ Ps +1
M(s) +

⎡ 1 ⎤ KP 1
Y ( s ) = [U ( s ) K P + M ( s )] ⎢ ⎥ = U ( s) + M ( s)
⎣τ P s + 1 ⎦ τ Ps +1 τ Ps +1

Õ Õ Õ

16.2 The Closed Loop Block Diagram

As shown in Fig. 16.2, the closed loop block diagram in the time domain is

d(t)
Disturbance

ysp(t) e(t) Final + y(t)

u(t) m(t)
Controller Control Process
+ +
– Element

ym(t) y(t)
Measurement
Sensor/Transmitter

Figure 16.2 – Closed Loop Block Diagram in Time Domain

In the Laplace domain, the closed loop block diagram is as shown in Fig. 16.3:

158
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 D(s)
GD(s)
Disturbance

Ysp(s) E(s) U(s) +

M(s) Y(s)
GC(s) GF(s) GP(s)
+ +
–
Controller Process
Final
Control
Element

YM(s) Y(s)
GM(s)

Measurement
Sensor/Transmitter

Figure 16.3 – Closed Loop Block Diagram in Laplace Domain

Notice that it is not only the process and controller that have transfer functions describing their
dynamic behavior. As shown in the block diagram of Fig. 16.3, final control element (e.g. valve,
pump) and measurement sensor also have transfer functions:

U (s) M ( s) YM ( s )
GC ( s) = G F ( s) = G M (s) = (16.15)
E (s) U ( s) Y ( s)
controller final control element measurement sensor

The block diagram shows that:

E(s) = Ysp(s) − YM(s) (16.16)

The block diagram also shows that the transfer function for the measured process variable is
somewhat more complicated when a disturbance variable is included. Following the block diagram
rules presented above, the transfer function is:

Y(s) = M(s)GP(s) + D(s)GD(s) (16.17)

16.3 Closed Loop Block Diagram Analysis

Building on the principles discussed earlier in this chapter, we can write a series of equations as we
step around the closed loop block diagram in an orderly fashion. A convenient place to start in the
balance is with process variable Y(s) as it exits the block diagram on the right. The equations thus
develop as:

159
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Y(s) = M(s)GP(s) + D(s)GD(s) (16.18a)
M(s) = U(s)GF(s) (16.18b)
U(s) = E(s)GC(s) = [Ysp(s) − YM(s)]GC(s) (16.18c)
YM(s) = Y(s)GM(s) (16.18d)

Substituting Eq. 16.18b into Eq. 16.18a, and Eq. 16.18d into Eq. 16.18c yields:

Y(s) = U(s)GF(s)GP(s) + D(s)GD(s) (16.19a)

U(s) = [Ysp(s) − Y(s)GM(s)]GC(s) (16.19b)

Substituting Eq. 16.19b into Eq. 16.19a yields:

Y(s) = [Ysp(s) − Y(s)GM(s)]GC(s)GF(s)GP(s) + D(s)GD(s) (16.20)

= Ysp(s)GC(s)GF(s)GP(s) − Y(s)GM(s)GC(s)GF(s)GP(s) + D(s)GD(s) (16.21)

Rearrange to obtain
Y(s)[1+ GM(s)GC(s)GF(s)GP(s)] = Ysp(s)GC(s)GF(s)GP(s) + D(s)GD(s) (16.22)

Combining these equations and solving for Y(s) produces the following closed loop Laplace equation:

GC ( s )G F ( s )G P ( s ) G D ( s)
Y ( s) = Ysp ( s ) + D(s) (16.23)
1 + GC ( s )G F ( s )G P ( s )G M ( s ) 1 + GC ( s )G F ( s )G P ( s )G M ( s )

Here we realize that a complex transfer function can be constructed from a combination of simpler
transfer functions. As this analysis reveals, the closed loop transfer functions are

Process Variable to Set Point (when disturbance is constant):

Y ( s) GC ( s )G F ( s )G P ( s )
=
Ysp ( s ) 1 + GC ( s )G F ( s )G P ( s )G M ( s )

Process Variable to Disturbance (when set point is constant):

Y (s) G D ( s)
=
D( s ) 1 + GC ( s )G F ( s )G P ( s )G M ( s )

With the controller in automatic (closed loop), if the dynamics are disturbance driven or set point
driven, the characteristic equation that reveals the inherent dynamic character of the system is the
denominator of the transfer function, which in this case is

1 + GC ( s )G F ( s )G P ( s )G M ( s ) = 0 (16.24)

Recall that the roots of the characteristic equation (the poles of the transfer function) indicate whether
or not a system is stable and the degree to which it has tendency to oscillate. The analysis above

160
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
reveals that the roots of Eq. 16.24 will provide this same important information for a closed-loop
control system.

16.4 Simplified Block Diagram

While the final control element, process and sensor/transmitter have individual dynamics, from a
controller’s viewpoint it is impossible to separate these different behaviors. A controller sends a
signal out on one wire and sees the response in the process variable when the measurement returns on
another wire. As a consequence, the individual gains, time constants and dead times all lump together
into a single overall dynamic response. A lumped or simplified block diagram can represent this as:

Ysp(s) E(s) U(s) Y(s)

GC(s) GP(s)
+
– Controller Process

Y(s)

Figure 16.3 – Simplified Closed Loop Block Diagram in Laplace Domain

As before, we write a balance around the closed loop block diagram of Fig. 16.3 that starts and ends
with Y(s):
Y(s) = U(s)GP(s)
U(s) = E(s)GC(s) = [Ysp(s) − Y(s)]GC(s)

Combining these equations and solving for Y(s) produces the closed-loop process variable to set point
transfer function that describes the dynamic response of the measured process variable in response to
changes in set point:

Y ( s) GC ( s )G P ( s )
= (16.25)
Ysp ( s ) 1 + GC ( s )G P ( s )

The characteristic equation for this closed-loop system is the denominator of the transfer function of
Eq. 16.25, or:

1 + GC ( s )G P ( s ) = 0 (16.26)

The roots of Eq. 16.26, which are the poles of the transfer function of Eq. 16.25, indicate whether or
not the closed-loop system is stable and the degree to which it has tendency to oscillate.

16.5 The Padé Approximation

Before we continue with our analysis of block diagrams, we recognize the need for a rational
expression for dead time in the Laplace domain, e−θ s . This will permit us to employ normal algebraic
manipulations during our analysis. The Taylor series expansion for e −θ s is:

161
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 θ 2s2 θ 3 s3 θ 4 s4
e −θ s = 1 − θ s + − + K + (16.27)
2! 3! 4!
For very small values of dead time we can truncate the series as:

e −θ s ≅ 1 − θ s (16.28)

A Padé approximation is a clever expression that more accurately approximates the Taylor series of
Eq. 16.27 while providing the rational expression we seek. There are a family of Padé expressions
that become increasingly accurate as they increase in complexity. A simple Padé form we use in the
next section is exact for the first three terms of the Taylor series expansion and quite close for the
fourth term:

2 −θ s
e −θs ≅
2 +θ s
which we can show using long division yields the series:

2 −θ s θ 2s 2 θ 3s3
= 1− θ s + − + K
2 +θ s 2 4

16.6 Closed Loop Analysis Using Root Locus

The poles of interest for our simplified closed loop system are the roots of Eq. 16.26:

1 + GC ( s )G P ( s ) = 0

This analysis assumes that the process behavior, and thus, the process transfer function, remains
constant. Adjustable controller tuning provides the ability to move the poles (root location), thereby
manipulating closed-loop system behavior.

Example 1: A true first order process without dead time, with a process gain KP = 1 and a time
constant τ P = 1 , is under P-Only control. What is the impact of controller gain, KC , on closed
loop system behavior?

KP
Solution: A first order process transfer function is G P ( s ) =
τ Ps +1

From the problem statement, we know that

1
G P ( s) =
s +1

The P-Only controller transfer function is

GC ( s ) = K C

162
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
Substituting GP(s) and GC(s) into the characteristic equation, we obtain

KC
1 + GC ( s )G P ( s ) = 1 + =0
s +1

Rearranging yields s + 1 + KC = 0

We recall that s is defined in the complex plane as s = a + bi , so

a + bi + 1 + K C = 0 + 0i

Equating like real term gives us a + 1 + K C = 0

and equating like imaginary terms: bi = 0i

We now see that the roots/poles as a function of KC are

KC a =−KC −1 b=0
0 −1 0
10 −11 0
100 −101 0

We can examine this result on the s plane of Fig. 16.4 and note that the single root always lies on
the real axis as long as KC ≥ −1. For increasing positive values of KC, the real root becomes
increasingly negative:

Im

KC = 0
KC increases
Re
–1

s plane
Figure 16.4 – P-Only root locus (root location) in the complex plane

All positive values of controller gain, KC, yield a solution with no imaginary part. Hence, a true
first order system under P-Only control cannot be made to oscillate, no matter how large a KC
value used. It is also unconditionally stable for all positive KC because the root always remains on
the left hand side of the s plane.

163
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 It is interesting to note that a true first order system can remain stable even when the
controller gain has the wrong sign. For example, if KC = −0.5, then a = −0.5 and b = 0. This root
is located on the left hand side of the s plane, and thus, the system will remain stable (though
control would be poor). A value of KC = −10 yields an a = 9.0 and b = 0, which produces a root
located on the right hand side of the s plane, indicating that the system is unstable. As the next
example illustrates, even a small value of process dead time dramatically changes the inherent
dynamic nature of a closed loop system.

Õ Õ Õ

Example 2: A first order plus dead time (FOPDT) process with a process gain, KP = 1, a time
constant, τ P = 1 , and a dead time, θ P = 0.1 , is under P-Only control. What is the impact of
controller gain, KC , on closed loop system behavior?

K P e −θ P s
Solution: A FOPDT process transfer function is GP ( s ) = .
τ Ps +1

From the problem statement, we know that

e −0.1s
G P (s) =
s +1

The P-Only controller transfer function is

GC ( s ) = K C

Substituting GP(s) and GC(s) into the characteristic equation, we obtain

K C e −0.1s
1 + GC ( s )G P ( s ) = 1 + =0
s +1

We can then employ the Padé approximation:

2 − 0.1s
e −0.1s =
2 + 0.1s

Substituting the Padé approximation into the characteristic equation gives us

⎛ 2 − 0.1s ⎞ K C
1+ ⎜ ⎟ = 0
⎝ 2 + 0.1s ⎠ s + 1

Rearranging yields 0.1s 2 + (2.1 − 0.1K C ) s + 2 + 2 K C = 0

164
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
We can then multiply both sides by 10 and then solve for the roots of the characteristic equation:

−(21 − K C ) ± (21 − K C ) 2 − 4(20 + 20 K C )

p1 , p2 =
2

−21 + K C ± (441 − 42 K C + K C 2 − 80 − 80 K C )
=
2

−21 + K C ± K C 2 − 122 K C + 361

When the roots are real: =
2

−21 + K C ± i − K C 2 + 122 K C − 361

When they have imaginary parts: =
2

We can now solve for the roots using various values of KC:

KC P1 P2
0 −1.0 −20
1 −2.25 −17
2 −4.0 −15
3 −8.0 −10
repeated real roots → 3.0345 −8.99 −8.99
3.04 −8.98 + 0.4i −8.98 − 0.4i
3.2 −8.90 + 2.19i −8.90 − 2.19i
4.0 −8.50 + 5.27i −8.50 − 5.27i
10.0 −5.50 + 13.8i −5.50 − 13.8i
20.0 −0.50 + 20.49i −0.50 − 20.49i
limit of stability → 21.0 0 + 20.98i 0 − 20.98i
25.0 2 + 22.72i 2 − 22.72i
50.0 14.5 + 28.46i 14.5 − 28.46i

The limit of stability is the point where the roots fall directly on the imaginary axis (the real part
of the root is zero). This is considered the limit of stability because as soon as the roots cross over
to the positive real part of the s plane, the system becomes unstable.
An important observation from this example is that the addition of a small amount of process
dead time is enough to transform a process that will not even oscillate (as shown in Example 1)
into a process that will oscillate and then go unstable as controller gain, KC, increases.
To gauge the accuracy of the Padé approximation, we could construct this problem in Custom
Process. There we will find that the limit of stability for this system is actually closer to a
controller gain KC = 16 rather than the KC = 21 predicted from the above anlaysis. The difference
arises because LOOP-PRO does not employ an approximation for dead time in its calculations.

165
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Im

21

Re
-10 -8.99 −1

−21

Õ Õ Õ

16.7 Exercises

Q-16.1 Showing all steps, derive the closed loop “set point to measured process variable” transfer
function for this block diagram.

Ysp(s) E(s) U(s) Y(s)

GC(s) GP(s)
+
– Controller Process

YM(s) Y(s)
GM(s)
Measurement
Sensor/Transmitter

166
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
Q-16.2 Draw and label the block diagram for this process. Please use the notation given.

Fi(s) Fd (s)

hm(s)

hsp(s)

Fc(s)
h(s)

Fo(s)

167
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 17. Deriving PID Controller Tuning Correlations
17.1 The Direct Synthesis Design Equation
The PID tuning correlations we have used in the book and that are summarized in Appendix C can be
derived using the theoretical foundation we have established to this point. Direct synthesis, though
challenging to extend to the complete family of PID algorithms, is perhaps the most straightforward
method for deriving tuning correlations.
The derivation is based on the simplified block diagram of Fig. 17.1:

Ysp(s) E(s) U(s) Y(s)

GC(s) GP(s)
+
– Controller Process

Y(s)

Figure 17.1 − Simplified block diagram

As derived in Chapter 15, the closed loop “process variable to set point” transfer function for this
block diagram is:
Y (s) GC ( s )GP ( s )
= (17.1)
Ysp ( s ) 1 + GC ( s )GP ( s )

Solve Eq. 17.1 for the controller transfer function:

1 Y (s)
GC ( s ) = (17.2)
GP ( s ) ⎡Ysp ( s ) − Y ( s ) ⎤
⎣ ⎦

Divide through by YSP(s) to arrive at the controller design equation:

Y ( s)
1 Ysp ( s )
GC ( s ) = (17.3)
GP ( s ) 1 − Y ( s )
Ysp ( s )

Next, we specify that in closed loop, the measured process variable will rise to meet a step change in
set point following a FOPDT shape, or:

Y ( s) K e −θC s
= CL (17.4)
Ysp ( s ) τC s +1

168
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
Figure 17.2 illustrates this desired closed loop response shape:

YSP(t)

Y(t) Y(t) ∆y (t )
KCL = =1
∆ysp (t )

τC
θC

Time, t

Figure 17.2 − Desired closed loop response of process variable to set point changes

where for Eq. 17.4 and illustrated in Fig. 17.2:

KCL [=] closed loop process variable to set point response gain
θ C [=] closed loop dead time
τ C [=] closed loop time constant

• Specify KCL: We want the measured process variable always to equal the set point. Whenever there
is a set point change, ∆ysp(t), then the process variable, ∆y(t), should respond quickly and ultimately
change in equal magnitude. Hence, recalling that gains are computed from one steady state to the
next, we conclude:
∆y (t )
K CL = =1
∆y sp (t )

• Specify θ C : Dead time is always undesirable. Whenever possible, we avoid adding more dead time
to a loop. Yet if we are tuning a controller for a process where dead time exists, we cannot ignore it.
Consequently, we set the closed loop dead time to the minimum value possible without adding to the
process dead time, so:
θC (t ) = θ P (t )

• Specify τ C : The closed loop time constant indicates the speed of the response of a process to set
point changes. A popular heuristic for achieving a 10% to 15% overshoot to step changes in set point
is:

τ C is the larger of 0.1τ P or 0.8θ P

A heuristic for a more conservative “no overshoot” response to set point changes is:

τ C is the larger of 0.5τ P or 4θ P

169
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
These rules indicate that for small values of dead time, the closed loop process should respond from
two to ten times faster than the open loop process. To understand how this is possible, consider that
when accelerating a car from one velocity to another, you do not move the gas pedal once and wait
for the car to speed up and then steady out at the desired velocity (this is how the process time
constant is computed in open loop).
Rather, you push the gas pedal past the level where it will ultimately end up and then ease off
the pedal as the car approaches the desired speed. Similarly, a controller will send a controller output
signal that is beyond its ultimate or final value and then ease off as the measured process variable
approaches the new set point. As such, a closed loop process can indeed respond faster than its natural
open loop response.
Hence, the desired closed loop response of the measured process variable to changes in set
point expressed in Eq. 17.4 becomes:

Y (s) e −θ P s
= (17.5)
Ysp ( s ) τ C s + 1

Substitute Eq. 17.5 into the controller design equation, Eq. 17.3, yields:

e −θ P s
1 τC s +1
GC ( s ) = (17.6)
GP ( s ) e −θ P s
1−
τC s +1

which gives our final controller design equation:

1 ⎛ e −θ P s ⎞
GC ( s ) = ⎜⎜ ⎟⎟ (17.7)
GP ( s ) ⎝ τ C s + 1 − e −θ P s ⎠

17.2 Deriving Controller Tuning Correlations Using Direct Synthesis

The method of approach for deriving tuning correlations using design Eq. 17.7 is summarized as:
- transform the process ODE from the time domain into the Laplace domain to determine GP(s),
- substitute resulting GP(s) into Eq. 17.7,
- assume a rational expression for θ P ( s ) (e.g. the Padé approximation) and substitute into Eq. 17.7,
- rearrange the updated Eq. 17.7 to match the form of a controller transfer function, GC(s), from the
PID family of algorithms,
- compare forms and deduce the correlations for the tuning parameters.

Example 1: Derive the PI Controller Tuning Correlations

The PI controller tuning correlations are derived assuming our process is reasonably described
with a FOPDT model:
K e −θ P s
G P ( s) = P
τ Ps +1

170
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
Substitute the process model into Eq. 17.7, the controller design equation:

⎛ τ s +1 ⎞⎛ e −θ P s ⎞
GC ( s ) = ⎜ P −θ s ⎟⎟⎜⎜ ⎟⎟
⎜K e P −θ P s
⎝ P ⎠⎝ τ C s + 1 − e ⎠

⎛ τ s +1 ⎞⎛ 1 ⎞
= ⎜ P ⎟ ⎜⎜ −θ s
⎟⎟
⎝ KP ⎠⎝ τC s +1 − e P ⎠

Next assume a very small value of dead time and employ the simplifying approximation:

e −θ P s ≅ 1 − θ P s

The controller design equation becomes:

⎛ τ s +1⎞⎛ 1 ⎞
GC ( s ) = ⎜ P ⎟⎜ ⎟
⎝ K P ⎠⎝ τ C s + 1 − 1 + θ P s ⎠

⎛ τ s +1⎞⎛ 1 ⎞
= ⎜ P ⎟⎜ ⎟
⎝ K P ⎠⎝ τ C s + θ P s ⎠

Factoring gives:
τP ⎛ 1 ⎞
GC ( s ) = ⎜⎜1 + ⎟
K P (τ C + θ P ) ⎝ τ P s ⎟⎠

Recall the general PI controller transfer function:

⎛ 1 ⎞
GC ( s ) = K C ⎜⎜1 + ⎟⎟
⎝ τIs⎠

Comparing these equations reveals that we have derived PI controller tuning correlations if we
specify:

τP
KC = and τ I =τ P
K P (τ C + θ P )

Õ Õ Õ

Example 2: Derive the Interacting PID with Filter Controller Tuning Correlations
The Interacting PID with Filter controller tuning correlations can be derived assuming our
process can be described with an FOPDT model:

K P e −θ P s
G P (s) =
τ Ps +1

171
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
Substituting the process model into the controller design equation gives:

⎛ τ s +1 ⎞ e −θ P s
GC ( s ) = ⎜ P −θ s ⎟
⎜ K e P ⎟ τ s + 1 − e −θ P s
⎝ P ⎠ C
⎛τ s + 1⎞ 1
= ⎜⎜ P ⎟⎟
−θ s
⎝ KP ⎠τCs +1− e P

Assume a small value of dead time and use the Padé 1-1 approximation:

2 −θP s
e −θ P s ≅
2 +θP s
so:
⎛ τ s + 1⎞ 1
GC ( s ) = ⎜⎜ P ⎟⎟
⎝ K P ⎠ ⎛ 2 −θPs ⎞
τ C s + 1 − ⎜⎜ ⎟⎟
⎝ 2 +θPs ⎠

⎛τ s + 1⎞ 2 +θPs
= ⎜⎜ P ⎟⎟
2
⎝ K P ⎠ 2τ C s + τ Cθ P s + 2θ P s
Factoring gives:

⎛ τ s ⎞⎛ 1 ⎞ 2 +θ P s
G C ( s ) = ⎜⎜ P ⎟⎟⎜⎜1 + ⎟⎟
⎝ K P ⎠⎝ τ P s ⎠ (2τ C + τ C θ P s + 2θ P ) s

⎛τ ⎞⎛ 1 ⎞ 2 +θ P s
= ⎜⎜ P ⎟⎟⎜⎜1 + ⎟⎟
⎝ KP ⎠⎝ τ P s ⎠ ⎛ ⎛ τ Cθ P ⎞ ⎞
(τ C + θ P )⎜⎜ 2 + ⎜⎜ ⎟ s⎟
⎟ ⎟
⎝ ⎝τ C +θ P ⎠ ⎠

⎛ τP ⎞⎛ 1 ⎞ 2 +θ P s
= ⎜⎜ ⎟⎜1 +
⎟⎜ ⎟⎟
⎝ K P (τ C + θ P ) ⎠⎝ τ P s ⎠ ⎛⎜ ⎛ τ θ ⎞ ⎞
2 + ⎜⎜ C P ⎟ s⎟
⎜ ⎟ ⎟
⎝ ⎝τ C +θ P ⎠ ⎠

⎛ θP ⎞
⎜ ⎟
⎜ 1 + s ⎟
⎛ τP ⎞⎛ 1 ⎞ 2
⎜
GC ( s ) = ⎜ ⎟⎜
⎟⎜1 + ⎟
⎟⎜ ⎟
⎝ K P (τ C + θ P ) ⎠⎝ τ P s ⎠⎜ 1 + ⎛⎜ τ C ⎞⎟ θ P s⎟
⎜ ⎜τ + θ ⎟ 2 ⎟
⎝ ⎝ C P ⎠ ⎠

172
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Recall the general Interacting PID with Filter controller transfer function:

⎛ 1 ⎞⎛ 1 + τ D s ⎞
GC ( s ) = K C ⎜⎜1 + ⎟⎟⎜⎜ ⎟⎟
⎝ τ I s ⎠⎝ 1 + ατ D s ⎠

Comparing these equations reveals that we have derived Interacting PID with Filter controller
tuning correlations if we specify:

τP θP ⎛ τC ⎞
KC = τ I =τ P τD = α = ⎜⎜ ⎟⎟
K P (τ C + θ P ) 2 ⎝τ C + θ P ⎠

Õ Õ Õ

17.3 Internal Model Control (IMC) Structure

Internal Model Control (IMC), like direct synthesis, can be used to derive PID controller tuning
correlations. Figure 17.3 shows a simplified block diagram of the IMC structure. The unique aspect
of IMC construction is the placement of a process model, GP*(s), in parallel with the actual process it
represents.

D(s)
GD(s)

Ysp(s) E(s) U(s) + Y(s)

GC*(s) GP(s)
+ +
–
Process

–
Y*(s)
GP*(s)
+
Process Model

Y(s) – Y*(s)

Figure 17.3 – IMC Structure

As shown in the diagram, process model GP*(s) receives the actual controller output signal,
U(s), and uses it to compute Y*(s), a prediction of the measured process variable, Y(s). While in
theory, the parallel process model must be derived and programmed as part of the controller, we show
in the following sections that with certain assumptions, the structure of Fig. 17.3 can be recast into a
traditional feedback control architecture. Thus, for the specific cases of interest here, this model is
never actually created as a separate entity.

173
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
17.4 IMC Closed Loop Transfer Functions
As with direct synthesis, the controller tuning correlations are based on the closed-loop transfer
functions. To derive the closed-loop transfer functions, we perform balances on the IMC structure
shown in Fig. 17.3 by writing:

Y ( s ) = U ( s )GP ( s) + D( s)GD ( s ) (17.8)

Y * ( s ) = U ( s )GP* ( s ) (17.9)

U ( s ) = E ( s )GC* ( s ) = ⎡Ysp ( s ) − Y ( s ) + Y * ( s ) ⎤ GC* ( s ) (17.10)

⎣ ⎦

Substituting Eqs. 17.8 and 17.9 into Eq. 17.10 yields:

U ( s ) = ⎡Ysp ( s ) − U ( s )GP ( s ) − D( s )GD ( s ) + U ( s )GP* ( s ) ⎤ GC* ( s )

⎣ ⎦

= Ysp ( s )GC* ( s ) − U ( s )GP ( s )GC* ( s ) − D( s )GD ( s )GC* ( s ) + U ( s )GP* ( s )GC* ( s )

We solve for U(s):

GC* ( s ) GD ( s )GC* ( s )
U (s) = Ysp ( s ) − D(s) (17.11)
1 + GC* ( s ) ⎡GP ( s ) − GP* ( s ) ⎤ 1 + GC* ( s ) ⎡GP ( s ) − GP* ( s ) ⎤
⎣ ⎦ ⎣ ⎦

Substitute Eq. 17.11 into Eq. 17.8 to obtain:

GC* ( s )GP ( s ) GD ( s )GC* ( s )GP ( s )

Y (s) = Ysp ( s ) − D ( s ) + D ( s )GD ( s )
1 + GC* ( s ) ⎡GP ( s ) − GP* ( s ) ⎤ 1 + GC* ( s ) ⎡GP ( s ) − GP* ( s ) ⎤
⎣ ⎦ ⎣ ⎦

And rearrange into the closed loop transfer function:

GC* ( s )GP ( s ) GD ( s ) ⎡1 − GC* ( s )GP* ( s ) ⎤

Y (s) = Ysp ( s ) + ⎣ ⎦ D(s) (17.12)
1 + GC ( s ) ⎡GP ( s ) − GP ( s ) ⎤
* *
1 + GC ( s ) ⎡GP ( s ) − GP ( s ) ⎤
* *
⎣ ⎦ ⎣ ⎦

Equation 17.12 yields a set-point tracking (servo response) transfer function assuming a constant
disturbance, and disturbance rejection (regulator response) transfer function assuming a constant set
point:

174
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Y (s) GC* ( s )GP ( s )
Set-Point Tracking: = (17.13)
Ysp ( s ) 1 + GC* ( s ) ⎡GP ( s ) − GP* ( s ) ⎤
⎣ ⎦

Y ( s) GD ( s ) ⎡1 − GC* ( s )GP* ( s ) ⎤
Disturbance Rejection: = ⎣ ⎦ (17.14)
D( s ) 1 + GC ( s ) ⎡GP ( s ) − GP ( s ) ⎤
* *
⎣ ⎦

17.5 Deriving Controller Tuning Correlations Using the IMC Method

Three basic steps are used to derive IMC tuning correlations; the first two steps detail the creation of
the IMC model. The last step relates the IMC model to a classical feedback controller transfer
function to obtain controller tuning correlations.

Step 1: Recall the discussion from Chapter 15 where we established that the poles of a transfer
function (the roots of the characteristic equation in the denominator of the transfer function) indicate
system stability. If the real part of any root is positive (lies in the right hand side of the complex
plane), the system is unstable.
This concept plays a central role in the IMC analysis. The approach we take is to invert the
process model to create the controller. One problem with such an approach is that any roots in the
numerator (analogous to poles, roots in the numerator of a transfer function are called zeros) of the
process model that lie in the right hand of the complex plane, when inverted, become unstable poles.
If we permit this to occur, our controller will be unstable.
To avoid creating an unstable controller, factor the process model, GP* ( s ) , into an invertible
and noninvertible part. The classification is based on the numerator of the transfer function because
this becomes the denominator when the model is inverted in Step 2.
The noninvertible part, GP* + ( s ) , contains all right-hand plane zeros (roots in the numerator of
a transfer function that have positive real parts) and the dead time term. The invertible part, GP* − ( s ) ,
contains left hand plane zeros (roots in the numerator that have negative real parts) that produce
stable behavior when in the denominator of a transfer function. Using this notation, the process model
is factored as:
GP* ( s ) = GP* + ( s )GP* − ( s ) (17.15)

Step 2: Specify the controller transfer function as:

1
GC* ( s ) = F ( s) (17.16)
GP* − ( s )

where F(s) is a low-pass filter with gain equal to 1. The term “low-pass” is used to indicate that high
frequencies (rapid controller output changes) are lost. For deriving tuning correlations, the IMC filter
has the form:
1
F (s) = (17.17)
(τ C s + 1)
As discussed in section 17.1, the closed loop time constant, τ C , indicates the speed of the response of
a process to set point changes. A popular heuristic for achieving a 10% to 15% overshoot to step
changes in set point is:

175
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 τ C is the larger of 0.1τ P or 0.8θ P

A heuristic for a more conservative “no overshoot” response to set point changes is:

τ C is the larger of 0.5τ P or 4θ P

Step 3: Relate the IMC transfer function models to those from classical feedback control. We recall
that the closed loop transfer function for a classical feedback control architecture is:

GP ( s )GC ( s ) GD ( s )
Y (s) = Ysp ( s ) + D( s)
1 + GP ( s )GC ( s ) 1 + GP ( s )GC ( s )

We compare set point tracking forms:

Y (s) GP ( s )GC* ( s ) Y ( s) GP ( s )GC ( s )

IMC: = Classical: =
Ysp ( s ) 1 + GC* ( s ) ⎡GP ( s ) − GP* ( s ) ⎤ Ysp ( s ) 1 + GP ( s )GC ( s )
⎣ ⎦

And equate the two:

( )
GP ( s )GC* ( s ) [1 + GP ( s )GC ( s ) ] = GP ( s )GC ( s ) ⎡1 + GP ( s ) − GP* ( s ) GC* ( s ) ⎤
⎣ ⎦

GC* ( s) + GC* ( s)GP ( s )GC ( s) = GC ( s ) + GC* ( s )GP ( s )GC ( s ) + GC* ( s )GP* ( s )GC ( s )

GC* ( s )
Rearranging, we obtain GC ( s ) = (17.18)
1 − GC* ( s )GP* ( s )

We can use Eq. 17.18 to obtain a classical feedback controller from one derived from the IMC
structure. This enables us to determine the controller tuning parameters KP , τI , τD , and α.

Example: Derive the PI Controller Tuning Correlations using the IMC Method
Assume a FOPDT process model:

K P e −θ P s
GP* ( s ) =
τ Ps +1

Substitute the first-order expansion for e−θ P s :

e−θ P s ≈ 1 − θ P s

K P (1 − θ P s )
so GP* ( s ) = (17.19)
τ Ps +1

176
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
Factor GP* ( s ) into invertible and noninvertible parts:

GP* ( s ) = GP* + ( s)GP* − ( s )

so following the discussion above:

GP* + ( s ) = (1 − θ P s )

KP
GP* − ( s ) = (17.20)
τ Ps +1

We can now express the IMC controller model, GC* ( s ) , in terms of the invertible process model
term and a first-order filter term, F(s):

1
GC* ( s ) = F (s) (17.21)
GP* − ( s )

where the IMC filter has the form:

1
F ( s) = (17.22)
τC s +1

Substituting Eqs. 17.20 and 17.22 into Eq. 17.21 yields the controller:

⎛ τ s +1⎞⎛ 1 ⎞ τ Ps +1
GC* ( s ) = ⎜ P ⎟⎜ ⎟= (17.23)
⎝ K P ⎠⎝ τ C s + 1 ⎠ K P (τ C s + 1)

We can relate this IMC controller model, GC* ( s ) , to a classical feedback controller model via Eq.
17.18::

GC* ( s )
GC ( s ) =
1 − GP* ( s )GC* ( s )

We substitute Eq. 17.19 and 17.23 into Eq. 17.18 and simplify:
τ Ps +1
K P (τ C s + 1)
GC ( s ) =
⎛ K (1 − θ P s ) ⎞ ⎛ τ P s + 1 ⎞
1− ⎜ P ⎟⎜ ⎟
⎜ τ P s + 1 ⎟ ⎜ K P (τ C s + 1) ⎟
⎝ ⎠⎝ ⎠
τ Ps +1
K P (τ C s + 1)
=
⎛ K P (1 − θ P s ) ⎞
1− ⎜ ⎟
⎜ K P (τ C s + 1) ⎟
⎝ ⎠

177
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 τ Ps +1

=
(
KP τC s + 1 )
τC s + 1 −1 + θPs
τC s +1

τ Ps +1
=
K P s (τ C + θ P )

τP s 1
= +
K P s (τ C + θ P ) K P s (τ C + θ P )

τP ⎡ 1 ⎤
GC ( s ) = ⎢1 + ⎥ (17.24)
K P (τ C + θ P ) ⎣ τ P s ⎦

Compare Eq. 17.24 to the classical feedback model for a PI controller,

⎡ 1 ⎤
GC ( s ) PI = K C ⎢1 + ⎥
⎣ τIs⎦

we obtain the following controller tuning parameters:

τP
KC = and τI =τP
K P (τ C + θ P )

Õ Õ Õ

Additional tuning correlation derivations for several controllers from the PID family can be found in
Appendix A.

17.6 Exercises

Q-17.1 Showing all steps, use the Direct Synthesis method to derive controller tuning correlations for
the PID control algorithm

⎛ 1 ⎞
GC ( s ) = K C ⎜⎜1 + + τ D s ⎟⎟
⎝ τIs ⎠

Assume the process is best described by a second order plus dead time model of the form

K P e −θ P s
G P (s) =
(τ 1 s + 1)(τ 2 s + 1)

178
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 and that the process has very little dead time, or e −θ P s ≅ 1 − θ P s .

Hint: the resulting correlations will not match those in the Controller Tuning Guide.

Q-17.2 Showing all steps, use the Direct Synthesis method to derive PI controller tuning
correlations, assuming the dynamic behavior of a process is described by a first-order
without dead time transfer function.

179
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Combining Theory and Practice

18. Cascade Control

18.1 Architectures for Improved Disturbance Rejection
The most popular architectures for improved regulatory performance are cascade control and the feed
forward with feedback trim architecture discussed in the next chapter. Both architectures trade off
additional complexity in the form of instrumentation and engineering time in return for a controller
better able to reject the impact of disturbances on the measured process variable. Neither architecture
benefits nor detracts from set point tracking performance.
As we will soon learn, the feed forward with feedback trim architecture requires an additional
sensor and the programming of a dynamic model. The sensor is installed to directly measure changes
in the disturbance variable. The model receives the disturbance measurement, computes control
actions to counter or negate its impending impact on the measured process variable, and transmits the
result to the controller for execution. Because of this architecture, feed forward is useful when one
specific disturbance variable is responsible for repeated, costly disruptions to stable operation. It is
also useful when a secondary measured process variable cannot be identified to construct a control
cascade as discussed in the next section.

18.2 The Cascade Architecture

A feed forward element seems reasonably straightforward in concept; directly measure a nasty
disturbance and use a model to instruct the controller how and when to take actions to negate its
impact on the measured process variable. Developing and programming a proper dynamic model,
however, can be a challenging task that should not be underestimated.

cascade control can improve but can not help rejection

rejection of this disturbance
of this disturbance
disturbance
variable II Disturbance
Process II
disturbance
variable I Disturbance
Process I

secondary primary
primary secondary process process
set point set point Final variable Primary variable
Primary Secondary +
++
Secondary
+– Controller +– Controller Control +
Process Process
Element

secondary process variable

primary process variable

Figure 18.1 − Block diagram of the cascade architecture

180
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Cascade control is more difficult to conceptualize, yet implementation is a familiar task
because the architecture is comprised of two ordinary controllers from the PID family. Like feed
forward, cascade provides minimal benefits for set point tracking. Before you begin your design, be
sure your goal is improved disturbance rejection.
In a traditional feedback loop, a controller adjusts a manipulated variable so the measured
process variable remains at set point. The cascade design requires that you identify a secondary
process variable (we will henceforth call the main process variable associated with original control
objective the primary variable). This secondary process variable has specific requirements:
• it must be measurable with a sensor,
• the same final control element (e.g. valve) used to manipulate the primary variable must also
manipulate the secondary variable,
• the same disturbances that are of concern for the primary variable must also disrupt the
secondary variable, and
• the secondary variable must be inside the primary process variable, which means it responds
well before the primary variable to disturbances and final control element manipulations.
With a secondary process variable identified, a cascade is constructed as shown in Fig. 18.1.
The block diagram shows how both Disturbance I and the final control element impact the secondary
variable before they affect the primary variable. Notice that the secondary loop has a traditional
feedback control structure, except here it is literally nested inside the primary loop.
A cascade requires two sensors and two controllers but only one final control element
because the output of the primary controller, rather than going to a valve, becomes the set point of the
secondary controller. Because of the nested architecture, success in a cascade implementation
requires that the settling time of the (inner) secondary loop is significantly faster than the settling
time of the primary (outer) loop.
As the above discussion implies, one advantage of a control cascade is that it is not tied to a
single disturbance. Rather, the same cascade can address multiple disturbances as long as each
impacts the inner secondary variable well before it impacts the outer primary variable. Also, as
mentioned before, implementation uses our existing skills because the architecture is comprised of
two ordinary controllers from the PID family.

18.3 An Illustrative Example

The Flash Drum Process
To better understand the construction and use of cascade control, consider the flash drum process
shown in Fig. 18.2. This process is not a case study in LOOP-PRO, but its behavior is reasonably
intuitive, and this makes it useful for a qualitative discussion of the important concepts.
As shown in the figure, the feed stream entering the flash valve is a hot liquid under high
pressure. A flash valve produces a large and sudden drop in pressure as the liquid flows toward the
flash drum, permitting some of the hot liquid to vaporize. As a result, a vapor phase forms over the
liquid inside the drum. Figure 18.2 shows that a vapor stream exits overhead from the drum while a
liquid stream exits from the bottom. It is essential that the liquid level never falls so low that vapor is
sent down the liquid drain nor rises so high that liquid enters the vapor line.
Given this scenario, our design objective is to control liquid level in the drum. Liquid drain
flow rate is selected as the manipulated variable in this example. If liquid level is too high, the
controller should increase liquid drain flow rate. If level is too low, the controller should decrease
liquid drain flow rate.

181
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 The disturbance of concern in this example is the pressure in the vapor phase. As indicated in
the figure, the overhead vapor phase pressure changes without warning due to the behavior of some
unidentified down stream unit.
pressure set
overhead vapor
down stream
(a disturbance)
P

vapor
hot liquid feed
LC Lsetpoint
flash
valve liquid

liquid
drain
valve position manipulated
to control liquid level

Figure 18.2 − Single-loop level control of a flash drum

Problems with Single Loop Control

The controller of Fig. 18.2 attempts to achieve the design objective by adjusting valve position in the
liquid drain stream. If the measured level is higher than set point, the controller signals the valve to
open by an appropriate percentage with the expectation that this will increase drain flow rate
accordingly. But drain flow rate is a function of several variables, including
- valve position,
- the hydrostatic head (height of the liquid), and
- the pressure of the vapor pushing down on the liquid (a disturbance).
Assume for a moment that the pressure of the vapor phase is constant over time (and the feed flow
rate and composition are also constant) then as the drain valve opens and closes, the liquid drain flow
rate increases and decreases in a predictable fashion. Hence, the single loop architecture of Fig. 11.2
should provide satisfactory liquid level control performance.
However, as Fig. 18.2 indicates, the pressure of the vapor phase is controlled by a unit down
stream of our flash drum. Rather than remaining constant, the pressure of the vapor phase changes
over time and this acts as a disturbance to the level control process.
Suppose the vapor phase pressure starts decreasing. This disturbance will cause the pressure
pushing down on the liquid interface to decrease. If the valve position remains constant, the liquid
drain flow rate will similarly decrease. If the pressure decrease occurs quickly enough, the controller
can actually be opening the valve yet the liquid drain flow rate can continue to decrease.
Alternatively, if the pressure in the vapor phase starts to increase, the controller can be closing the
valve yet the liquid drain flow rate can actually increase.
The pressure change disturbance causes a contradictory outcome that can confound the
controller and result in poor control. As the preceding discussion indicates, it is not valve position but
liquid drain flow rate which must be adjusted for high performance disturbance rejection.

182
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
A Cascade Control Solution
A first step in cascade design is to ensure that our control objective is disturbance rejection. The
scenario presented supports that the controller is not intended for set point tracking. In fact, the set
point will be constant at mid-drum level during normal operation. Our goal is to maintain liquid level
at set point while rejecting the disturbance of pressure changes in the overhead vapor phase.
To implement a cascade we must be able to identify a secondary process variable. Liquid
level becomes the primary process variable and controlling it remains the central design objective of
our strategy. For the secondary process variable we propose liquid drain flow rate. As required by the
cascade design criteria:

• liquid drain flow rate is measurable with a sensor,

• the same valve used to manipulate liquid level (the primary variable) also manipulates the
liquid drain flow rate,
• changes in vapor phase pressure that disturb liquid level control also impact the drain flow
rate, and
• drain flow rate is inside the liquid level in that it responds well before liquid level to changes
in valve position and changes in vapor phase pressure.

Figure 18.3 shows a cascade architecture with two controllers (level control and drain flow
rate control), two measurement sensors (measuring liquid level and liquid drain flow rate) and one
final control element (a valve in the liquid drain stream). Figure 11.4 shows a block diagram of this
same level-to-flow cascade.

pressure set
overhead vapor
down stream
(a disturbance)
P

vapor
hot liquid feed
LC Lsetpoint
flash
valve liquid

Fsetpoint

FC
liquid
drain
flow rate manipulated
to control liquid level

Figure 18.3 − Level-to-flow cascade control of liquid level

Liquid level control, our main objective, is the primary or outer loop. The output of the
primary controller is the set point of the secondary controller, which controls liquid drain flow rate by
adjusting the valve position. Flow control dynamics are much faster than level control dynamics.
Hence, this configuration is consistent with the previously mentioned design criteria that the settling
time of the secondary loop must be significantly faster than the settling time of the primary loop.

183
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 vapor phase
pressure Pressure to
Drain Flow
Relationship

primary secondary liquid drain liquid

set point set point Liquid flow rate Drum Level level
Secondary
++
Primary Liquid Drain
+– Controller +– Controller
Drain
Process Process
Lsetpoint Fsetpoint Valve

secondary process variable (liquid drain flow rate)

primary process variable (liquid level)

Figure 18.4 − Block diagram of the level-to-flow cascade architecture

With this cascade architecture, if the liquid level is too high, the level controller now
specifically calls for an increased liquid drain flow rate rather than simply an increase in valve
opening as was the case in the single loop configuration of Fig. 18.2. It is the flow controller that then
determines whether this means opening or closing the valve and by how much. Thus, a pressure
disturbance in the vapor phase gets addressed quickly by the secondary flow controller and this
dramatically improves the disturbance rejection performance of primary control loop.

18.4 Tuning a Cascade Implementation

Cascade loop tuning uses the skills we have developed in previous chapters:
1) Begin with both the primary and secondary controllers in manual mode.
2) Select a P-Only controller for the inner secondary loop (integral action increases settling
time and offset is rarely an issue for the secondary process variable).
3) Tune the secondary P-Only controller using a set point tracking criterion (its main job is to
respond to set point commands from the primary controller). Test it to ensure a satisfactory
set point tracking performance.
4) Leave the secondary controller in automatic; it now literally becomes part of the primary
process. Select a controller with integrating action for the primary loop (PI or PID). Use a
disturbance rejection design criterion as this is the main job of the primary controller.
5) Tune the primary controller using methods discussed in the previous chapters and test it to
ensure acceptable performance.
6) With both controllers in automatic, tuning of the cascade is complete.

18.5 Exploring the Jacketed Reactor Process

As described in Section 2.6 and shown in Fig. 18.5 for the single feedback loop case, the jacketed
reactor is a continuously stirred vessel in which an exothermic (heat producing) reaction occurs.
Residence time is constant in this well mixed reactor, so the conversion of reactant feed to desired
product can be inferred from the temperature of the reactor exit stream.

184
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 To control reactor exit stream temperature (the measured process variable), the vessel is
enclosed with a jacket through which a cooling liquid flows. The controller manipulates a valve to
adjust the cooling liquid flow rate. If the exit stream temperature (and thus conversion) is too high,
the controller opens the valve. This increases cooling liquid flow rate, which cools off the reactor and
causes the heat producing reaction to slow. Ultimately, the measured temperature of the stream
exiting the reactor drops in response to the control action. As shown in Fig. 18.5, the disturbance
variable of interest for this process is the temperature of cooling liquid entering the jacket.

Figure 18.5 − Jacketed reactor process with single loop feedback control architecture

In the study explored here, the control objective is to maintain the reactor exit stream
temperature at set point by rejecting disturbances caused by changes in the cooling jacket inlet
temperature. The disturbance rejection performance of a single loop PI controller configuration is first
presented and then compared to that of a PI to P-Only controller cascade architecture.
For both the single loop and cascade investigation, the design level of operation is a reactor
exit stream temperature (measured process variable) of 86°C. The cooling jacket inlet temperature
(disturbance) is normally at its design value of 46°C, but our concern is that on occasion, it is known
to unexpectedly spike as low as 40°C. An open loop study establishes that a controller output of 50%
causes the reactor to steady at the design measured exit stream temperature of 86°C when the cooling
jacket inlet temperature is at its expected or design value of 46°C.

18.6 Single Loop Disturbance Rejection in the Jacketed Reactor

Tuning the single loop PI controller shown in Fig. 18.5 follows the same design procedure we have
always used in this book. We perturb the controller output, record the data as the process responds, fit
a FOPDT (first order plus dead time) model to the data, and use the resulting model parameters in a
tuning correlation to determine initial controller tuning values.

185
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 As shown in Fig. 18.6, we use a controller output doublet to generate measured process
variable data both above and below the design level of operation. Although a variety of dynamic tests
would produce an equally useful data set, here the controller output is stepped from the design value
of 50% up to 53%, then down to 47%, and finally back to 50%. The measured reactor exit stream
temperature exhibits a clear response after each controller output step that dominates any
measurement noise.

Modeling Controller Output to Process Variable Dynamics

Model: First Order Plus Dead Time (FOPDT) File Name: Process.txt
Process Variable

87

86

85 process data and

FOPDT model fit
84

controller output
Controller Output

54
52 steps in manual mode
50
48
46

0 5 10 15 20 25
Time

Gain (K) = -0.36, Time Constant (T1) = 1.58, Dead Time (TD) = 0.88
SSE: 0.4316

Figure 18.6 − FOPDT model fit of single loop controller output to measured process variable data

A FOPDT fit of the dynamic process data as computed by Design Tools is also shown in Fig. 18.6.
The model appears to be reasonable and appropriate based on visual inspection, thus providing the
following design parameters:
Process Gain, KP = − 0.36°C/%
Time Constant, τP = 1.6 min
Dead Time, θP = 0.88 min
To use the IMC correlations, we first compute the closed loop time constant. Here we choose
aggressive tuning:
τC = larger of 0.1τP or 0.8θP = larger of 0.1(1.6) or 0.8(0.88) = 0.70 min.
Substituting this closed loop time constant and the above FOPDT model parameters into the IMC
tuning correlations of Eq. 8.5 yield the following tuning values:

Controller Gain, KC = − 2.8 %/°C

Reset Time, τI = 1.6 min

The disturbance rejection performance of the single loop PI controller using these tuning
parameters is shown in Fig. 18.7. The controller label to the upper right of the plot confirms that the
PI controller has a direct acting proportional term, an anti-reset windup integral term, and a derivative
term that is off.

186
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 As shown in Fig. 18.7, the measured reactor exit stream temperature is initially steady at the
design set point value of 86°C. To test the controller, the cooling jacket inlet temperature is stepped
from its design value of 46°C down to 40°C and back again. As shown, the single loop PI controller
is able to maintain reactor exit stream temperature near the constant set point of 86°C, with deviations
ranging as high as 2.5°C during the event.

Disturbance Rejection Performance of Single Loop PI Controller

Process: Single Loop Jacketed Reactor Cont.: PID ( P= DA, I= ARW, D= off, F = off)
88
PV/Setpoint

86

84 constant
reactor exit
set point
60
temperature
Contr Output

50
40
30

48
disturbance
46
Disturbance

44
variable steps
42
40

15 20 25 30 35 40 45
Time (mins)
Tuning: Gain = -2.80, Reset Time = 1.60, Sample Time = 1.00

Figure 18.7 − Disturbance rejection in the jacketed reactor under single loop PI control
show process variable deviations reach 2.5°C (compare to Fig. 18.13)

18.7 Cascade Disturbance Rejection in the Jacketed Reactor

The open loop dynamic behavior of the cascade jacketed reactor is identical to that of the single loop
case. The process graphic for the cascade controller architecture is shown in Fig. 18.8.

The Cascade Architecture

As discussed in the single loop study, our control objective is disturbance rejection, so it is
appropriate to consider a cascade architecture. The primary process variable remains the reactor exit
stream temperature. To construct a cascade, we need to identify a secondary process variable. As
shown in Fig. 18.8, LOOP-PRO uses the cooling jacket outlet temperature. As required for a cascade
design:
• cooling jacket outlet temperature is measurable with a sensor,
• the same valve used to manipulate reactor exit stream temperature (the primary variable) also
manipulates the cooling jacket outlet temperature,
• changes in cooling jacket inlet temperature that disturb the reactor exit stream temperature
also impact the cooling jacket outlet temperature, and
• the cooling jacket outlet temperature is inside the reactor exit temperature in that it responds
first to changes in valve position and changes in the cooling jacket inlet temperature.

187
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
A block diagram of this architecture is shown in Fig. 18.9. Like all cascades, there are two
measurements, two controllers and one final control element; the same final control element as in the
single loop case.

Figure 18.8 − Jacketed reactor process with cascade control architecture

The primary (outer) process is still the reactor and the primary measured process variable is
the reactor exit stream temperature. The output of the primary controller is the set point of the
secondary controller. The inner (secondary) process is the cooling jacket. The manipulated variable of
the secondary loop is the cooling jacket liquid flow rate and the secondary measured process variable
is the cooling jacket outlet temperature.
jacket inlet
temperature Inlet to Outlet
Jacket Temp
Relationship

primary secondary jacket outlet reactor exit

set point set point Secondary Jacket temperature temperature
++
Primary Cooling Jacket Reactor
+– Controller +– Controller Flow
Process Process
Tsetpoint Tsetpoint Valve

secondary process variable (cooling jacket outlet temperature)

primary process variable (reactor exit stream temperature)

Figure 18.9 − Block diagram of jacketed reactor cascade architecture

188
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
Secondary P-Only Controller
To implement a cascade, the secondary controller is tuned while the primary controller is in manual
mode. The design operating conditions are the same as those used for the single loop PI controller
study. That is, with the cooling jacket inlet temperature at 46°C and the controller output at 50%, the
reactor exit stream temperature steadies at the design value of 86°C. We note at these design
conditions that the cooling jacket outlet temperature, the secondary measured process variable,
steadies at 69°C. Hence, for the secondary controller:
ysetpoint = 69°C

The bias is the value of the controller output that, in open loop, causes the measured process
variable to steady at its design condition when the disturbances are at their design or expected value.
So for the secondary P-Only controller:
ubias = 50%

Starting from steady state at the design level of operation, a doublet is used to generate
controller output to secondary process variable dynamic data as shown in Fig. 18.10. The controller
output is stepped from the design value of 50% up to 55%, down to 45%, and back to 50%. After
each control action, the secondary process variable displays a clear response that dominates
measurement noise.

Model Dynamic Data for Secondary P-Only Controller Design

Model: First Order Plus Dead Time (FOPDT) File Name: secondary.txt

71
Secondary PV

70
69
68 secondary process variable
67 data and FOPDT model fit

secondary controller
Secondary CO

55
output steps in manual mode
50

45

0 5 10 15 20
Time

Gain (K) = -0.374, Time Constant (T1) = 1.93, Dead Time (TD) = 0.2527
SSE: 1.14

Figure 18.10 − FOPDT fit of controller output to secondary process variable dynamic data

A FOPDT dynamic model fit to the data, also shown in Fig. 11.10, yields the following secondary
control loop model parameters.
Process Gain, KP = − 0.37°C/%
Time Constant, τP = 1.9 min
Dead Time, θP = 0.25 min

189
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Although disturbance rejection is the overall objective, the goal of the inner secondary loop is
to track set point changes computed by the primary controller. Using these FOPDT model parameters
in the ITAE for set point tracking correlation (recall that IMC correlations do not exist for P-Only
controllers) yields the following P-Only tuning parameter:

Controller Gain, KC = − 6.4 %/°C

P-Only set point tracking performance is shown in Fig. 18.11. The primary loop is still in
manual mode at this point. As expected for a P-Only controller, offset exists when the set point is not
at the design value. The secondary process variable responds quickly and settles rapidly to set point
changes so we consider the design of the secondary loop to be complete. The secondary loop is left in
automatic and literally becomes part of the primary process. We now tune the primary controller.

Step Secondary Set Point to Generate Primary PV Dynamic Data

Process: Cascade Jacketed Reactor Pri: Manual Mode
Sec: PID ( P= DA, I= off, D= off, F = off)

88
Primary PV

86

84

82
secondary P-Only
set point steps
primary process variable
72
dynamic data
Secondary PV

70

68

66

0 5 10 15 20 25 30
Time (mins)

Tuning: Gain = -6.40, Sample Time = 1.00

Figure 18.11 − Set point tracking performance of the secondary loop under P-Only control

Primary Loop Control

In previous chapters, we have always generated dynamic process data for controller design by
stepping, pulsing or otherwise perturbing the controller output signal. In the cascade architecture,
however, the controller output of the primary loop is the set point of the secondary controller. So to
design the primary controller, we must step, pulse or otherwise perturb the secondary set point and
then fit a model to the corresponding response in the primary measured process variable. Figure 18.11
contains such data and this is used for the design of the primary PI controller.
The set point of the secondary P-Only controller, as shown in Fig. 18.12, is stepped in a
doublet from its design value of 69°C up to 72°C, down to 66°C, and back to 69°C. The measured
reactor exit stream temperature displays a clear response after each change that dominates the
measurement noise. A FOPDT fit of the dynamic data, shown in Fig. 18.12, yields the following
primary control loop model parameters.

190
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 °C of reactor exit stream
Process Gain, KP = 0.70
°C of cooling jacket outlet stream
Time Constant, τP = 0.55 min
Dead Time, θP = 0.71 min

We first compute the closed loop time constant. Here we choose aggressive tuning:

τC = larger of 0.1τP or 0.8θP = larger of 0.1(0.55) or 0.8(0.71) = 0.57 min.

Substituting this closed loop time constant and the above FOPDT model parameters into the IMC
correlations for PI control yields the following tuning values:

°C of cooling jacket outlet stream

Controller Gain, KC = 0.61
°C of reactor exit stream
Reset Time, τI = 0.55 min

These tuning values are implemented on the primary loop and the design of the cascade is complete.

FOPDT Model Fit for Design of Primary Controller

Model: First Order Plus Dead Time (FOPDT) File Name: secondary.txt
Sec: Manual Mode

88
Primary PV

86
primary process data
84
and FOPDT model fit
82
secondary P-Only
set point steps
Secondary SetPoint

72

70

68

66

0 5 10 15 20 25 30
Time

Gain (K) = 0.7049, Time Constant (T1) = 0.5497, Dead Time (TD) = 0.7135
SSE: 13.78

Figure 18.12 − FOPDT model fit of dynamic data from a primary loop doublet test

Figure 18.13 shows the disturbance rejection performance of the cascade using these tuning
parameters for the primary PI controller while the secondary loop remains under P-Only control as
previously described. Just as in Fig. 18.7, the primary measured process variable (reactor exit stream
temperature) is initialized at the design set point value of 86°C. To test the controller, the cooling
jacket inlet temperature is again stepped from its design value of 46°C down to 40°C and back again.

191
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 One interesting outcome of this cascade implementation is the offset that occurs in the
secondary P-Only loop when the process moves away form the design level of operation. Of more
importance, however, is that the cascade shows improved performance over the single loop case in
maintaining the reactor product temperature near the constant set point of 86°C.
Specifically, while the measured reactor exit temperature deviations for the single loop PI
controller range as high 2.5°C during the event, here the cascade limits the maximum deviation to
about 1.0°C. This improved performance did not come free, however, as the cascade architecture
requires an additional sensor, controller and tuning effort.

Disturbance Rejection Performance of Cascade Architecture

Process: Cascade Jacketed Reactor Pri: PID ( P= RA, I= ARW, D= off, F = off)
Sec: PID ( P= DA, I= off, D= off, F = off)
88
Primary PV

86
84 constant set point
P-Only control offset for primary PV
72
Sec PV & CO

71
70
69
68
48 disturbance
Disturbance

46
44 variable steps
42
40

50 55 60 65 70 75 80 85
Time (mins)
Tuning: Gain = 0.61, Reset Time = 0.55, Sample Time = 1.00

Tuning: Gain = -6.40, Sample Time = 1.00

Figure 18.13 − Disturbance rejection in the jacketed reactor with a P-Only to PI cascade
Process variable deviations reach 1.0°C (compare to Fig. 11.7)

18.8 Set Point Tracking Comparison of Single Loop and Cascade Control
The cascade architecture does not provide benefit in tracking set point changes and this is illustrated
in Fig. 18.14 for the jacketed reactor. The plot on the left shows the set point tracking performance of
the single loop PI controller while the plot on the right shows that for the cascade. The lower trace on
each plot is the controller output signal sent to the valve located on the jacket cooling outlet stream.
The performance of the single loop controller on the left may reasonably be considered
superior to that of the cascade on the right. Do not forget, however, that the single loop PI controller
was tuned using the IMC for set point tracking and disturbance rejection correlation. The primary
loop of the cascade was tuned using the ITAE for disturbance rejection correlation, which generally
provides more aggressive tuning values.
This single example is not sufficient to support a claim that one architecture performs better
than the other for set point tracking. We close this chapter, however, by restating once again that
before considering a cascade architecture, be sure the controller design objective is the rejection of
disturbances.

192
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Set Point Tracking Performance Under PI Control Set Point Tracking Performance Under Cascade Control
Process: Single Loop Jacketed Reactor Controller: PID ( P= DA, I= ARW, D= off ) Process:
Process: Cascade
Cascade Jacketed
Jacketed Pri:Reactor
Reactor PIDPri:
( P=PIDRA,
( P=I=
RA,ARW, D=D=off
I= ARW, off)) / /Sec:
Sec:PIDPID
( P=(DA,
P= I=
DA,
off, I=
D= off,
off)

91 91
90 90
PV/Setpoint

Primary PV
89 89
88 88
87
87
86
86
85 set point tracking
85
set point tracking 60 performance
60 performance

Secondary CO
50
Controller Output

55
50 40
45
40 30
35
30
100 105 110 115 120 125 130 135
Time (mins)
20 25 30 35 40 45 50 55
Time (mins)
Tuning: Gain = 1.00, Reset Time = 0.95, Deriv Time = 0.0, Sample Time = 1.00
Tuning: Gain = -2.70, Reset Time = 1.60, Deriv Time = 0.0, Sample Time = 1.00 Tuning: Bias = 50.0, Gain = -5.80, Sample Time = 1.00

Figure 18.14 − Comparing set point tracking of single feedback loop to that of the cascade

18.9 Exercises
Q-18.1 Repeat the cascade design study presented in this chapter, only for comparison,
a) use a PID algorithm for the outer primary loop
b) use a PI algorithm for the inner secondary loop
Comment on how these changes impact controller performance. Include plots that support
your conclusions.

193
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 19. Feed Forward Control
19.1 Another Architecture for Improved Disturbance Rejection
As discussed in the previous chapter, the most popular architectures for improved regulatory
performance are cascade control and the feed forward with feedback trim architecture discussed here.
Like cascade, a feed forward implementation requires additional instrumentation and engineering
time and neither benefits nor detracts from set point tracking performance.
Before considering feed forward for your application, be sure your control objective is
disturbance rejection. Choose feed forward over cascade if one specific disturbance variable is
responsible for repeated, costly disruptions to stable operation, or if an appropriate secondary
measured process variable as required for a cascade implementation cannot be identified.

19.2 The Feed Forward Architecture

The process block diagram used in the cascade discussion is shown in Fig. 19.1 with the secondary
cascade control loop removed. As we have learned, a cascade architecture can improve the rejection
performance of Disturbance I but not Disturbance II because there is no secondary process variable
associated with that disturbance. As indicated in this figure, feed forward with feedback trim holds
potential for improved disturbance rejection for both classes of disturbance.

feed forward can improve

rejection of either disturbance
disturbance
variable II Disturbance
Process II
disturbance
variable I Disturbance
Process I

measured
process
set point Final variable
++
Feedback
++
Secondary Primary
+– Controller
Control
Process Process
Element

measured process variable

Figure 19.1 - Feed forward controllers do not require a secondary process variable

When a traditional feedback controller works to reject a disturbance, corrective action begins
only after the measured process variable has been forced away from set point. Damage to stable
operation is in progress before a traditional feedback controller even begins to respond.
Consider that many disturbances originate in some other part of the plant. A measurable
series of events occur that cause that “distant” event to ultimately impact your process. From this
view, the traditional feedback controller simply starts too late to be effective in reducing or negating
the impact of a disturbance.

194
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 A feed forward controller gains advantage by using a sensor to directly measure the
disturbance before it reaches the process. As shown in Fig. 19.2, a feed forward element receives the
disturbance measurement signal and uses it to compute and schedule preemptive control actions that
will counter the impact of the disturbance just as it reaches the measured process variable.

d(t) Disturbance to
Process Variable
Behavior

sensor “disturbance”

utotal is the feedback controller output

minus the feed forward output signal
Feed Forward Element
Disturbance Model
Process Model
ydisturb
– ufeedforward
ysetpoint Traditional Controller Output to y(t)
+– Feedback ++ Process Variable
yprocess ++
ufeedback utotal Behavior
Controller
“process”

measured process variable

Figure 19.2 - Architecture of a feed forward controller with feedback trim

A feed forward implementation requires the purchase and installation of a sensor and the construction
of a feed forward model element. This element is comprised of a disturbance model and a process
model. Both models are linear in form. The computation performed by the feed forward element may
be thought of as a two step procedure:
• The disturbance model receives disturbance measurement, d(t), and predicts an “impact
profile,” or when and by how much the measured process variable, y(t), will be impacted.

• Given this predicted sequence of disruption to y(t), the process model then back calculates a
series of control actions, ufeedforward(t), that will exactly counteract the disturbance as it arrives
so the measured process variable remains constant at set point.

Implementation requires that these linear models be programmed into the control computer. As we
have learned, linear models never exactly describe real process behavior. So although a feed forward
element can dramatically reduce the impact of a disturbance, it never will succeed in providing
perfect disturbance rejection.

195
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 To account for model inaccuracies, the feed forward signal is combined with traditional
feedback control action, ufeedback(t), to create a total controller output, utotal(t). The feedback controller
provides trim. That is, it rejects those portion of the measured disturbance that make it past the feed
forward element and reach the measured process variable. The feedback controller also works to
reject all other unmeasured disturbances affecting plant operation and provides set point tracking
capabilities as needed.
Notice in Fig. 19.2 that the computed feed forward control action, ufeedforward, is assigned a
negative sign. It is subtracted from the feedback output signal to create a total controller output:
utotal(t) = ufeedback(t) − ufeedforward(t) (19.1)

This makes sense because if the disturbance model predicts that a particular disturbance will cause the
measured process variable to, say, move up by a certain amount over a period of time, the process
model must compute feed forward control actions that cause the measured process variable to move
down in the same fashion. The negative sign enables “action opposite to prediction” to be taken.

19.3 An Illustrative Example

Flash Drum Process Revisited
To better understand the construction and use of feed forward with feedback trim, we revisit the flash
drum process, shown again in Fig. 19.3. We used this example of a disturbance rejection objective in
Section 18.3. Since feed forward is an alternative strategy for improved disturbance rejection, we use
the same process here so we can compare the two techniques.
As shown in Fig. 19.3, a hot liquid under high pressure flashes as it enters the drum
(experiences a large and sudden pressure drop). The result is a vapor and liquid phase. The design
objective is to control liquid level in the drum. If level is too high, the controller should increase
liquid drain flow rate. If level is too low, the controller should decrease liquid drain flow rate. The
disturbance of concern is the pressure in the overhead vapor phase, which can change without
warning due to the behavior of some unidentified down stream unit.

pressure set
overhead vapor
down stream
(a disturbance)
P

vapor
hot liquid feed
LC Lsetpoint
flash
valve liquid

liquid
drain
valve position manipulated
to control liquid level

Figure 19.3 − Single-loop level control of a flash drum

196
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
Problems with Single Loop Control
The liquid drain flow rate is a function of several variables, including valve position, hydrostatic head
(height of the liquid), and pressure of the vapor pushing down on the liquid (a disturbance). Suppose
the vapor phase pressure starts decreasing. This disturbance will cause the pressure pushing down on
the liquid interface to decrease. If the pressure decrease occurs quickly enough, the controller can
actually be opening the valve yet the liquid drain flow rate can continue to decrease. Alternatively, if
the pressure in the vapor phase starts increasing, the controller can be closing the valve yet the liquid
drain flow rate can actually increase.

A Feed Forward Solution

As shown in Fig. 19.4, a feed forward control strategy includes a sensor that directly measures
changes in a specific disturbance variable, and a feed forward model element to compute corrective
control actions based on that measurement. In this case, a sensor is installed to detect changes in the
pressure of the vapor phase.
overhead vapor pressure set down stream
(a disturbance)

P
Feed Forward Element
Disturbance Model
Process Model
vapor
hot liquid feed Lsetpoint
LC
flash
valve liquid

liquid
drain

Figure 19.4 − Controlling liquid level with a feed forward with feedback trim architecture

As the pressure signal is received by the feed forward element, the disturbance model
continually predicts a liquid level impact profile, or how far and how fast the liquid level will change
for a given change in the vapor phase pressure. For example, if the pressure begins to increase, the
model will predict how the drum level will fall as liquid is forced out the drain faster.
Based on this impact profile prediction, the process model then computes a precise sequence
of control actions (adjustments to the liquid drain valve), as well as the timing of these actions. The
goal is to adjust the valve so that drain flow rate does not change as pressure changes, and thus the
impending upset to drum level will be canceled.
With reasonably accurate models, the feed forward controller can substantially reduce the
impact of pressure changes on liquid level. Perfect elimination of the disturbance is not likely because
the linear feed forward dynamic models will not exactly describe the nonlinear and time varying
behavior of this (or any real) process.

197
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 To make up for plant-model mismatch, the traditional feedback level controller, represented
by the LC in the circle attached to the tank, trims the control system. That is, the feedback controller
rejects those portions of any pressure disturbance that make it past the feed forward controller and
succeed in impacting the liquid level. It also enables rejection of all other disturbances to the liquid
level control process (such as changes in feed composition) and provides set point tracking
capabilities when required.

19.4 Feed Forward Control Design

Design Criteria
Implementation of a feed forward element requires the installation of a sensor that directly measures
the disturbance variable and the programming of a feed forward element comprised of a process and
disturbance dynamic model. To benefit from this technology, two design criteria for success are:
1) The process and disturbance dynamic models must reasonably describe the controller output to
measured process variable behavior and disturbance to measured process variable behavior
respectively, and
2) The process dead time (controller output to measured process variable dead time) must be
shorter than the disturbance dead time (disturbance to measured process variable dead time).
The first criterion is rather obvious. If the models don’t describe the behavior of the plant, then the
feed forward computations will be of questionable value.
The second criterion is more subtle. Suppose a plant has a disturbance dead time that is
shorter than the process dead time. Further suppose a disturbance occurs and the feed forward
controller instantly responds with appropriate control actions. Because of the dead time difference,
the disturbance will reach the measured process variable before the controller manipulations, even
though both happen at the same time in this example. The disturbance will already be disrupting the
process before the first disturbance rejection control actions even arrive.
At the limit, the control actions need to arrive at the same time as the disturbance (equal dead
times) for reasonable disturbance rejection. If conditions are such that control actions can arrive first
(if the process dead time is shorter), the feed forward controller can be most effective in rejecting a
measured disturbance.

19.5 Feed Forward Control Theory

Obtaining Process and Disturbance Models

Laplace transforms are used in the next sections to make the derivation of the feed forward element
mathematically correct. The presentation uses basic math rules, so even if you are not familiar with
Laplace transforms, the logic of the presentation will (hopefully) make sense.
Methods for obtaining a process model have been discussed in detail in previous chapters. In
short, a process data set is generated by stepping, pulsing or otherwise perturbing the controller
output signal, u(t), and recording the measured variable, y(t), as the process responds. The process
should initially be at steady state at the design level of operation and the response of the measured
process variable should clearly dominate the noise in the measurement signal.
A process model is obtained by fitting the data with a linear dynamic equation ranging from
first order without dead time (FO) up through second order with dead time and lead time (SOPDT w/
L). If we call this process model of unspecified form GP(s), then in the Laplace domain we can say

Y ( s ) = GP ( s)U ( s ) (19.2)

198
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 That is, given knowledge of the controller output, Eq. 19.2 permits the expected behavior of
the measured process variable to be computed. This equation can rearranged to say that, given a
change in the measured process variable, the controller output signal sequence that would cause that
change can be back-calculated:
U ( s ) = [1/ GP ( s ) ]Y ( s ) (19.3)

The disturbance model is created in the identical fashion to the process model except it is the
disturbance variable, d(t), that must be perturbed in some fashion. Since disturbance variables are
beyond the control of a loop (which is what makes them disturbances), it is not always possible to
step or pulse the disturbance at will. Hence, it may be necessary to wait for moments of opportunity
and collect data during an actual disturbance event. An alternative is to sift through data logs to find a
disturbance event that produced data suitable for model fitting. In any case, a linear model ranging
from FO up through SOPDT w/ L is fit to this data. In the Laplace domain, this disturbance model,
GD(s), is expressed:
Y ( s ) = GD ( s) D( s ) (19.4)

So with knowledge of changes in the disturbance variable (provided by the added sensor), Eq. 19.4
permits the impact profile of each disturbance on the measured process variable to be computed.

Deriving the Feed Forward Element

Once online, the new sensor measures the disturbance variable and the signal is fed through Eq. 19.5,
the disturbance model of the feed forward element. This model continually updates Ydisturb as labeled
in Fig. 19.2, which is a prediction of the impact profile (how and when the measured process variable
will be impacted by the disturbance) as:

Ydisturb ( s ) = GD ( s ) D ( s ) (19.5)

This predicted sequence of disruption, Ydisturb(s), is then fed to Eq. 19.6, the process model of
the feed forward element, to back calculate a series of control actions, Ufeedforward(s). The result is a
sequence of control actions that will cause the measured process variable to behave just like this
predicted disturbance impact profile (we add a negative sign in Eq. 19.8 to make the feed forward
actions oppose the disturbance):
U feedforward ( s ) = [1/ GP ( s ) ]Ydisturb ( s ) (19.6)

Substituting Eq. 19.5 into Eq. 19.6 completes the “disturbance model divided by process model” feed
forward element:
U feedforward ( s ) = [GD ( s ) / GP ( s )] D ( s ) (19.7)

Eq. 19.7 computes a series of control actions that cause the measured process variable to
duplicate the predicted disturbance impact profile. A negative sign is added so the feed forward
controller acts in a manner opposite to this, thus canceling the impact of the disturbance. This
negative feed forward action is combined with the traditional feedback control output signal to yield
the total controller output:
U total ( s ) = U feedback ( s ) − U feedforward ( s ) (19.8)

199
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
19.6 Limits on the Form of the Feed Forward Model
Highest Model Order
The linear model range of FO (first order) up through SOPDT w/ L (second order plus dead time with
lead time) are mentioned in the preceding discussion because they are the choices available in LOOP-
PRO. Feed forward theory permits third, fourth and higher order linear models to be used in the feed
forward element.

When working with a real plant, obtaining data that will yield accurate values for the three
parameters of a FOPDT model is surprisingly challenging, and an accurate FOPDT model is usually
capable of providing effective feed forward disturbance rejection. Obtaining a data set so rich in
dynamic information and absent of disturbance influences that it can yield accurate values for the five
parameters of a SOPDT w/ L model is very difficult in real applications and pushes practical
implementation near the limit. Only the rarest of applications would benefit from a model more
complex than the choices available in LOOP-PRO.
A SOPDT w/ L process model in the time domain has the following form:

d 2 y (t ) d y (t ) d u (t − θ P )
τ P1 τ P 2 + ( τ P1 + τ P 2 ) + y (t ) = K P u (t[− θ P ) + τ PL ] (19.9)
dt2 dt dt

In the Laplace domain this same process model is expressed as follows:

K P (τ PL s + 1) e − θ P s
Y (s) = U ( s) (19.10)
( τP1s + 1)(τ P 2 s + 1)

The model of Eq. 19.10 is identical to that of Eq. 19.9. Both are linear ordinary differential equations
with constant coefficients. If a SOPDT w/ L model is used for both the process and disturbance
models in Eq. 19.7, the feed forward element becomes
⎧⎛ K ⎫
U feedforward ( s ) = ⎪⎨⎜⎜ D ⎞⎟ ⎡⎢ ( τ P1s + 1)( τ P 2 s + 1)( τ DL s + 1) ⎤⎥ e− (θ D −θ P ) s ⎪⎬ D( s) (19.11)
⎪⎩⎝ K P ⎟⎠ ⎣⎢ ( τ D1s + 1)( τ D 2 s + 1)( τ PL s + 1) ⎦⎥ ⎪⎭

Notice that both process time constants are in the numerator along with the disturbance lead time and
the both disturbance time constants are in the denominator along with the process lead time. This
form is referred to as a dynamic feed forward element because the time dependent variables, including
time constants, lead times and dead times, are included in the feed forward computation.

Dead Time Difference

We discussed previously that the process dead time must be shorter than the disturbance dead time for
effective feed forward disturbance rejection. In fact, LOOP-PRO does not even permit entry of a
larger process dead time. This is not a limitation of the software but a requirement of the math.
If θP > θD, then Eq. 19.11 would be required to compute corrective control actions that are to
be implemented before the disturbance change was even first detected. Knowledge of the future, an
outcome possible in the sterile world of mathematics, is simply not possible in the real world of
process control. Formulations that require unknowable information or impossible actions are said to
be unrealizable.

200
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Suppose for a particular application, you determine that the disturbance dead time is indeed
shorter than the process dead time. Best practice is to arbitrarily set the process dead time equal to the
disturbance dead time when entering values into the feed forward input form. Disturbance rejection
performance will suffer but at least the feed forward calculations will yield control actions that make
physical sense.

Model Order Ratio

A second way an impossible or unrealizable feed forward model can result is through mismatching of
the time constant and lead time terms for the process and disturbance models. As mentioned
previously, both process time constants are in the numerator along with the disturbance lead time and
the both disturbance time constants are in the denominator along with the process lead time.
A model that is physically realizable requires that the number of “time” terms in the
numerator be less than or equal to the number of such terms in the denominator. Again, this is not a
limitation of LOOP-PRO but a consequence that the mathematics must compute actions that are
physically possible to implement. There are many ways to create a realizable feed forward element. A
few examples of permitted and prohibited forms are illustrated below.

Permitted: FO or FOPDT Process model; SO or SOPDT Disturbance Model

⎧⎛ K (τP1s + 1) ⎫
U feedforward ( s ) = ⎪⎨⎜⎜ D ⎞⎟ ⎡⎢ ⎤
⎥ e− (θ D −θ P ) s ⎪⎬ D( s) (19.12)
⎪⎩⎝ K P ⎠ ⎢⎣ ( τ D1s + 1)( τ D 2 s + 1) ⎥⎦
⎟ ⎪⎭

Permitted: FO or FOPDT Process model; SO w/ L or SOPDT w/L Disturbance Model

⎧⎛ K ⎫
U feedforward ( s ) = ⎪⎨⎜⎜ D ⎞⎟ ⎡⎢ ( τ P1s + 1)( τ DL s + 1) ⎤⎥ e− (θ D −θ P ) s ⎪⎬ D( s) (19.13)
⎪⎩⎝ K P ⎟⎠ ⎢⎣ ( τ D1s + 1)( τ D 2 s + 1) ⎥⎦ ⎪⎭

Permitted: SO or SOPDT Process model; SO or SOPDT Disturbance Model

⎧⎛ K ⎫
U feedforward ( s ) = ⎪⎨⎜⎜ D ⎞⎟ ⎡⎢ ( τ P1s + 1)( τ P 2 s + 1) ⎤⎥ e− (θ D −θ P ) s ⎪⎬ D( s) (19.14)
⎪⎩⎝ K P ⎟⎠ ⎢⎣ ( τ D1s + 1)( τ D 2 s + 1) ⎥⎦ ⎪⎭

NOT Permitted: SO or SOPDT Process model; SO w/ L or SOPDT w/L Disturbance Model

⎧⎛ K ⎫
U feedforward ( s ) = ⎪⎨⎜⎜ D ⎞⎟ ⎡⎢ ( τ P1s + 1)( τ P 2 s + 1)(τ DL s + 1) ⎤⎥ e− (θ D −θ P ) s ⎪⎬ D( s) (19.15)
⎪⎩⎝ K P ⎟⎠ ⎢⎣ (τD1s + 1)(τD 2s + 1) ⎥⎦ ⎪⎭

This last form is unrealizable because there are more “time” terms in the numerator than denominator,
requiring computation of physically impossible actions.

Positive Process Lead Required

The model of Eq. 19.11, called a Laplace domain transfer function, can also be expressed as a rather
complicated linear ordinary differential equation with constant coefficients in the time domain. The
solution of this differential equation is unstable (diverges) if any of the “time” terms in the
denominator are negative.
There is no concern that τD1 or τD2 in the denominator will be negative as time constants by
definition are always positive. The process lead time, τPL, would be negative if the measured process
variable displays an inverse (nonminimum phase) response to a step change in the controller output
signal. But since τPL is in the denominator, it cannot be negative or the feed forward computations
will go unstable.

201
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 If your process displays an inverse response and thus yields a negative τPL after a model fit,
best practice is to use a FOPDT model and approximate the inverse portion of the response as a long
dead time. Interestingly, the disturbance can have a negative lead element such as that exhibited by
the heat exchanger because the disturbance lead term is in the numerator of the feed forward element.

19.7 Feed Forward Disturbance Rejection in the Jacketed Reactor

The control objective of this investigation is the same as that used in the cascade study of Chapter 18.
We seek to maintain the reactor exit stream temperature at set point by rejecting disturbances caused
by changes in the cooling jacket inlet temperature. The design level of operation is a measured reactor
exit stream temperature of 86°C. The cooling jacket inlet temperature (disturbance), normally at
46°C, is known to spike as low as 40°C. Open loop studies establish that a controller output of 50%
causes the reactor to steady at the design measured exit stream temperature of 86°C when the cooling
jacket inlet temperature is at its normal (design) value of 46°C.

Single Loop Disturbance Rejection

As detailed in cascade control chapter and shown again in Fig. 19.5, the controller output is stepped
from the design value of 50% up to 53%, then down to 47%, and finally back to 50%, forcing the
measured reactor exit stream temperature to exhibit a clear response after each controller output step
that dominates any measurement noise.

Modeling Controller Output to Process Variable Dynamics

Model: First Order Plus Dead Time (FOPDT) File Name: Process.txt
Process Variable

87

86

85 process data and

FOPDT model fit
84

54 controller output
Controller Output

52 steps in manual mode

50
48
46

0 5 10 15 20 25
Time

Gain (K) = -0.36, Time Constant (T1) = 1.58, Dead Time (TD) = 0.88
SSE: 0.4316
Figure 19.5 − FOPDT model fit of single loop controller output to measured process variable data

A FOPDT fit of the dynamic process data as computed by Design Tools yields the model parameters:
Process Gain, KP = − 0.36 °C/% (19.16a)
Time Constant, τP = 1.58 min (19.16b)
Dead Time, θP = 0.88 min (19.16c)

202
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
These parameters were used in the standard IMC tuning correlation to obtain the PI tuning values:
Controller Gain, KC = − 2.7 %/°C (19.17a)
Reset Time, τI = 1.6 min (19.17b)

The disturbance rejection performance of the single loop PI controller using these tuning
parameters is shown in Fig. 19.6. The label at the upper right of the plot confirms that the PI controller
has a direct acting proportional term, an anti-reset windup integral term, and a derivative term that is
off.

Disturbance Rejection Performance of Single Loop PI Controller

Process: Single Loop Jacketed Reactor Controller: PID ( P= DA, I= ARW, D= off )
88
PV/Setpoint

87
86
85
84 constant
83 reactor exit set point
55 temperature
Contr Output

50
45
40
35
30

disturbance
Disturbance

45
variable steps
40

65 70 75 80 85 90 95 100
Time (mins)
Tuning: Gain = -2.70, Reset Time = 1.60, Deriv Time = 0.0, Sample Time = 1.00

Figure 19.6 − Disturbance rejection in the jacketed reactor under single loop PI control
Process variable deviations reach 2.5°C (compare to Fig. 19.9)

The measured reactor exit stream temperature is initially steady at the design set point value
of 86°C in Fig. 19.6. To test the controller, the cooling jacket inlet temperature is stepped from its
design value of 46°C down to 40°C and back again. As shown, the single loop PI controller is able to
maintain reactor exit stream temperature near the constant set point of 86°C, with deviations ranging
as high as 2.5°C during the event.

Feed Forward Disturbance Rejection

To construct a feed forward controller, the cooling jacket inlet temperature (the disturbance) is
measured as shown in Fig. 19.7. The signal from this disturbance temperature sensor is sent to a feed
forward element comprised of a process model and disturbance model.
A FOPDT model reasonably describe the dynamics of the controller output to measured
process variable as shown in Fig. 19.5. The parameter values of Eq. 19.16 are thus used to construct
the process model of the feed forward element.

203
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 FF

Figure 19.7 − Jacketed reactor with disturbance sensor for feed forward control

Generating disturbance driven data can be problematic on real processes if the disturbance
variable cannot be manipulated at will. LOOP-PRO permits such disturbance manipulations. As
shown in Fig. 19.8, the cooling jacket inlet temperature is stepped from the design value of 46°C up
to 49°C, down to 43°C and back to 46°C. The measured reactor exit stream temperature exhibits a
clear response after each step that dominates measurement noise.

Modeling Disturbance to Process Variable Dynamics

Model: First Order Plus Dead Time (FOPDT) File Name: Disturb.txt
Process Variable

88

86

84 process data and

FOPDT model fit
82

50 disturbance variable
steps in manual mode
Disturbance

48
46
44
42

30 35 40 45 50 55
Time

Gain (K) = 0.95, Time Constant (T1) = 1.92, Dead Time (TD) = 1.30
SSE: 9.81

Figure 19.8 − FOPDT model fit of disturbance to measured process variable data

204
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
A FOPDT model reasonably describes the disturbance to measured process variable dynamics as
shown in Fig. 19.8. The disturbance model parameters used to construct the disturbance model of the
feed forward element are thus:
o
C of reactor exit stream temperature
Disturbance Gain, KD = 0.95 o (19.18a)
C of cooling jacket inlet temperature
Disturbance Time Constant, τD = 1.92 min (19.18b)

Disturbance Dead Time, θD = 1.30 min (19.18c)

Using the process model of Eq. 19.16 and the disturbance model of Eq. 19.18, the feed forward
element is constructed:

⎧⎪ 0.95 ⎞ ⎡ (1.58s + 1) ⎤ − (1.30−0.88) s ⎫⎪

U feedforward ( s ) = ⎨⎛⎜ ⎟⎢ ⎥e ⎬ D( s ) (19.19)
⎪⎩⎝ −0.36 ⎠ ⎣ (1.92 s + 1) ⎦ ⎪⎭

Equation 19.19 is physically realizable because there are as many “time” terms in the denominator as
in the numerator and because θD > θP.
This feed forward element is combined with the PI feedback controller used earlier in this
chapter for the single loop PI controller study (see Fig. 19.5) to yield a feed forward with feedback
trim architecture. Figure 19.9 shows the disturbance rejection performance of this controller. Just as
in Fig. 19.5, the measured process variable (reactor exit stream temperature) is initialized at the
design set point value of 86°C. To test the controller, the cooling jacket inlet temperature is stepped
from its design value of 46°C down to 40°C and back again.

Disturbance Rejection Performance of PI With Feed Forward

Process: Single Loop Jacketed Reactor Controller: PID with Feed Forward
88
PV/Setpoint

87
86
85
constant
84
rapid control action
83
set point
55 from feed forward
Contr Output

50
45
40
35
30

disturbance
Disturbance

45
variable steps
40

200 205 210 215 220 225 230 235

Time (mins)
Tuning: Gain = -2.70, Reset Time = 1.60, Deriv Time = 0.0, Sample Time = 1.00
Process Model: Gain(Kp) = -0.36, T1 = 1.58, T2 = 0.0, TD = 0.88, TL = 0.0
Disturbance Model: Gain(Kd) = 0.95, T1 = 1.92, T2 = 0.0, TD = 1.30, TL = 0.0

Figure 19.9 − Disturbance rejection in jacketed reactor under feed forward with feedback trim
Process variable deviations reach 1.0°C (compare to Fig. 19.5)

The feed controller with feedback trim architecture performs better than the single loop case
in maintaining the reactor product temperature near the constant set point of 86°C. Specifically, while

205
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
the measured reactor exit temperature deviations for the single loop PI controller range as high as
high as 2.5°C during the event, this advanced architecture limits the maximum deviation to less than
1.0°C.
The improved performance did not come free, however, as an additional sensor, controller
and tuning effort were required. As shown in Fig. 19.9, the feed forward controller initiated rapid
compensating controller just after the disturbance event to minimize its impact on the measured
process variable. Perfect disturbance rejection was not achieved because the FOPDT models only
approximate the higher order and nonlinear behavior of the jacketed reactor process.

19.8 Static Feed Forward Control

The preceding discussion detailed the implementation of a dynamic feed forward controller. A
dynamic feed forward controller employs the time dependent model variables in the feed forward
computation when calculating a corrective action. These include the time constant and lead time
terms that describe the speed of response of the process, and the dead time terms that describe the
time delay before the response of the process begins.
A static feed forward controller does not consider the time dependent information in the feed
forward computation. Only the size of response as described by the ratio of the steady state process
gains is used. Hence, the general dynamic form of Eq. 19.11 reduces rather substantially to the static
feed forward equation:
⎛K ⎞
U feedforward ( s ) = ⎜⎜ D ⎟⎟ D ( s ) (19.20)
⎝ KP ⎠
As Eq. 19.20 shows, the feed forward control action is simply a constant multiplied by the current
value of the disturbance signal. The constant is computed as the ratio of the disturbance model gain
divided by the process model gain.
The benefit of static feed forward is that many commercial controllers are sophisticated
enough to permit implementation of this simple algorithm, and hence, there is no need for
programming models on the control computer. It should come as no surprise, however, that the loss of
time dependent information in computing corrective actions implies some degradation in disturbance
rejection performance.
Figure 19.10 shows a side-by-side comparison of disturbance rejection performance in the
jacketed reactor for three controller architectures. For all three cases, the jacketed reactor inlet
temperature disturbance is stepped from 46°C down to 40°C and back again.
The left most portion of the plot shows disturbance rejection performance using a lone PI
controller with no feed forward element. The center of the plot shows rejection of the same
disturbance for a PI controller with a static feed forward element. The right most portion shows the
performance for a PI controller with a full dynamic feed forward element.
Disturbance rejection performance improves from left to right in Fig. 19.10 because the
controller is provided more information in the form of models that describe the behavior of the
process and disturbance. And while we are cautious to avoid making broad assumptions from this
single example, it is safe to assume that this general trend will hold true in a broad range of
applications.

206
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Comparing Feedback with Static and Dynamic Feed Forward
Process: Single Loop Jacketed Reactor Controller: PID with Feed Forward
88

Disturbance Contr Output PV/Setpoint

87
86
85
84
83 feedback only static f. f. dynamic f. f.
55
50
45
40
35
30
disturbance variable steps
45

40

20 30 70 40 80 50
90 60
100 110 120 130
Time (mins)
Tuning: Gain = -2.70, Reset Time = 1.60, Deriv Time = 0.0, Sample Time = 1.00
Process Model: Gain(Kp) = -0.36, T1 = 1.58, T2 = 0.0, TL = 0.0, TD = 0.88
Disturbance Model: Gain(Kd) = 0.95, T1 = 1.92, T2 = 0.0, TL = 0.0, TD = 1.30

Figure 19.10 − Comparing set point tracking of single feedback loop to that of the cascade

19.9 Set Point Tracking Comparison of Single Loop and Feed Forward Control
The feed forward with feedback trim architecture does not provide benefit in tracking set point
changes and this is illustrated in Fig. 19.11 for the jacketed reactor. The trace to the left of the plot
shows the set point tracking performance of the single loop PI controller while the trace to the right
shows the performance when the feed forward controller is added.

Set Point Tracking Performance: Feedback vs Feed Forward

Process: Single Loop Jacketed Reactor Controller: PID with Feed Forward

90
PV/Setpoint

88
86
84
feedback only feed forward with feedback trim
60
Controller Output

50

40

10 20 30 40 50 60 70
Time (mins)
Tuning: Gain = -2.70, Reset Time = 1.60, Deriv Time = 0.0, Sample Time = 1.00
Process Model: Gain(Kp) = -0.36, T1 = 1.58, T2 = 0.0, TL = 0.0, TD = 0.88
Disturbance Model: Gain(Kd) = 0.95, T1 = 1.92, T2 = 0.0, TL = 0.0, TD = 1.30

Figure 19.11 − Comparing set point tracking of single feedback loop to that of the cascade

207
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 As shown in Fig. 19.11, both architectures provide identical performance in tracking set point
changes. This makes sense because for the feed forward with feedback trim architecture shown in Fig.
19.2, as long as the disturbance is constant, the feedback loop works alone to control the process.
Since the same feedback loop PI controller tuning was used in both cases, we expect the set point
tracking performance to be identical. Based on this observation, be sure your objective is the rejection
of one particular disturbance before considering a feed forward architecture.

208
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 20. Multivariable Controller Interaction and Loop Decoupling

20.1 Multivariable Process Control

Except for cascade control, the control systems discussed to this point have all been single-loop
architectures. That is, the control loops have had a single controller output signal that adjusts a single
manipulated variable to impact a single measured process variable. In many real processes, however,
there are two or more manipulated variables that when adjusted, each impact more than one measured
process variable. These multivariable process control systems, with their loop interactions, present a
new level of complexity for controller design.
LOOP-PRO's distillation column process, shown in Fig. 20.1, offers such a multivariable
control challenge. It is a binary distillation column that separates benzene and toluene. The objective is to
send a high percentage of benzene (and thus low percentage of toluene) out the top stream, and a low
percentage of benzene (and thus high percentage of toluene) out the bottom stream. The column dynamic
model employs tray-by-tray mass and energy calculations similar to that proposed by McCune and Gallier
[ISA Transactions, 12, 193, (1973)].

Figure 20.1 - Distillation column has top and bottom control loops that interact

To achieve the desired benzene-toluene separation, the top controller manipulates the reflux rate
to control the top (distillate) composition. The bottom controller adjusts the rate of steam to the reboiler to
control the bottom stream composition. Any change in feed rate to the column acts as a disturbance to the
process. With two manipulated variables and two measured process variables, this is commonly called a
two-by-two multi-input multi-output (2x2 MIMO) process.

209
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 To illustrate the loop interaction in this MIMO process, suppose the composition (or purity) of
benzene in the top stream is below set point. The top controller will respond by increasing the flow rate of
cold reflux into the column. This increased reflux flow will indeed increase the purity of benzene in the
top stream. However, the additional cold liquid will work its way down the column, begin to cool the
bottom, and as a result permit more benzene to flow out the bottom stream.
As the bottom composition moves off set point and produces a controller error, the bottom
controller will compensate by increasing the flow of steam into the reboiler to heat up the bottom of the
column. While this works to restore the desired purity to the bottom stream, unfortunately, it also results
in an increase of hot vapors traveling up the column that eventually will cause the top of the column to
begin to heat up.
As the top of the column heats up, the purity of benzene in the top stream again becomes too low.
In response, the top controller compensates by further increasing the flow of cold reflux into the top of the
column. The controller “fight," or multivariable interaction, begins.

20.2 Control Loop Interaction

Decouplers are essentially feed forward elements designed to reduce controller interaction in MIMO
processes. The only difference between a feed forward element and a decoupler is that with a decoupler,
the disturbance to be rejected is the action of another control loop on the process.

TOP
top
ysetpoint utop PROCESS ytop
- CONTROL
Top Comp
+ Top
response to
+
+
Composition
Top CO
GTT (s)

INTERACT
Top Comp
response to
Bottom CO
GTB (s)

INTERACT
Bottom Comp.
response to
Top CO
GBT (s)

BOTTOM
CONTROL PROCESS
+
+- Bottom Bottom Comp. +
bottom
ysetpoint ubottom response to ybottom
Composition
Bottom CO
GBB (s)

Figure 20.2 - Block diagram of top and bottom distillation control loops with “cross loop”
interaction

210
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
As shown in Fig 20.2 for the distillation column, the "cross-loop disturbance" of the top stream
composition is the bottom controller manipulations of the steam flow rate. The cross-loop disturbance of
the bottom composition is the top controller manipulations of the cold reflux rate to the top of the column.

20.3 Decouplers are Feed Forward Controllers

Recall that a feed forward controller gains advantage by using a sensor to directly measure a
disturbance while it is still distant from the measured process variable. A feed forward element
receives this measurement and uses it to compute preemptive control actions designed to counter the
impact of the disturbance just as it reaches the measured process variable.
A decoupler is a feed forward element where the disturbance is the cross-loop controller signal.
Thus, a decoupler is comprised of a process model and a cross-loop disturbance model
• The cross-loop disturbance model receives the cross-loop controller signal and predicts an
“impact profile,” or when and by how much the process variable will be impacted.
• Given this predicted sequence of disruption, the process model then back calculates a series
of control actions that will exactly counteract the cross-loop disturbance as it arrives so the
measured process variable remains constant at set point.
Implementation of a decoupler does not require an added sensor to measure the disturbance
because the current value of the cross-loop controller signal is readily available for use by the
decoupler. However, developing and programming the dynamic process and cross-loop disturbance
models is required for decoupler implementation.
To better understand the decoupling computation and its relation to feed forward control, we
focus here on the top control loop of the distillation column. Though we concentrate on the logic of
the computation in this discussion, Laplace transforms are used to make the derivation of the
decoupler mathematically correct.
To develop a process model for the top control loop, a data set is generated by stepping,
pulsing or otherwise perturbing the top controller output signal, utop(t), and recording the measured
variable, ytop(t), as the process responds. Both loops should be in manual during this data collection
exercise. As always, the process should initially be at steady state at the design level of operation and
the response of the top process variable should clearly dominate the noise in the measurement signal.
This process data set is fit with a linear dynamic model ranging from first order without dead
time (FO) up through second order with dead time and lead time (SOPDT w/ L). If we call this top
composition response to top controller output model GTT(s), then in the Laplace domain we can say:

Ytop(s) = GTT(s)Utop(s) (20.1)

That is, given knowledge of the top controller output, Eq. 20.1 permits the expected behavior of the
top process variable to be computed. This equation can be rearranged to say that, given a change in
the top process variable, the top controller output signal sequence that would cause that change can be
back-calculated as:
Utop(s) = [1/GTT(s)] Ytop(s) (20.2)

The cross-loop disturbance model is created by perturbing the cross-loop controller output,
which in this case is the bottom controller, ubottom(t), and recording how ytop(t) responds. Again, both
controllers should be in manual during this data collection exercise. We fit a linear dynamic model to
this cross-loop disturbance data and call the resulting top composition response to bottom controller
output model GTB(s).

211
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
In the Laplace domain we can then say:

Ytop(s) = GTB(s)Ubottom(s) (20.3)

Hence, with knowledge of controller signal ubottom(t), Eq. 20.3 lets us compute the impact profile of
this cross loop disturbance on the top process variable, ytop(t).

TOP
top top top
utotal
ysetpoint ufeedback PROCESS ytop
- CONTROL
++ Top Comp.
+ Top
response to
+
+
Composition
Top CO
GTT (s)
TOP
DECOUPLER
Top Comp. INTERACT
decoupled from top
udecouple Top Comp.
Bottom CO response to
Dtop (s) Bottom CO

GTB ( s)

BOTTOM INTERACT
bottom Bottom Comp.
DECOUPLER udecouple
Bottom Comp. response to
decoupled from Top CO
Top CO GBT ( s)
Dbottom (s)

BOTTOM
CONTROL PROCESS
+
bottom +- Bottom ++ Bottom Comp. +
ybottom
ysetpoint Composition bottom
ufeedback bottom
utotal
response to
Bottom CO
GBB ( s)

Figure 20.3 - Block diagram of top and bottom distillation control loops
with cross loop interaction and decouplers

We now have the models needed to construct the decoupler Dtop(s) as shown in Fig 20.3. As
bottom
the bottom controller makes changes in u feedback (t ) , the signal is sent to Eq. 20.4, which is the cross-

loop disturbance model portion of the top decoupler. This model continually updates y *top (t ) , which
is a prediction of the impact profile or how and when the top process variable will be impacted by the
actions of the bottom controller. That is:
* bottom
Ytop ( s ) = GTB ( s) U feedback (t ) (20.4)

212
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 This predicted sequence of disruption, y *top (t ) , is then fed to Eq. 20.5, the process model
top
portion of the top decoupler, to back calculate a series of control actions, u decouple (t ) . This results in a
sequence of control actions that will cause the top process variable to behave just like the predicted
cross-loop disturbance impact profile (we add a negative sign in Eq. 20.7 to make the decoupling
actions oppose the disturbance):

top ⎡ 1 ⎤ *
U decouple (s) = ⎢ Ytop ( s ) (20.5)
⎣ GTT ( s ) ⎥⎦

Substituting Eq. 20.4 into Eq. 20.5 completes the top decoupler, Dtop(s), that as shown in Eq. 20.6,
bottom
computes decoupling actions based on knowledge of the bottom controller actions, u feedback (t ) :

top ⎡ G ( s) ⎤ bottom bottom

U decouple ( s ) = ⎢ TB ⎥⎦ U feedback ( s ) = D top ( s) U feedback ( s ) (20.6)
⎣ GTT ( s)

Eq. 20.6 computes a series of control actions that cause the top process variable to duplicate
the predicted cross-loop disturbance impact profile. A negative sign is added so the decoupler acts in
a manner opposite to this, thus canceling the impact of the cross loop disturbance. This negative
decoupling action is combined with the traditional feedback control signal of the top controller to
yield the total controller output:
top top top
U total ( s) = U feedback ( s ) − U decouple (s) (20.7)

The bottom decoupler can be derived in an analogous manner as:

bottom ⎡ G ( s ) ⎤ top top

U decouple ( s) = ⎢ BT ⎥ U feedback ( s) = Dbottom ( s) U feedback (s) (20.8)
G (
⎣ BB ⎦ s )

And thus, to implement a 2x2 top and bottom decoupler, we need to develop four dynamic
process models. For the top decoupler, the FOPDT process and disturbance model parameters entered
are:
Process model: GTT(s) = top composition response to top controller output model
Disturbance model: GTB(s) = top composition response to bottom controller output model

For the bottom decoupler, the FOPDT process and disturbance model parameters entered are:
Process model: GBB(s) = bottom composition response to bottom controller output
Disturbance model: GBT(s) = bottom composition response to top controller output

These models are used to develop two decouplers that must be programmed into the control computer
and are implement as shown in Fig 20.3:

⎡ GTB ( s ) ⎤
Top decoupler: D top ( s ) = (20.9)
⎢⎣ GTT ( s ) ⎥⎦

213
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 ⎡ G ( s) ⎤
Bottom decoupler: D bottom ( s ) = ⎢ BT (20.10)
⎣ GBB ( s) ⎥⎦

20.4 Distillation Study - Interacting Control Loops

We will study decouplers in the next section. Here we explore the multivariable interactions that
occur between the top and bottom controllers on the distillation column when they are designed and
implemented as independent or stand-alone loops.
The control objective is the design and tuning of a PI controller that can track set point steps
in the top composition from 92% and 94% while the bottom composition remains constant at 1.5%.
The design level for this study includes a feed flow rate to the column of 547 Kg/min (note that this is
not the default startup value) and the steady state operating conditions:

Top Î utop(t) = 52 % ytop(t) = 92 %

Bottom Î ubottom(t) = 48 % ybottom(t) = 1.5 %

Fig 20.1 shows the distillation column at these design conditions, only in that figure the control loops
are in automatic.
Before tuning the loops, we investigate the dynamic behavior of the column using open loop
doublet tests. The top controller output is stepped from 52% up to 55%, down to 49% and back to
52%. After the process settles, the bottom controller is stepped from 48% up to 51%, down to 45%
and back to 48%. Fig. 20.4 shows the process variable responses to these tests.
Open Loop Response Reveals Very Nonlinear Behavior
Process: Distillation Column Top: Manual Mode / Bot: Manual Mode

94
Top % %

92
Distillate

90
88

56
Top CO

52

48
2.5
top CO step up
Bottoms %

2.0 gives large KP

1.5
1.0

52
bottom CO step down
gives large KP
Bot CO

48

44
500 1000 1500 2000 2500 3000
Time (mins)

Figure 20.4 - Open loop step tests on the distillation column's top and bottom controller

The upper most trace of Fig. 20.4 shows that the top composition process variable is mildly
nonlinear. That is, ytop(t) responds both up and down from the design level of operation in roughly the
same magnitudes whether forced by utop(t) or ubottom(t).
The bottom composition loop, on the other hand, is extremely nonlinear. In particular, the
response of the bottom process variable is three times larger when ybottom(t) is increasing from the
design level of operation compared to when it is decreasing. As shown in Fig. 20.4, this is true
whether it is utop(t) or ubottom(t) that is forcing the response.

214
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 This extreme nonlinear behavior provides added control challenges beyond the loop
interaction issue. To meet the challenge, we take advantage of the information in the above plot to
design the dynamic tests for modeling and tuning. Rather than the standard doublet, we will pulse
each variable in one direction only. We choose the direction that yields parameters leading to the
most stable controller design.
Recall that tuning correlations compute controller gain, KC, as proportional to the inverse of
process gain, KP. So for a conservative design, we want to use large values of KP in the correlations to
obtain small (and thus less aggressive) values of KC. Exploiting the information contained in Fig.
13.4, we achieve the largest values of KP by pulsing utop(t) up and ubottom(t) down when generating our
dynamic test data.

Top Composition Control

We perform a pulse test to generate dynamic process data for the design of the top composition
controller. As shown in Fig. 13.5, the top controller output, utop(t), is stepped from its design value of
52% up to 54% and back. The bottom controller is in manual mode during this test. Design Tools is
used to fit a FOPDT model to the data. The process data and model fit are shown in Fig. 20.5.

Top
TopProcess:
Process: Top Composition
Distillate Response
Composition to to
Response Top
TopCO
CO
Model: First Order Plus Dead Time (FOPDT) File Name: top.txt
Top Composition
Composition

94

93
Distillate

92 distillate responseresponse
top composition data
and
dataFOPDT
and FOPDTmodelmodel
fit fit
55

54
Top CO

53

52

51
0 100 200 300 400 500 600 700 800 900
Time

Gain (K) = 1.12, Time Constant (T1) = 62.16, Dead Time (TD) = 24.26
SSE: 0.1661

Figure 20.5 - Top composition response to top controller output pulse

and FOPDT fit of process data

The FOPDT dynamic model provides a good approximation of the process data. Hence,
GDT(s), the top composition response to top controller output model has the parameters:
Process Gain: KP, TT = 1.1 %/%
Overall Time Constant: τP, TT = 62 minutes
Apparent Dead Time: θP, TT = 24 minutes

215
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
The FOPDT dynamic model parameters are used in the IMC correlation to obtain initial estimates for
PI controller tuning. Recalling that the standard IMC correlation uses a closed loop time constant, τC,
as the larger of 0.1τP or 0.8θP, then Design Tools computes the PI tuning parameters:

Controller Gain, KC, top = 1.3 %/%

Reset Time, τI, top = 62 minutes

Returning to the distillation process, a PID controller is selected and these tuning values are entered.
We turn derivative action off on the controller design menu, resulting in the desired PI controller
form.

Top
Top Loop Performs
Distillate Well Well
Loop Peforms WithWith
Bottom Loop
Bottom in Manual
Loop Mode
in Manual Mode
Process: Distillation Column Top: PID ( P= RA, I= ARW, D= off ) / Bot: Manual Mode
Top % %

94
Distillate

93
92

56 good set point tracking performance…

Top CO

54
52
50

2.0
Bottoms %

1.8
1.6
1.4

49
but bottoms process variable
bottom process variable isisdriven
driven
Bot CO

from the design level of operation

48
47 the design level of operation
46
200 300 400 500 600 700 800
Time (mins)
Tuning: Gain = 1.30, Reset Time = 62.0, Deriv Time = 0.0, Sample Time = 1.00

Figure 20.6 - Set point tracking capability of top loop under PI control
when bottom loop is in manual mode

Controller performance is tested in tracking set point steps between 92% and 94% as shown
in Fig 20.6. The bottom loop remains in manual mode during the evaluation. The top loop shows
desirable set point tracking performance, with a rapid response, minimal overshoot and rapid decay.
However, the corresponding movement in the bottom composition from the design level of operation
shown in Fig. 20.6 highlights the need for a second controller on the bottom composition stream to
properly operate the distillation column.

Bottom Control and Loop Interaction

The design of the bottom PI controller is analogous to that of the top controller. A pulse test is
performed as shown in Fig. 20.7 by stepping the bottom controller output, ubottom(t), from its design
value of 48% down to 46% and back. The top controller is in manual mode during this test. Design
Tools is used to fit a FOPDT model to the data and the model fit is also shown in Fig. 20.7.

216
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 The FOPDT dynamic model provides a good approximation of the process data. Hence,
GBB(s), the bottom composition response to bottom controller output model has the parameters:
Process Gain: KP, BB = −0.22 %/%
Overall Time Constant: τP, BB = 53 minutes
Apparent Dead Time: θP, BB = 14 minutes

The FOPDT dynamic model parameters are used in the IMC correlation to obtain initial estimates for
PI controller tuning. Using the standard IMC correlation, Design Tools computes the PI tuning
parameters:
Controller Gain, KC, bottom = −9.7 %/%
Reset Time, τI, bottom = 53 minutes

Returning to the distillation process, a PID controller is selected and the above tuning values are
entered. We turn derivative action off on the controller design menu, resulting in the desired PI
controller form.

Bottom
Bottom Process:Bottoms
Process: Bottom Composition
Composition Response to Bottom
Response CO CO
to Bottom
Model: First Order Plus Dead Time (FOPDT) File Name: bottom.txt

2.0
Bottoms Composition

1.8

1.6
bottom response data
1.4 and FOPDT model fit
49
Bottom CO

48

47

46

45
0 100 200 300 400 500 600 700 800
Time

Gain (K) = -0.2181, Time Constant (T1) = 53.34, Dead Time (TD) = 14.02
SSE: 0.0204

Figure 20.7 - Bottom composition response to bottom controller output pulse

and FOPDT fit of process data

The performance of the bottom PI controller in tracking set point steps in bottom composition
(not shown) while the top control loop is in manual mode reveals desirable performance. Thus, both
the top and bottom control loops perform well when implemented individually while the cross loop
controller of each is in manual mode.

217
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Both the top and bottom PI controllers are put in automatic while leaving the controller
tuning values unchanged. The same top composition set point step made in Fig 20.6 is repeated.
Recall from that figure that when the bottom controller was in manual mode, the top controller was
able to complete the set point step response in about 200 minutes. Fig. 20.8 shows that with both
controllers in automatic, the set point step response of the top composition drags on for well over
1500 minutes.
The reason for this slow response is evident in Fig. 20.8. Notice how both utop(t) and ubottom(t)
continually climb during the transient as each tries to compensate for the actions of the other. That is,
as the top controller sends more and more cold reflux down the column to raise the purity of the top
composition, the bottom controller responds by sending more and more steam up the column in an
attempt to maintain the bottom composition at set point. One is working to cool while the other
attempts to heat. This controller fight, or interaction, is a clear detriment to high-performance control.

Control Loop Interaction Degrades Set Point Tracking Performance

Process: Distillation Column Top: PID ( P= RA, I= ARW, D= off ) / Bot: PID ( P= DA, I= ARW, D= off )
Top % %

94
Distillate

93
92 set point tracking performance degrades…
70
Top CO

60
50
because control loops
1.7 fight each other
Bottoms %

1.6
1.5

70
Bot CO

60
50
40
500 1000 1500 2000 2500
Time (mins)
Tuning: Gain = 1.30, Reset Time = 62.0, Deriv Time = 0.0, Sample Time = 1.00

Tuning: Gain = -9.70, Reset Time = 53.0, Deriv Time = 0.0, Sample Time = 1.00

Figure 20.8 - Top and bottom loop fight each other, thus degrading set point tracking
performance of top loop

20.5 Distillation Study - Decoupling the Loops

As discussed previously, decouplers are feed forward elements designed to reduce this loop
interaction. Each decoupler design requires a process model and cross-loop disturbance model. For
the process models, we can use the FOPDT model fits already developed during PI controller tuning
and shown in Figs. 20.5 and 20.7. The task of developing cross-loop disturbance models remains.
No additional testing is needed to generate data for the cross-loop disturbance models
because during the previous dynamic tests, both the top and bottom loops were in manual mode.
Hence, these data sets already include information that describe how the top controller impacts the
bottom composition and how the bottom controller impacts the top composition.

218
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 The top controller pulse test shown in Fig 20.5 not only forced a response in the top
composition, but also impacted the bottom composition. We read that file into Design Tools, label the
proper columns of data and fit a FOPDT model. The result is shown in Fig. 20.9. The FOPDT
dynamic model provides a good approximation of this cross-loop disturbance data. Hence, GBT(s), the
bottom composition response to top controller output model has the parameters:
Process Gain: KD, BT = 0.24 %/%
Overall Time Constant: τD, BT = 54 minutes
Apparent Dead Time: θD, BT = 22 minutes

Similarly, the bottom controller pulse test shown in Fig. 20.7 not only forced a response in
the bottom composition, but also impacted the top composition. That data is read into Design Tools,
the proper columns of data are labeled and a FOPDT model is fit. The result is shown in Fig. 20.10.

Interact:Interact:
BottomsBottom Composition
Composition Responseto
Response to Top
Top CO
CO
Model: First Order Plus Dead Time (FOPDT) File Name: top.txt

2.0
Bottoms Composition

1.8

1.6

1.4
bottom response to top CO pulse
and FOPDT model fit of data
55

54
Top CO

53

52

51
0 100 200 300 400 500 600 700 800 900
Time

Gain (K) = 0.2365, Time Constant (T1) = 54.29, Dead Time (TD) = 22.03
SSE: 0.0435

Figure 20.9 - Bottom composition response to top controller output pulse

and FOPDT fit of process data

The FOPDT dynamic model provides a good approximation of this cross-loop disturbance data.
Hence, GBD(s), the top composition response to bottom controller output model has the parameters:

Process Gain: KD, TB = −1.0 %/%

Overall Time Constant: τD, TB = 63 minutes
Apparent Dead Time: θD, TB = 21 minutes

219
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Interact:
Interact: Top Composition
Distillate Response
Composition to Bottom
Response CO
to Bottom CO
Model: First Order Plus Dead Time (FOPDT) File Name: bottom.txt

94

Distillate Composition
Top Composition
93

92

topdistillate
top composition response
response to bottom
to bottom CO
CO pulse
pulse
and FOPDTand FOPDT
model fitmodel
of datafit of data
49
Bottom CO

48

47

46

45
0 100 200 300 400 500 600 700 800
Time

Gain (K) = -1.01, Time Constant (T1) = 63.02, Dead Time (TD) = 20.96
SSE: 0.126

Figure 20.10 - Top composition response to bottom controller output pulse

and FOPDT fit of process data

All process models and interaction models are now defined in FOPDT form. Returning to
LOOP-PRO, the PID with Decoupler controllers are chosen for both top and bottom loops. The PI
tuning parameters for the feed back loops remain as before. The FOPDT decoupler model values
required by LOOP-PRO are entered.
For the top decoupler, the FOPDT process model entered is GTT(s) and the cross-loop
disturbance model entered is GTB(s). For the bottom decoupler, the FOPDT process model entered is
GBB(s) and the cross-loop disturbance model entered is GBT(s).
When the top loop is closed, an error message results telling us that the theory requires that the
process dead time must be less than or equal to the disturbance dead time. This is a requirement of the
theory and not a limitation of LOOP-PRO (for more information, please review the design criteria
discussed in the feed forward chapter). Hence, we set θP, TT = θP, TB = 21 minutes to result in a
mathematically rational design. That dead time change is underlined in Fig. 20.11 below.
Controller performance is again tested by stepping the top set point between 92% and 94% as
shown in Fig. 20.11. The bottom loop remains closed during the evaluation. The decouplers succeed
in restoring a set point tracking performance that is nearly equal to that of the lone controller as
shown in Fig. 20.6. With two controllers and two decouplers, however, the bottom loop is now able
to maintain the bottom composition close to set point throughout the event.
While we have achieved our goal of decoupling the control loop interactions, a somewhat
disturbing result is the constant rapid movement in the controller output as shown in Fig. 20.11. This
control signal behavior will cause the mechanical valves to "chatter" in almost a nervous fashion. It is
a situation that likely will not be tolerated because valve wear and the consequent maintenance costs
will become an issue in many plants.

220
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Decouplers Restore Set Point Tracking Peformance
Process: Distillation Column Top: PID /w Decoupler / Bot: PID /w Decoupler
95

Top % %
94

Distillate
93
92 set point tracking
91
performance restored
80
Top CO

60

40

1.7
Bottoms %

1.6
1.5
1.4
1.3

80
bottoms composition remains near set point
Bot CO

60

40

0 500 1000 1500

Time (mins)
Tuning: Gain = 1.30, Reset Time = 62.0, Deriv Time = 0.0, Sample Time = 1.00
Process Model: Gain(Kp) = 1.10, T1 = 62.0, T2 = 0.0, TL = 0.0, TD = 21.0
Disturbance Model: Gain(Kd) = -1.00, T1 = 63.0, T2 = 0.0, TL = 0.0, TD = 21.0
Tuning: Gain = -9.70, Reset Time = 53.0, Deriv Time = 0.0, Sample Time = 1.00
Process Model: Gain(Kp) = -0.22, T1 = 53.0, T2 = 0.0, TL = 0.0, TD = 14.0
Disturbance Model: Gain(Kd) = 0.24, T1 = 54.0, T2 = 0.0, TL = 0.0, TD = 22.0

Figure 20.11 - Improved set point tracking capability of top loop and reduced interaction with bottom
loop when both are under PI control with decouplers

The reason for this behavior is quite subtle and understanding why it occurs is important. For
the models used in the bottom decoupler, the process model gain, KP, BB, is −0.22 %/% and the cross
loop disturbance gain, KD, BT, is 0.24 %/%. In absolute value, |KD, BT| > |KP, BB|, which says that the top
controller (the disturbance) has a larger influence on the bottom composition than does the bottom
controller.
The relative size of the gains displayed by the process and used in the decouplers is
fundamental to stable control. We will learn more in the next chapter, but it is reasonable to postulate
that a decoupler must have at least as much influence on its own process variable as does any
disturbance.
Following this assumption, we lower the cross-loop disturbance gain of the bottom loop
decoupler so that in absolute value |KD, BT| = |KP, BB| as underlined in Figure 20.12. Repeating the top
set point step test as before, we observe in Fig. 13.12 that decoupling performance is maintained yet
the controller output chatter is eliminated.

221
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Relative Size of Decoupler Gains Impacts Performance
Process: Distillation Column Top: PID /w Decoupler / Bot: PID /w Decoupler
95

Top %%
94

Distillate
93
92
91 same performance but…
80
Top CO

60

40

1.7 controllers no longer “chatter”

Bottoms %

1.6
1.5
1.4
1.3

80
Bot CO

60

40

0 500 1000 1500

Time (mins)
Tuning: Gain = 1.30, Reset Time = 62.0, Deriv Time = 0.0, Sample Time = 1.00
Process Model: Gain(Kp) = 1.10, T1 = 62.0, T2 = 0.0, TL = 0.0, TD = 21.0
Disturbance Model: Gain(Kd) = -1.00, T1 = 63.0, T2 = 0.0, TL = 0.0, TD = 21.0
Tuning: Gain = -9.70, Reset Time = 53.0, Deriv Time = 0.0, Sample Time = 1.00
Process Model: Gain(Kp) = -0.22, T1 = 53.0, T2 = 0.0, TL = 0.0, TD = 14.0
Disturbance Model: Gain(Kd) = 0.22, T1 = 54.0, T2 = 0.0, TL = 0.0, TD = 22.0

Figure 20.12 - Decoupled loops do not chatter with slight adjustment to one model parameter

222
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 21. Modeling, Analysis and Control of Multivariable Processes

21.1 Generalizing 2x2 Multivariable Processes

As we learned when studying the distillation column in the previous chapter, multivariable loop
interaction presents a substantial control challenge. To explore in more detail the nature and
challenges of multivariable control, we generalize all two-by-two multi-input multi-output (2x2 MIMO)
dynamic processes into a standard form.
Following the nomenclature established in the popular texts, Gij represents the dynamic
behavior of the ith measured process variable response to the jth controller output signal. Hence, as
shown in Fig. 21.1, process G11 describes the direct dynamic response of measured process variable
PV1 to changes in controller output CO1, and interaction G21 describes the cross-loop dynamic
response of PV2 to changes in CO1.
G11
CO1 PV1

G12 G21
CO2 PV2
G22

Figure 21.1 - General multivariable architecture

The multi-loop Custom Process in LOOP-PRO enables the simulation of a broad range of
2x2 MIMO processes that follow this general form. Figure 21.2 shows the multi-loop Custom Process
graphic for building such simulations.

G11

G21

G12

G22

Figure 21.2 - LOOP-PRO's Multi-Loop Custom Process

223
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
21.2 Relative Gain as a Measure of Loop Interaction
Before exploring different multivariable process behaviors, we introduce the concept of relative gain.
Relative gain, λ, is popular because it:
- provides a convenient measure of loop interaction,
- is easy to compute, and
- is dimensionless so it is not affected by the units of the process data.

Relative gain is computed from the steady state gains of the process and cross-loop interaction models
that best describe observed process behavior (that result from model fits of process data). Following
the nomenclature we have established, relative gain is computed:

K11 K 22
λ= (21.1)
K11 K 22 − K12 K 21

In the following sections we explore what the size and sign of λ implies for multivariable loop
interaction and the ease with which a process can be controlled.
Before starting that study, consider that a 2x2 MIMO process has two controllers (COA and
COB) that adjust two final control elements (FA and FB) to regulate two process variables (PVA and
PVB). The controllers are connected to the process variables and final control elements by wires.
Consequently, the control loops can be wired two ways:

1) COA manipulates FA to control PVA ; COB manipulates FB to control PVB

2) COA manipulates FA to control PVB ; COB manipulates FB to control PVA

Each combination yields a different value of λ for a multivariable process. What we will learn is that:

control loops should always be paired (wired) where possible

so the relative gain is positive and as close as possible to one.

21.3 Effect of KP on Control Loop Interaction

To appreciate the usefulness of relative gain as a measure of cross-loop interaction, consider the
variety of 2x2 MIMO cases listed in Table 21.3. These cases are simulated and studied here using
LOOP-PRO's Custom Process feature shown in Fig 21.2.
All of the direct process and interaction models used in the Custom Process simulations are
first order plus dead time (FOPDT), and all time constant and dead time parameters for all the
simulations in Table 21.3 are:
Process time constant, τP = 10
Dead time, θP = 1

Also, all of the investigations use two PI controllers with no decoupling and with:
Controller gain, KC = 5
Reset time, τI = 10

For all examples, when one PI controller is put in automatic while the other is in manual mode, that
controller tracks set point changes quickly and with little oscillation. The issue we study here is
process behavior when both PI controllers are put in automatic at the same time.

224
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
For each simulation case study, the direct process and cross-loop interaction gains are:

direct cross-loop cross-loop direct

CO1 Î PV1 CO1 Î PV2 CO2 Î PV1 CO2 Î PV2

Case K11 K21 K12 K22 λ

1 1 1.1 1.1 1 - 4.8

2a 0 1.1 0.5 1 0.0
2b -1 3.0 1.1 1 0.2
2c 1 -3.0 0.5 1 0.4
3 1 -1.1 0.5 1 0.6
4 1 0 0.5 1 1.0
5 1 1.1 0.5 1 2.2
6 1 1.1 .85 1 15.4

Table 21.3 - Exploring relative gain, λ, as a measure of loop interaction

Case 1: λ < 0
When the cross-loop interaction gains are larger then the direct process gain, as is true for Case 1 in
Table 21.3, then each controller has more influence on its cross-loop measured process variable than
it does on its own direct measured process variable. As listed in the table, the relative gain, λ,
computed by Eq. 21.1 for this case is negative.
Figure 21.4 shows the set point tracking performance of the Case 1 process when both loops
are under PI control with no decoupling (remember that for all simulations, τP = 10 and θP = 1, with
KC = 5 and τI = 10). As each controller works to keep its direct measured process variable on set
point, every control action causes an even larger disruption in the cross-loop process variable. And
the harder each controller works, the worse the situation. As shown in Fig. 21.4, the result is an
unstable, diverging system.
The Process is Unstable When Relative Gain, λ < 0
( )

50
PV 1

48

100
CO 1

75
50
54
PV 2

52
50

50
CO 2

25

25 30 35 40 45 50 55 60 65
Time (time units)
Tuning: Gain = 5.00, Reset Time = 10.0, Deriv Time = 0.0, Sample Time = 0.10

Tuning: Gain = 5.00, Reset Time = 10.0, Deriv Time = 0.0, Sample Time = 0.10

Figure 21.4 - λ < 0 indicates incorrect loop pairing and an unstable process under PI control

225
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
A negative relative gain implies that the loop pairing is incorrect. That is, each controller is wired to
the wrong measured process variable. The best course of action is to switch the controller wiring.
This switches the cross-loop gains in Table 4.3 to direct process gains and vice versa.
Switching the loop pairing recasts Case 1 into a process with a relative gain of λ = 5.8. This
loop interaction behavior is somewhere between Case 5 and Case 6. As we will learn, a process with
this relative gain is challenging to control, but it is closed-loop stable and the loops can be decoupled
following standard methodologies.

Case 1 Can Be Stabilized By Detuning and Reversing Control Action

51
PV 1

50
49
55
CO 1

50
45
40
51
PV 2

50
49
60
CO 2

55
50
45
20 40 60 80 100 120 140
Time (time units)
Tuning: Gain = -3.00, Reset Time = 10.0, Deriv Time = 0.0, Sample Time = 0.10

Tuning: Gain = 5.00, Reset Time = 10.0, Deriv Time = 0.0, Sample Time = 0.10

Figure 21.5 - Case 1 can be stabilized by detuning one of the loops and reversing its action

As an alternative to switching the pairing, one of the controllers can be detuned (the KC can
be lowered) and the control action on that controller reversed (e.g. if it is direct acting make it reverse
acting). This tricks the control system into making reasonable corrective actions. Figure 21.5 shows
that the Case 1 process can be stabilized by detuning and reversing the action of loop 1 as indicated
by the KC circled in the figure. The resulting set point tracking performance is rather unsatisfactory,
however, and we do not recommend this approach as a general solution.

Case 2: 0 < λ ≤ 0.5

For the relative gain to be exactly zero (λ = 0), one of the direct process gains must be zero. A direct
process gain of zero means that a controller has no impact on the measured process variable it is
wired to. Clearly there can be no regulation if a controller has no influence.
Case 2a in Table 21.4 has K11 = 0, implying that CO1 has no influence on PV1. Yet because
the cross-loop gain K12 is not zero, changes in CO2 will disrupt PV1. If a measured process variable
can be disrupted but there is no means to control it, the result is an unstable process under PI control
(no figure shown). Because both cross-loop gains are not zero in Case 2a, the loop pairing should be
switched in this case to give each controller direct influence over a measured process variable. This
would recast Case 2a into a process with a λ = 1.0, which is the interaction measure most desired. We
study such a process in Case 4 below.

226
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 When the relative gain is near zero (0 < λ ≤ 0.5), then at least one of the cross-loop gains is
large on an absolute basis (e.g. Case 2b and 2c). Under PI control with no decoupling and using the
base tuning values of KC = 5 and τI = 10, both of these processes are unstable and show considerable
loop interaction (no figure shown). Detuning both controllers to KC = 2 and τI = 10 restores stability
but control-loop interaction is still significant.
Again, the best course of action is to switch loop pairing. With the wiring switched, Case 2b
yields λ = 0.8 and Case 2c yields λ = 0.6, putting both relative gains closer to the desired value of
one. While both processes still display loop interaction, the processes become stable under PI control
with no decoupling, even with fairly aggressive PI controller tuning.

Case 3: 0.5 ≤ λ < 1

When the relative gain is between 0.5 and one, the cross-loop interactions cause each control action to
be reflected and amplified in both process variables. As shown in the left most set point steps of Fig.
21.6 for a case where λ = 0.6, this interaction leads to a measured process variable response that
includes significant overshoot and slowly damping oscillations.
This amplifying interaction exists when stepping the set point of either loop. It grows more
extreme and ultimately leads to an unstable process as λ approaches zero (see Case 2). The interaction
becomes less pronounced as λ approaches one (see Case 4).

Relative Gain and Multi-Loop Interaction

51
PV 1

50
λ= 0.6 λ= 1 λ= 2.2 λ = 15.4 (case 6)
60 (case 3) (case 4) (case 5)
CO 1

50
40
51
PV 2

50

60
CO 2

50
40

50 100 150 200 250 300 350

Time (time units)
Tuning: Gain = 5.00, Reset Time = 10.0, Deriv Time = 0.0, Sample Time = 0.10

Tuning: Gain = 5.00, Reset Time = 10.0, Deriv Time = 0.0, Sample Time = 0.10

Figure 21.6 - Impact of λ on PI control loop interaction with no decoupling

Case 4: λ = 1
A relative gain of one occurs when either or both of the cross-loop gains are zero. In Case 4, K21 is
zero so controller output CO1 has no impact on the cross-loop measured process variable PV2.
However, since K12 is not zero as listed in Table 21.4, changes in CO2 will impact PV1.

227
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 The second set point steps in Fig. 21.6 show the control performance of the Case 4 process
when the set point of PV1 is changed. As expected, the set point tracking actions of CO1 have no
impact on PV2. While not shown, a set point step in PV2 would cause a some cross-loop disruption in
PV1 because of loop interaction.
When both cross-loop gains are zero, the loops do not interact. Such a system is naturally and
completely decoupled and the controllers should be designed and tuned as single loop processes.

Case 5: λ > 1
Opposite to the observations of Case 3, when the relative gain is greater than one, the control loops
fight each other. Specifically, the cross-loop interactions act to restrain movement in the measured
process variables, prolonging the set point response. The third set point steps in Fig. 21.6 illustrate
this behavior for a case where λ = 2.2. We also observed this prolonged interaction in the distillation
case study in the previous chapter.
As stated earlier, a process with a relative gain that is positive and close to one displays the
smallest loop interactions (is better behaved). For Case 5, switching the loop pairing would yield a
very undesirable negative λ. This means that the loops are correctly paired and the significant loop
interaction is unavoidable.

Case 6: λ >> 1
As the cross-loop gain product, K12K21, approaches the direct process gain product, K11K22, the
relative gain grows and the restraining effect on movement in the measured process variables
discussed in Case 5 becomes greater. This is illustrated in the right most set point steps in Fig. 21.6
for a case where λ = 15.4. Again, switching the loop pairing would yield a negative λ so the loops are
correctly paired and the significant loop interaction is unavoidable. Interestingly, as the cross-loop
gains grow to the point that their product is larger then the direct process gain product (when K12K21 >
K11K22), then λ becomes negative and we circle back to Case 1.

21.4 Effect of τP on Control Loop Interaction

The relative gain provides a convenient means for predicting how the size and sign of the direct
process and cross-loop gains will impact loop interaction on a 2x2 process. Here we learn that while
τP is fundamental to loop interaction dynamics, it rarely plays as important a role as the gains. There
is no simple equation analogous to relative gain we can use for this study, so we learn by example.
To explore the role of cross-loop time constants on loop interaction, we define a set of 2x2
processes where for all cases:
Steady State Gains: K11 = 1.0, K21 = 0.5, K12 = 0.5, K22 = 1.0
Dead Times: θ11 = 1.0, θ21 = 1.0, θ12 = 1.0, θ22 = 1.0
All of the investigations use two PI controllers with no decoupling and with:
Controller gain, KC = 5
Reset time, τI = 10

Based on Eq. 21.1, all three cases have a relative gain λ = 1.3, which describes a process with some
loop interactions, but as explained in Case 5, is generally well-behaved.

228
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 direct cross-loop cross-loop direct
CO1 Î PV1 CO1 Î PV2 CO2 Î PV1 CO2 Î PV2

Case τ11 τ21 τ12 τ22

base 10 10 10 10
7 10 30 10 10
8 10 3 10 10

Table 21.7 - Exploring the impact of cross-loop time constants on loop interaction

With the gains and dead times as listed above, we investigate a base case and two variations
that focus on time constants. In particular, the direct process and cross-loop interaction time constants
explored here are listed in Table 21.7.

Base Case
The base case in Table 21.7 establishes the dynamic behavior when the direct process and cross-loop
time constants have equal influence. As shown in the left most set point steps of Fig. 21.8, the loop
interactions for the base case are modest and display some restraint in the movement of the measured
process variables. This prolongs the set point response as discussed in Case 5 and shown to the right
in Fig. 21.6 above.

Impact of Time Constant on Multi-Loop Interaction

51.0
PV 1

50.5
50.0 τ21= 10 τ21 = 30 τ21 = 3
60 (base case) (case 7) (case 8)
CO 1

55
50
45
51
PV 2

50
49
55
CO 2

50
45

40 60 80 100 120 140 160 180

Time (time units)
Tuning: Gain = 5.00, Reset Time = 10.0, Deriv Time = 0.0, Sample Time = 0.10

Tuning: Gain = 5.00, Reset Time = 10.0, Deriv Time = 0.0, Sample Time = 0.10

Figure 21.8 - Impact of τP on PI control loop interaction with no decoupling

229
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
Case 7: Cross-Loop τP Large
This process, with τ21 = 30 while τ11 = τ22 = τ12 = 10, considers a case where one of the cross-loop
time constants is large relative to the direct process time constants. A short time constant implies a
fast reaction time. Hence, CO1 and CO2 can correct for errors in their direct process variables quickly,
while the disruptive influence of CO1 on cross-loop process variable PV2 proceeds slowly in
comparison.
Because each controller can quickly impact its own direct process variable, each is capable of
correcting for the disruptive effects of the slow moving cross-loop interaction as it arrives. And
consequently, as shown in the second set point steps of Fig. 21.8, performance is improved relative to
the base case when using two PI controllers with no decoupling and with KC = 5 and τI = 10.
Specifically, the set point tracking performance of PV1 in case 7 shows less oscillation as it moves to
the new set point, and the size and speed of disruption of CO1 on PV2 is lessened in comparison to the
base case.

Case 8: Cross-Loop τP Small

Opposite to Case 7, with τ21 = 5 while τ11 = τ22 = τ12 = 10, Case 8 in Table 21.7 considers a process
where one of the cross-loop time constants is small relative to the direct process time constants. Thus,
CO1 is able to disrupt cross-loop process variable PV2 faster than CO2 is able to correct the problem.
Because the cross-loop disruption is fast relative to the controller's ability to regulate its own
direct process variable, the third set point steps of Fig. 21.8 show that control performance degrades
relative to the base case. As shown in the figure, the set point tracking performance of PV1 in case 8
shows greater oscillations and a larger overshoot as it moves to the new set point. Also shown is that
the disruption of CO1 on PV2 is significantly greater here than the base case.

21.5 Effect of θP on Control Loop Interaction

Similar to the time constant, dead time plays a secondary role in loop interaction dynamics compared
to the relative gains. We continue to learn by investigation with a set of 2x2 processes where for all
cases:

Steady State Gains: K11 = 1.0, K21 = 0.5, K12 = 0.5, K22 = 1.0
Time Constants: τ11 = 10, τ21 = 10, τ12 = 10, τ22 = 10

controlled by two PI controllers with no decoupling and with:

Controller gain, KC = 5
Reset time, τI = 10

Based on Eq. 21.1, all three cases have a relative gain λ = 1.3, which describes a process with some
loop interactions, but as explained in Case 5, is generally well-behaved. We again explore a base case
and two variations where:

Base Case
The base case in Table 21.9 establishes the dynamic behavior when the direct process and cross-loop
dead times are equal. As shown in the left most set point steps of Fig. 21.10, the loop interactions for
the base case are identical to those shown in the base case of Fig. 21.8 above.

230
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 direct cross-loop cross-loop direct
CO1 Î PV1 CO1 Î PV2 CO2 Î PV1 CO2 Î PV2

Case θ11 θ21 θ12 θ22

base 1 1 1 1
9 1 3 1 1
10 1 0.3 1 1

Table 21.9 - Exploring the impact of cross-loop dead time on loop interaction

Case 9: Cross-Loop θP Large

This case, with θ21 = 3 while θ11 = θ22 = θ12 = 1, considers a process where one of the cross-loop dead
times is large relative to the direct process dead times. Hence, there is a short delay before a change in
CO1 impacts PV1 and a long delay before it begins disrupting cross-loop variable PV2.

Impact of Dead Time on Multi-Loop Interaction

51.0
PV 1

50.5
50.0
θ21= 1 θ21 = 3 θ21 = 0.3
60 (base case) (case 9) (case 10)
CO 1

55
50
45
50.5
PV 2

50.0
49.5
52
CO 2

50
48

20 40 60 80100 120 140 160 180 200

Time (time units)
Tuning: Gain = 5.00, Reset Time = 10.0, Deriv Time = 0.0, Sample Time = 0.10

Tuning: Gain = 5.00, Reset Time = 10.0, Deriv Time = 0.0, Sample Time = 0.10

Figure 21.10 - Impact of θP on PI control loop interaction with no decoupling

When the second set point steps occur in the middle of Fig. 21.10, CO1 begins moving PV1 to
track the change. As a consequence, CO1 also imparts a disruptive event to PV2. Because of the cross-
loop dead time, there is a relatively long delay before the disruption finally impacts PV2.
When the disruption to PV2 does begin, CO2 takes action to reject the disturbance. And as
CO2 changes, the cross-loop interaction causes the control actions to be reflected back to PV1. The
extended delay in one of the steps of this back and forth interaction produces an "echo effect" in the
set point response of PV1. This can be seen in Fig. 21.10 as a more pronounced second dip in the
response compared to that evident in the base case.

231
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
Case 10: Cross-Loop θP Small
Opposite to case 9, with θ21 = 0.3 while θ11 = θ22 = θ12 = 1, we consider a case where one of the cross-
loop dead times is small relative to the direct process dead times. Hence, there is a relatively short
delay before a change in CO1 imparts a cross-loop disruption to PV2.
When the disruption to PV2 begins, CO2 takes action to reject the disturbance and these
actions are reflected back to PV1. As shown in the right most right most set point steps of Fig. 21.10,
the shorter delay has little impact on loop interaction. The plot reveals that controller performance is
very similar to that of the base case.

21.6 Decoupling Cross-Loop KP Effects

Recall that a decoupler is a feed forward element where the measured disturbance is the actions of a
cross-loop controller. Analogous to a feed forward controller, a decoupler is comprised of a process
model and a cross-loop disturbance model. The cross-loop disturbance model receives the cross-loop
controller signal and predicts an “impact profile,” or when and by how much the process variable will
be impacted. Given this predicted sequence of disruption, the process model then back calculates a
series of control actions that will exactly counteract the cross-loop disturbance as it arrives so the
measured process variable, in theory, remains constant at set point.
Here we explore how perfect decouplers can reduce cross-loop interaction. A perfect
decoupler employs the identical models in the decoupler as is used for the process simulation. Be
aware that in real-world applications, no decoupler model exactly represents the true process
behavior. Hence, the decoupling capabilities shown here must be considered as the best possible
performance. We revisit the "no decoupler" cases explored earlier in this chapter as summarized in
Table 21.3.

Case 1: λ < 0
A negative relative gain implies that the loop pairing is incorrect. Decoupling is not explored for Case
1 in Table 21.3 because the best course of action is to switch the controller wiring to produce a 2x2
process with a relative gain of λ = 5.8. This loop interaction behavior is somewhere between Case 5
and Case 6 discussed below.

Case 2: 0 < λ ≤ 0.5

A relative gain of exactly zero (λ = 0) implies that at least one controller has no impact on the
measured process variable it is wired to. There can be no regulation if a controller has no influence.
Hence, decoupling becomes meaningless for this case.
When the relative gain is near zero (0 < λ ≤ 0.5), PI controllers with no decoupling must be
detuned to stabilize the 2x2 system. Even when the PI controllers are detuned, perfect decouplers (the
identical models are used in the decouplers as are used for the process simulation) result in an
unstable system (no figure shown). Detuning the decouplers (lowering the disturbance model gain)
will restore stability but interaction remains significant and general performance is poor. Again, the
best course of action is to switch loop pairing.

Case 3: 0.5 ≤ λ < 1

When the relative gain is between 0.5 and one, the cross-loop interactions cause each control action to
be reflected and amplified in both process variables. As shown in the left most set point steps of Fig.
21.11 for the case of λ = 0.6, PI controllers with perfect decouplers virtually eliminate cross loop
interactions. This is not surprising since the relative gain is positive and close to one.

232
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
Case 4: λ = 1
A relative gain of one occurs when either or both of the cross-loop gains are zero. In Case 4 of Table
21.3, K21 is zero so controller output CO1 has no impact on the cross-loop measured process variable
PV2. Consequently, a perfect decoupler will provide no benefit for this loop. And as shown in Fig.
21.11 for the middle set point steps, while a perfect decoupler causes no harm, a decoupler
implemented on a real process will likely have imperfect models and would then actually create loop
interaction.
Table 21.3 shows that K12 is not zero, however, so changes in CO2 will impact PV1. A perfect
decoupler will virtually eliminate cross loop interaction for information flow in this direction (no
figure shown). Thus, the Case 4 MIMO process can address the 2x2 loop interaction with a single
decoupler on the CO2 to PV1 loop.

Case 5: λ > 1
When the relative gain is greater than one, the cross-loop interactions act to restrain movement in the
measured process variables. The third set point steps in Fig. 21.11 for the case where λ = 2.2 illustrate
that perfect decouplers substantially eliminate both this restraining effect and the level of loop
interaction. Again, this is not surprising since the relative gain is positive and reasonably close to one.
Perfect Decouplers Minimize Interaction for Moderate λ
Process: Multi-Loop Custom Process Cont. 1: PID w/ Decoupler
Cont. 2: PID w/ Decoupler
51.0
PV 1

50.5
50.0
λ= 0.6 λ= 1 λ= 2.2
60 (case 3) (case 4) (case 5)
CO 1

50
40
50.5
PV 2

50.0
49.5
60
CO 2

50
40

20 40 60
80 100 120 140 160 180 200
Time (time units)
Tuning: Gain = 5.00, Reset Time = 10.0, Sample Time = 0.10
Process Model: Gain(Kp) = 1.00, T1 = 10.0, T2 = 0.0, TL = 0.0, TD = 1.00
Disturbance Model: Gain(Kd) = 0.50, T1 = 10.0, T2 = 0.0, TL = 0.0, TD = 1.00
Tuning: Gain = 5.00, Reset Time = 10.0, Sample Time = 0.10
Process Model: Gain(Kp) = 1.00, T1 = 10.0, T2 = 0.0, TL = 0.0, TD = 1.00
Disturbance Model: Gain(Kd) = …..
1.10, T1 = 10.0, T2 = 0.0, TL = 0.0, TD = 1.00

changes with case

Figure 21.11 - Perfect decouplers can virtually eliminate cross-loop interaction when λ is near 1.

233
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
Case 6: λ >> 1
As relative gain grows larger, the restraining effect on movement in the measured process variables
due to loop interaction becomes greater. Case 6 in Table 21.3 is interesting because K21 is greater than
both K11 and K22. That is, controller output CO1 has a greater influence on cross-loop process variable
PV2 then it does on its own direct process variable PV1. Further, the influence of this disturbance on
PV2 is large compared to the actions CO2 must take to reject it. Switching loop pairing offers no
benefit as this makes the relative gain negative.
With perfect decouplers as shown in the right set point steps of Fig. 21.12 (the decoupler
employs the identical models as are used for the process simulation), the system is unstable. This
cannot be addressed by detuning the PI controller. Even with lower values for controller gain, KC, the
system is unstable. As we postulate at the end of the distillation case study in the previous chapter, a
decoupler must have as much influence on its own process variable as does any disturbance.
Hence, we detune the decoupler by lowering the cross-loop disturbance gain of the bottom
loop so that in absolute value, K21 ≤ K22 and K21 ≤ K11. Repeating the set point steps in the left set
point steps of Fig. 21.12 reveal a stable and reasonably decoupled system.
Detuning Station:
Control the Decouplers for Case
Case 6 (λ >> 1)
Studies
Process: Multi-Loop Custom Process Cont. 1: PID w/ Decoupler
Cont. 2: PID w/ Decoupler
51
PV 1

50
detuned cross-loop decoupler perfect cross-loop decoupler
100 (K21 = 1.0) (K21 = 1.1)
CO 1

50
0
51
PV 2

50

100
CO 2

50
0
40 60 80 100
120 140 160 180 200 220
Time (time units)
Tuning: Gain = 5.00, Reset Time = 10.0, Sample Time = 0.10
Process Model: Gain(Kp) = 1.00, T1 = 10.0, T2 = 0.0, TL = 0.0, TD = 1.00
Disturbance Model: Gain(Kd) = 0.85, T1 = 10.0, T2 = 0.0, TL = 0.0, TD = 1.00
Tuning: Gain = 5.00, Reset Time = 10.0, Sample Time = 0.10
Process Model: Gain(Kp) = 1.00, T1 = 10.0, T2 = 0.0, TL = 0.0, TD = 1.00
Disturbance Model: Gain(Kd) = 1.10, T1 = 10.0, T2 = 0.0, TL = 0.0, TD = 1.00

lower cross-loop gain

Figure 21.12 - Perfect decouplers must be detuned to produce a stable system for large λ

234
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
21.7 Decoupling Cross-Loop τP Effects
We continue exploring how perfect decouplers can reduce cross-loop interaction and now focus on
perfect decoupling of time constant effects. We revisit the "no decoupler" cases explored earlier in
this chapter as summarized in Table 21.7. Recall that all three cases have a relative gain λ = 1.3,
which describes a process with some loop interactions but is generally well-behaved.

Base Case
The base case in Table 21.7 is a system where the direct process and cross-loop time constants have
equal influence. As shown in the left most set point steps of Fig. 21.13, perfect decouplers virtually
eliminate cross-loop interaction for this process.

Case 7: Cross-Loop τP Large

Case 7 in Table 21.7, with τ21 = 30 while τ11 = τ22 = τ12 = 10, considers a case where one of the cross-
loop time constants is large relative to the direct process time constants. Here, the disruptive influence
of CO1 on cross-loop process variable PV2 proceeds relatively slowly. This enables perfect
decouplers, as shown in the middle set point steps of Fig. 21.13, to effectively eliminate cross-loop
interaction for this process.

Perfect Decouplers
Control Reduce Time
Station: CaseConstant Interaction
Studies
Process: Multi-Loop Custom Process Cont. 1: PID w/ Decoupler
Cont. 2: PID w/ Decoupler

51.0
PV 1

50.5
50.0
τP = 10 τP = 30 τP = 3
75 (base case) (case 7) (case 8)
CO 1

50
25
51
PV 2

50
49

75
CO 2

50
25

40 50 60 70 80 90 100 110 120 130 140 150 160

Time (time units)
Tuning: Gain = 5.00, Reset Time = 10.0, Sample Time = 0.05
Process Model: Gain(Kp) = 1.00, T1 = 10.0, T2 = 0.0, TL = 0.0, TD = 1.00
Disturb Model: Gain(Kd) = 0.50, T1 = 10.0, T2 = 0.0, TL = 0.0, TD = 1.00
Tuning: Gain = 5.00, Reset Time = 10.0, Sample Time = 0.10
Process Model: Gain(Kp) = 1.00, T1 = 10.0, T2 = 0.0, TL = 0.0, TD = 1.00
Disturb Model: Gain(Kd) = 0.50, T1 =…..
3.00, T2 = 0.0, TL = 0.0, TD = 1.00

changes with case

Figure 21.13 - Decoupling loop interaction due to time constant effects

235
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
Case 8: Cross-Loop τP Small
With τ21 = 5 while τ11 = τ22 = τ12 = 10, Case 8 in Table 21.7 considers a process where CO1 is able to
disrupt cross-loop process variable PV2 faster than CO2 is able to correct the problem. In spite of this
challenge, perfect decouplers are able to dramatically reduce the loop interaction as shown in the right
most set point steps of Fig. 21.13. We learn from this study that small cross-loop time constants
present the greatest decoupling challenge.

21.8 Decoupling Cross-Loop θP Effects

We finish by exploring how perfect decouplers can reduce cross-loop interaction and decouple dead
time effects. We revisit the "no decoupler" cases explored earlier in this chapter as summarized in
Table 21.9. Recall that all three cases have a relative gain λ = 1.3, which describes a process with
some loop interactions but is generally well-behaved.

Base Case
The base case in Table 21.9 is identical to the base case of Table 21.7. That is, the dead times all have
equal influence. As shown in the left most set point steps of Fig. 21.14, perfect decouplers virtually
eliminate cross-loop interaction for this process.
Perfect Decouplers
Control Reduce Dead
Station: TimeStudies
Case Interaction
Process: Multi-Loop Custom Process Cont. 1: PID w/ Decoupler
Cont. 2: PID w/ Decoupler
51.0
PV 1

50.5
50.0
θP = 10 θP = 30 θP = 3
60 (case 9) (case 10)
(base case)
55
CO 1

50
45
50.5
PV 2

50.0
49.5

55
CO 2

50
45

40 60
80 100 120 140 160
Time (time units)
Tuning: Gain = 5.00, Reset Time = 10.0, Sample Time = 0.10
Process Model: Gain(Kp) = 1.00, T1 = 10.0, T2 = 0.0, TL = 0.0, TD = 1.00
Disturb Model: Gain(Kd) = 0.50, T1 = 10.0, T2 = 0.0, TL = 0.0, TD = 1.00
Tuning: Gain = 5.00, Reset Time = 10.0, Sample Time = 0.10
Process Model: Gain(Kp) = 1.00, T1 = 10.0, T2 = 0.0, TL = 0.0, TD = 1.00
Disturb Model: Gain(Kd) = 0.50, T1 = 10.0, T2 = 0.0, TL = 0.0, TD = 1.00
…..
changes with case

Figure 21.14 - Decoupling loop interaction due to dead time effects

236
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
Case 9: Cross-Loop θP Large
Case 9 in Table 21.9, with θ21 = 3 while θ11 = θ22 = θ12 = 1, considers a process where there is a short
delay before a change in CO1 impacts PV1 and a long delay before it begins disrupting cross-loop
variable PV2. This additional time enables perfect decouplers, as shown in the middle set point steps
of Fig. 21.14, to eliminate cross-loop interaction for this process.

Case 10: Cross-Loop θP Small

With θ21 = 0.3 while θ11 = θ22 = θ12 = 1, Case 10 considers a process where one of the cross-loop dead
times is small relative to the direct process dead times. We cannot implement perfect decouplers here
because the theory requires that we enter a process dead time in the decoupler model that is less than
or equal to the disturbance dead time. This is not a limitation of LOOP-PRO. To do otherwise would
be seeking to compute a corrective control action before the disturbance change first occurs.
Suppose for a particular application, you determine that the disturbance dead time is indeed
shorter than the process dead time. Best practice is to arbitrarily set the process dead time equal to the
disturbance dead time when entering values into the decoupler input form. An approach that yields
identical performance is to set the disturbance dead time equal to the process dead time (the approach
we take here). Decoupler performance will suffer but at least the calculations will yield control
actions that make physical sense.
Thus, even though the process simulation has the cross loop dead time of θ21 = 0.3, we enter
θ21 = θ22 = 1.0 in the decoupler input form. The right most set point steps of Fig. 21.14 show that
while loop interaction is reduced, the comprise we were forced to make with the decoupler model
parameters prevents complete decoupling. We learn from this study that small cross-loop dead times
present the greatest decoupling challenge.

237
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 22. Model Based Smith Predictor For Processes with Large Dead Time
22.1 A Large Dead Time Impacts Controller Performance
Dead time is considered large only in comparison to the time constant of a process. As dead time
grows large relative to the time constant (θP ≥ τP), it becomes increasingly difficult to achieve a tight
set point tracking performance with traditional PID control. One example of this performance loss is
that as dead time increases, the rise time and settling time must lengthen to maintain a desired peak
overshoot ratio.
Suppose a process has a dead time equal to the time constant (θP = τP) and the controller
sample time is ten times per time constant (T = 0.1τP). For such a process, 10 full samples (one dead
time) must pass after a control action before the sensor detects any impact. And every subsequent
control action encounters this tremendous delay. The controller tuning must be sluggish enough to
reflect this dead time. Otherwise, too much corrective action can accumulate in the dead time
“pipeline,” leading to large oscillations and even a dangerous unstable operation.
To visualize this idea, consider the tank temperature control process shown in Fig. 22.1. A
hot and cold liquid stream combine at the entrance to a pipe, travel its length, and spill into a tank.
The control objective is to maintain temperature in the tank by adjusting the flow rate of hot liquid
entering the pipe (this case study is not available in LOOP-PRO but is illustrative for this discussion).

2. The sensor does not see the result of the control action
cold liquid until the hot liquid travels down the pipe, and this
Fsetpoint dead time makes tight control difficult.
FC

hot liquid

Tsetpoint
TC

1. If the measured temperature is below set point,

the controller calls for more hot liquid.

Figure 22.1 – The large dead time from the pipe makes tight temperature control difficult

If tank temperature is below set point (too cold), the controller opens the hot liquid valve to
compensate and the temperature of the liquid entering the pipe increases. The temperature sensor in
the tank does not immediately see the result of this action, however, because the additional hot liquid
is still a pipe length away.
If the controller is tuned for a rapid response (a large KC and small τI), it will open the hot
liquid valve more and more in an attempt to raise tank temperature. The pipe fills with ever hotter
liquid until the first hot liquid makes its way down the pipe and finally spills into the tank.

238
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 When tank temperature eventually rises to set point, the controller steadies the hot liquid flow
rate entering the pipe to stop any additional temperature rise. But the pipe is now filled with hot liquid
and it continues spilling into the tank. As tank temperature rises yet further, the controller begins
closing the valve to cool the tank. Again, because of the delayed response, the controller will fill the
pipe with too-cold liquid before the actions are finally revealed in the tank.
The dead time from the pipe combined with an overly aggressive controller causes the tank
temperature to cycle between too-hot and too-cold. One solution is to detune the controller. That is,
decrease the controller gain and/or increase the reset time to make the action more sluggish.
A sluggish controller makes small control actions at a slow pace. This enables the controller
to literally wait out the delay between action and response. Detuning addresses the problem of over-
correction just described, but sluggish tuning is associated with a poor control performance.

22.2 Predictive Models as Part of the Controller Architecture

An alternative solution is model predictive control (MPC), which has the potential to markedly
improve closed loop performance in the presence of large dead time. Model predictive controllers
incorporate a dynamic process model as part of the architecture of the control algorithm. For success
in implementation, the model must reasonably describe the controller output to measured process
variable dynamic behavior. We already have experience developing dynamic process models because
the FOPDT (first order plus dead time) form used for controller tuning is suitable for MPC
implementation, though it is the simplest of the models available.
The function of the dynamic model is to predict the future value of the measured process
variable based on the current state of the process and knowledge of recent control actions. A control
action is “recent” if the dynamic response it caused in the process is still in progress.
If the predicted measured process variable does not match a desired set point, this future error
enables corrective control actions to be taken immediately, and before the predicted problem actually
occurs. Thus, the model predictive controller exploits process knowledge contained in the dynamic
model to compute current control actions based on a predicted future.
In theory, a perfect model can eliminate the negative influence of dead time on controller
performance. In the practical world, as we will see later in this chapter, MPC can certainly provide a
performance benefit in the presence of large dead time. This benefit comes at a price, however. The
designer must identify an appropriate dynamic model form, fit the model parameters to process data,
and program the result into the control computer. The model predictions must then be properly
sequenced with the computations of the feedback loop to create an integrated MPC architecture.

22.3 The Smith Predictor Control Algorithm

The Smith predictor is the simplest implementation of general MPC theory (a variant of the Smith
predictor can be created from MPC theory by choosing both the near and far prediction horizon as
θP/T + 1 samples, a control horizon of one, and a constant set point profile). The Smith predictor
architecture, illustrated in Fig. 22.2, is comprised of an ideal (or no dead time) process model block
and a separate dead time model block. The computation proceeds as follows:
1) The ideal process model receives the current value of the controller output, u(t), and computes
yideal(t), which is a model prediction of what the measured process variable, y(t), would be if
there were no dead time in the process (or alternatively, a prediction of what y(t) will be one
dead time into the future).

239
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 2) yideal(t) is then fed into a dead time model where it is stored until one dead time, θP, passes. At the
instant that the yideal(t) is stored, a previously stored yprocess(t) is released. This yprocess(t) is the
value of yideal(t) that was computed and stored one dead time ago. Hence, yprocess(t) is a model
prediction of the current value of y(t).

ysetpoint u(t) Actual y(t)

PID
+– Process
Controller

Predict y(t) Behavior yideal(t) +

As If There Were +- +
No Dead Time yideal(t) – yprocess(t)
Process Model
Add Time Delay
to “Ideal” yprocess(t)
Prediction of y(t)
Dead Time Model
Model Internal to
Controller Architecture

( y(t) – yprocess(t)) + yideal(t)

1. If the model is exact, 2. The controller then receives a
this term is zero zero dead time “measurement”
predicted by the model

Figure 22.2 - The Smith predictor model based controller architecture

The ideal and dead time model predictions are combined with the actual process measurement
prior to every control action to form the Smith controller error, e*(t), as:

e*(t) = ysetpoint(t) − [ y(t) − yprocess(t) + yideal(t)] (22.1)

If the model exactly describes the dynamic behavior of the actual process, then:

y(t) − yprocess (t) = 0 (22.2)

or
[ y(t) − yprocess(t)] + yideal(t) = yideal(t) (22.3)

and so for a perfect model, the Smith predictor error, e*(t), going to the controller is:

e*(t) = ysetpoint (t) − yideal(t) (22.4)

240
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
Hence, if the model exactly describes process behavior, then the Smith predicted error going to the
controller, rather than being the traditional e(t) = ysetpoint(t) − y(t), becomes the difference between the
set point and a prediction of what the measured process variable would be if there were no dead time
in the process.

Of course, the dynamic model will never exactly describe the true behavior of a process and
the logic presented above does not reflect this. It is safe to assume, however, that the ability of the
Smith predictor to reduce the influence of dead time on controller performance is directly related to
how well the model indeed describes the actual process dynamics. It should also be noted that a very
poor match between the model predictions and the actual process dynamics invites disaster, including
creating an unstable closed loop system for an otherwise well behaved process.

22.4 Exploring the Smith Predictor Controller

To isolate and explore the impact of dead time on controller performance and to establish the benefits
of the Smith predictor control algorithm, we use LOOP-PRO's Custom Process module. This study
proceeds as follows:
- a second order without dead time process is created in Custom Process and a PI controller is
designed and validated. The control objective is a 10% peak overshoot ratio and complete settling
of dynamics in one cycle of the measured process variable when the set point is stepped from
50% up to 55%.
- significant dead time is added to the original second order without dead time process and the
resulting degradation in controller performance is explored.
- a Smith predictor is designed and implemented to compensate for the added dead time and to
restore the original set point tracking performance.

PI Control of a Second Order Without Dead Time Process

The second order without dead time process is implemented in Custom Process with the parameters:

Process gain, KP = 1.2

First time constant, τP1 = 10 min
Second time constant, τP2 = 7 min
Process dead time, θP = 0 min

Thus, the process has a steady state gain of 1.2 (no units specified) and two time constants (which is
why it is second order) of 10 minutes and 7 minutes.
As shown in Fig 22.3, controller tuning is based on a pulse test where the controller output is
stepped from 50% up to 55% and back again. A Design Tools fit of a FOPDT model to the pulse test
data is also shown in the figure (a doublet test is normally recommended for controller design so data
is generated both above and below the design level of operation to average nonlinear effects. Since
Custom Process simulations are linear, a pulse test provides equally valid data.)
The FOPDT model parameters computed by Design Tools and shown at the bottom of the
Fig. 22.3 are:
Process gain, KP = 1.2 (unitless)
Overall time constant, τP = 13.7 min
Apparent dead time, θP = 3.7 min

241
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 FOPDT Fit of Second Order Without Dead Time Process
Model: First Order Plus Dead Time (FOPDT) File Name: NoDeadTime.txt
56 KP = 1.2

Process Variable
54 Process τP1 = 10
data τP2 = 7
52
θP = 0
50
FOPDT fit
Controller Output

56

54

52

50

0 50 100 150
Time

Gain (K) = 1.21, Time Constant (T1) = 13.69, Dead Time (TD) = 3.69
SSE: 8.82

Figure 22.3 - FOPDT model of second order without dead time pulse test data

It is interesting to note that a FOPDT model fit of data from this second order without dead time
process yields a rather significant apparent dead time (where “significant” dead time exists only in
comparison to the time constant of the process). This apparent dead time is the natural result of
approximating the dynamics of a higher than first order processes with the FOPDT form.

PI Controller Tuned for 2nd Order Without Dead Time Process

Process: Custom Process Controller: PID ( P= RA, I= ARW, D= off )
58
Process Variable

56
54
52 KP = 1.2
50 τP1 = 10
48 τP2 = 7
process has θP = 0
no dead time
Controller Output

80

70

60

50

0 20 40 60 80 100 120 140 160

Time (time units)
Tuning: Gain = 1.70, Reset Time = 13.7, Deriv Time = 0.0, Sample Time = 0.10

Figure 22.4 - IMC tuned PI controller produces the design 10% overshoot,
a rise time of 15 minutes and complete settling in 50 minute

242
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
Performance Degrades When Dead Time is Added
The Custom Process simulation is now modified by adding 5 minutes of dead time to the original
model. To appreciate the impact of dead time on controller performance, the PI tuning parameters
used to generate Fig. 22.4 are retained and the same set point step test is performed.
As shown in Fig. 22.5, the addition of dead time results in seriously degraded performance.
The peak overshoot ratio has climbed from 10% up to about 70%, and complete settling of dynamics
has increased from 50 minutes up to well over 150 minutes with multiple cycles of the measured
process variable.
Performance of Same Controller After Dead Time is Added
Process: Custom Process Controller: PID ( P= RA, I= ARW, D= off )
58
Process Variable

56
54
52 KP = 1.2
50 τP1 = 10
48 τP2 = 7
added dead time θP = 5
degrades performance
Controller Output

80

70

60

50

0 20 40 60 80 100 120 140 160 180

Time (time units)
Tuning: Gain = 1.70, Reset Time = 13.7, Deriv Time = 0.0, Sample Time = 0.10

Figure 22.5 – Using the PI tuning of Fig. 22.4, the added dead time causes
a 70% overshoot and complete settling in excess of 150 minutes

One method of achieving the design performance criteria of a 10% peak overshoot ratio and
complete settling within one cycle of the measured process variable is to detune the controller. Since
the process has been altered to include significant dead time, the most efficient way to determine the
more sluggish tuning values is to follow the identical design procedure just detailed.
Hence, as shown in Fig. 22.6, a FOPDT model is fit to pulse test data collected form the
second order plus dead time (SOPDT) process. At the bottom of the Fig. 22.6 are the FOPDT model
parameters computed by Design Tools:

Process gain, KP = 1.2 (unitless)

Overall time constant, τP = 13.7 min
Apparent dead time, θP = 8.7 min

Using IMC tuning in this study, the closed loop time constant, τC, is computed as the larger of
0.1τP or 0.8θP, and thus τC = 2.95 minutes. Substituting this τC and the above FOPDT model
parameters into the IMC correlations yields the PI tuning values:

243
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Controller Gain, KC = 1.7 (unitless)
Reset Time, τI = 13.7 min

The capability of the PI controller in tracking step changes in set point for the second order
without dead time process is shown in Fig. 22.4. For a set point step from 50% up to 55%, the
controller achieves the desired peak overshoot ratio of about 10% with a rise time of about 15 minutes
and complete settling in one cycle of the measured variable in about 50 minutes. This control
performance becomes the base case for the subsequent investigations.

FOPDT Fit of Second Order Plus Dead Time Process

Model: First Order Plus Dead Time (FOPDT) File Name: YesDeadTime.txt
56 KP = 1.2
Process Variable

54 Process τP1 = 10
data τP2 = 7
52
θP = 5
50
FOPDT fit
Controller Output

56

54

52

50

0 50 100 150
Time

Gain (K) = 1.21, Time Constant (T1) = 13.65, Dead Time (TD) = 8.69
SSE: 8.94

Figure 22.6 - FOPDT model of second order with dead time pulse test data

The new PI tuning values computed from the IMC correlations for the SOPDT process are:

Controller Gain, KC = 0.7 (unitless)

Reset Time, τI = 13.7 min

As shown in Fig. 22.7, this tuning produces the design performance criteria of a 10%
overshoot ratio and complete settling within one cycle of the measure variable. However, rise time
has doubled from the base case 15 minutes up to about 30 minutes. Complete settling now occurs in
80 minutes, a dramatic increase over the 50 minutes of the no dead time base case of Fig. 22.4.

The Smith Predictor Restores Performance

LOOP-PRO permits implementation of a PI with Smith predictor controller by selecting that option
from the list of controllers and entering the PI tuning values and the Smith predictor model values
where indicated in the controller design form. The model fit shown in Fig. 22.6 from our previous
controller design provides a FOPDT model that describes the dynamics of the SOPDT process. This
will serve as our predictive process model.

244
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 PI Controller is Detuned to Original Peak Overshoot Ratio (POR)
Process: Custom Process Controller: PID ( P= RA, I= ARW, D= off )
58

Process Variable
56
54
52 KP = 1.2
50 τP1 = 10
48
rise time increases
τP2 = 7
θP = 5
Controller Output

80

70

60

50

0 20 40 60 80 100 120 140 160

Time (time units)
Tuning: Gain = 0.70, Reset Time = 13.7, Deriv Time = 0.0, Sample Time = 0.10

Figure 22.7 - Detuning the controller produces the design 10% overshoot, but rise time has
doubled to 30 minutes and complete settling now takes 80 minutes

Tuning presents a challenge, however, because there are no correlations readily available for
PI controllers used in a Smith predictor architecture. Here we get creative by recognizing that the
Smith predictor theoretically eliminates the impact of dead time on control performance. Thus, we
will use the standard IMC tuning correlations, but in the calculation we will employ an artificial or
theoretical minimum value for dead time.
Recall that sample time, T, should ideally be less than or equal to 0.1τP. And for commercial
control equipment, one sample time always passes between a control action and when the controller
receives the next measurement. Hence, the minimum that dead time can be on a commercial control
system, in theory anyway, is one sample time.
So in the IMC tuning correlation, we will use the process gain and time constant from the
actual FOPDT fit of the process data as shown in Fig. 22.6, but for dead time in the correlation we
will use θP = 0.1τP = 1.4 min, or:

Process gain, KP = 1.2 (unitless)

Overall time constant, τP = 13.7 min
Apparent dead time, θP = 1.4 min

Using these parameters in the IMC correlation yields the tuning values for the PI with Smith predictor
architecture:

Controller Gain, KC = 4.1 (unitless)

Reset Time, τI = 13.7 min

245
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
As shown in Fig. 22.8, the PI with FOPDT Smith predictor produces very close to the design
performance criteria of a 10% overshoot ratio and complete settling within one cycle of the measure
variable. Rise time approaches 20 minutes, which equals the base case 15 minutes plus the 5 minutes
of dead time added to the base case process. Complete settling occurs in 55 minutes and this equals
our base case settling of 50 minutes plus the 5 minutes of dead time.

First Order Smith Predictor Improves Controller Performance

Process: Single Loop Custom Process Controller: PID with Smith Predictor
58
Process Variable

56
54
52 KP = 1.2
50 τP1 = 10
48
τP2 = 7
θP = 5
Controller Output

80

70

60

50

0 20 40 60 80 100 120 140 160

Time (time units)
Tuning: Gain = 4.10, Reset Time = 13.7, Deriv Time = 0.0, Sample Time = 0.10
Process Model: Gain(Kp) = 1.21, T1 = 13.65, T2 = 0.0, TL = 0.0, TD = 8.69

Figure 22.8 - PI controller with FOPDT Smith predictor produces the design 10% overshoot,
a rise time of 20 minutes and complete settling in 55 minutes

Though the design performance criteria are met, one observation is that the controller works
uncomfortably hard to achieve this success as evidenced by the large swings in the controller output
signal trace. As we learn in Fig. 22.9, this behavior is the result of using the FOPDT model to predict
the dynamics of a second order plus dead time process.
In Fig. 22.9 we use the original PI controller tuning values from Fig. 22.4, but now the
identical SOPDT model used to simulate the process in Custom Process is used as the predictive
process model in the Smith predictor architecture. The potential of the MPC architecture is
demonstrated in that the exact controller output signal and measured process variable traces evolve
here as they did in the base case of Fig. 22.4. The only difference is that there is a 5 minute shift in
the measured process variable in Fig. 22.9 because of the process dead time added to the simulation.

246
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Second Order Smith Predictor Restores Controller Performance
Process: Custom Process Controller: PID with Smith Predictor
58

Process Variable
56
54
52 KP = 1.2
50 τP1 = 10
48
τP2 = 7
θP = 5
Controller Output

80

70

60

50

0 20 40 60 80 100 120 140 160 180

Time (time units)
Tuning: Gain = 1.70, Reset Time = 13.7, Deriv Time = 0.0, Sample Time = 0.10
Process Model: Gain(Kp) = 1.20, T1 = 10.0, T2 = 7.00, TD = 5.00, TL = 0.0

Figure 22.9 - PI controller with SOPDT Smith predictor produces the design 10% overshoot,
a rise time of 20 minutes and complete settling in 55 minutes

247
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 23. DMC - Single Loop Dynamic Matrix Control
23.1 Model Predictive Control
Model predictive control (MPC) refers to a class of control algorithms that have a dynamic model of
the process programmed into the control architecture. The function of the model is to predict the
future behavior of the process based upon past controller moves and the current state of the process.
At each sample time, the next controller move is computed from a comparison of this predicted future
behavior with the desired set point trajectory.
Expanding this explanation, the MPC strategy begins with a performance objective function,
or mathematical equation that defines "good" control. This objective function typically combines
controller error and controller effort into a single formulation. If the measured process variable is
maintained at its set point over the predicted future (controller error is zero), the controller
performance is certainly good. If the control effort (the size of each control move) is small, then the
mechanical components of the final control element won't wear excessively and the process will not
experience unsettling sudden changes. Thus, by finding the mathematical minimum of the objective
function, control actions of modest size will be computed to drive the future predicted controller error
to zero.
Future set point tracking performance in the objective function is computed using the
dynamic process model. The model predicts the future of the measured process variable using past
and future (yet to be computed) controller output moves. The objective function is minimized by
computing a series of future controller output moves over the control horizon (the distance into the
future being considered) that balance set point tracking performance with controller effort. Only the
first of these controller output moves is implemented before repeating the entire procedure at the next
sample time.
Model predictive control is growing in popularity for its reasonably intuitive approach, its
ability to control processes with complex dynamics such as large dead time, inverse response and
unstable open loop dynamics, and its convenient handling of multivariable control challenges. MPC
offers additional benefits because process constraints and variable set point trajectories can be directly
addressed in the control calculations. Unfortunately, the design of a single loop MPC strategy
requires specifying at least five adjustable tuning parameters prior to implementation. For
multivariable implementations, the tuning burden grows substantially.

23.2 Dynamic Matrix Control

Dynamic matrix control (DMC) is one MPC strategy experiencing rapid growth in the chemical
process industries. DMC does not compete with classical PID controllers. Rather, it is typically
implemented in a hierarchy above a set of traditional PID loops. Such an approach is well suited for
multivariable control, where the actions of x controller outputs each affect y process variables. It also
permits the flexible handling of constraints. These can be specified explicitly as hard constraints (e.g.,
process limits) or implicitly as soft constraints (e.g., economic objectives).
Here we focus on the design and tuning of a single-loop DMC controller to illustrate design
concepts and provide a handle on the tuning strategy, both of which extend directly to
implementation of the multivariable case. The tuning strategy for a DMC controller can be reduced to
a ‘recipe’ as detailed in Table 23.1. After summarizing the theory, the use of this table is
demonstrated in an example.

248
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Current
Past Future

Process Variable / Set point

desired set point

y$ (n) set point tracking

y(n) error
predicted process variable profile
}d(n) y$ (n+j)
x x
x
x x
x
n n+1 n+2 n+3 n+P

Model horizon, N Prediction horizon, P

controller output
move implemented ⎧ controller output assumed
Controller output

∆u(n) ⎨ constant beyond control horizon

⎩

computed controller output profile, ∆u(n+j)

n n+1 n+2 n+3 n+M

Control horizon, M

Time (in samples)

Figure 23.1 - The moving horizon architecture

DMC is a ‘moving horizon’ MPC strategy. As shown in Fig. 23.1, moving horizon
controllers employ a model internal to the controller architecture to predict the future measured
process variable profile, y$ (n + j ) (j = 1, 2, …, P). P is the prediction horizon and represents the
number of sample times into the future over which DMC predicts the measured process variable.
The predicted process variable profile, yˆ (n +j ) , is computed using the recent history of
controller output moves, ∆u (n − j ) (j = 1, 2, …, N) and a finite step response process model:

j N -1
yˆ (n + j )
1 424 3
= yss + ∑ ai ∆u (n + j − i ) + ∑ ai ∆u (n + j − i ) + d{ ( n) (23.1)
i =1 i = j +1
Predicted process 144 42444 3 144424443 Prediction error
variable profile Effect of current and Effect of past moves
future moves

249
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
In Eq. 23.1, yss is the initial steady state of the measured process variable and ∆ui = ui − ui −1 is the
change in the controller output at the ith sample time. Also, a i (i = 1, 2, …, N) are the unit step
response coefficients as discussed in the next section. N is the model horizon and represents the
number of past controller output moves used by DMC to predict the future process variable profile.
The prediction error in Eq. 23.1, d(n), is the difference between the prediction of the current
value of the process variable and its actual measured value, or d (n) = y (n) − y$ (n) . This prediction
error is added to the predicted process variable profile, y$ (n + j ) , to correct for any unmeasured
disturbances or inaccuracies due to plant-model mismatch.
The predicted process variable profile computed above is subtracted from the set point
profile, ysp (n + j ) , over the next P sample times, squared, and then summed to yield the sum of the
squares of the set point tracking error. The DMC objective function, to be minimized at every sample
time, is then represented as

P M
∑{ysp (n + j ) − y(ˆ n + j )}
2
∑{∆u(n + i − 1)}
2
Min J= +λ (23.2)
j =1 i =1
144 42444
3
14444244443
Set point tracking Penalty on controller
error output move sizes

Here, λ is a positive constant that weighs the controller output move sizes relative to the sum of the
squares of the set point tracking error. The penalty on the controller output move sizes is introduced
into the DMC objective function to suppress otherwise aggressive control actions when the control
horizon, M, is greater than one.
Before the optimization problem posed in Eq. 23.2 can be solved, the set point tracking error
must be expressed in terms of the controller output moves that have already occurred and those to be
determined, as follows:

ysp (n + j ) − yˆ (n + j )
N −1 j
= ysp (n + j ) − yss − ∑ ai ∆u (n + j − i ) − d ( n) − ∑ ai ∆u (n + j − i) (23.3)
i = j +1 i =1
144 42444
3
144444444244444444
3
Predicted error based on Effect of current and future
past moves, e ( n +j) moves to be determined

where j = 1, 2, …, P. Eq. 3 is a linear system of P equations that can be represented in a matrix form,
assuming that the controller output is held constant beyond the control horizon, M:

250
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
y sp − ŷ =

⎡ e( n +1) ⎤ ⎡ a1 0 0 L 0 ⎤
⎢ e( n + 2 ) ⎥ ⎢a a1 0 0 ⎥ ∆u( n )
⎢ ⎥ ⎢ 2 ⎥ ⎡ ⎤
⎢ e( n + 3 ) ⎥ ⎢ a3 a2 a1 O 0 ⎥ ⎢ ∆u( n +1) ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ M ⎥ − ⎢ M M M 0 ⎥ ⎢ ∆u( n + 2 ) ⎥
⎢e( n + M ) ⎥ ⎢a M a M −1 a M − 2 ⎢ ⎥
a1 ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥
⎢ M ⎥ ⎢ M M M M ⎥ ⎢⎣∆u( n + M − 1) ⎥⎦
1444 4244M ×3
441
⎢ e( n + P ) ⎥ ⎢a ⎥
⎣ ⎦ P ×1 ⎣ P a P −1 a P − 2 L a P − M +1 ⎦ P × M Current and future controller
144244 3 144444444244444444 3 output moves to be determined, ∆u (23.4)
Predicted error based DMC Dynamic Matrix, A
on past moves, e

or in a compact matrix notation as:

y sp − ˆy = e − A ∆u (23.5)

Here, y sp is the vector of future set points, y$ is the vector comprising the predicted process variable
profile, e is the vector of predicted errors over the next P sampling intervals based on past controller
output moves, A is the DMC dynamic matrix, and ∆u is the vector of M controller output moves to
be determined.
With the transformation of the set point tracking error as in Eq. 23.5, the DMC objective
function can be alternatively written as:

Min J = [e − A∆u ]T [e − A∆u ]+ [∆u ]T λ[∆u ] (23.6)

∆u

In the unconstrained case, Eq. 23.6 represents a least squares optimization problem that has a closed
form solution representing the DMC control law (e.g., García and Morshedi, 1986):

∆u = ( AT A+λ I )−1 AT e (23.7)

Adding constraints to this classical formulation (Eq. 23.6) produces the Quadratic Dynamic Matrix
Control (QDMC) (e.g., García and Morshedi, 1986) algorithm. The constraints considered in this
work include:

ˆymin ≤ ˆy ≤ ˆymax (23.8a)

∆umin ≤ ∆u ≤ ∆umax (23.8b)
umin ≤ u ≤ umax (23.8c)

23.3 The DMC Process Model

DMC requires that the dynamic process model be expressed both as the finite step response model
(Eq. 23.1) and the DMC dynamic matrix (Eq. 23.4). Both forms rely on the unit step response
coefficients, ai (i = 1, 2, …, N), generated from the actual process.
Step response data is generated by introducing a positive step in the controller output with the
process at steady state and the controller in manual mode. From the instant the step change is made,

251
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
the process variable response is recorded as it evolves and settles at a new steady state. For a step of
arbitrary size, the response data is normalized by dividing through by the size of the controller output
step to yield the unit step response. For processes with a modest steady-state gain, a unit step in the
controller output can be used directly to obtain the unit step response. In either case, it is necessary to
make the controller output step large enough such that noise in the process variable measurement
does not mask the true process behavior.
Discrete points at every sample time along the unit step response are collected (Fig. 23.2) to
yield the unit step response coefficients:

Unit step response coefficients = [ a1 , a2 , a3 , L , aN ] (23.9)

The unit step response coefficients in Eq. 23.9 are used in Eq. 23.1 to compute the predicted process
variable profile at every sample time. Also, the first P of the N coefficients are cast in a matrix form
as shown in Eq. 23.4 to obtain the DMC dynamic matrix. This matrix can then be used directly in the
DMC control law (Eq. 23.7). Notice, however, that both the sample time, T, and the prediction
horizon, P, have to be known before the finite step response model or the DMC dynamic matrix can
be designed.

aN x x x
x x
Process Variable

x
a6 x new steady
a5 x State

initial steady
a4 x
State
a3 x
0 x x
0 1 2 3 4 5 6 N

a1 = a2 = 0 due to dead time

Controller Output

1
unit step in
controller output
0
0 1 2 3 4 5 6

Time (in sampling intervals)

Figure 23.2 - Generating the unit step response coefficients

23.4 Tuning DMC

Tuning DMC, even for a single-input single-output (SISO) process, is challenging because of the
number of adjustable parameters that affect closed loop performance. As listed in Appendix C, these

252
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
include: a finite prediction horizon, P; a model horizon, N; a control horizon, M; a move suppression
coefficient, λ; and a sample time, T.
Step 1 involves fitting a first order plus dead time (FOPDT) model to actual controller output
to measured process variable data. Reasonable estimates of the FOPDT model parameters, i.e., the
steady state gain, Kp, overall time constant, τp, and effective dead time, θp, are essential to the success
of this tuning strategy. It is important to stress that this model is used in the tuning strategy only and
that Eq. 23.1 through 23.7 used in the implementation of DMC are formulated from the actual process
data obtained as described above (Fig. 23.2 and Eq. 23.9).
Step 2 involves the selection of an appropriate sample time, T. The estimated FOPDT model
parameters provide a convenient way to select T. If the designer does not have complete freedom to
select sample time equal to the value computed, then it should be picked as close as possible to this
recommended value.
Step 3 computes a model horizon, N, and a prediction horizon, P, from τp, θp and T. Note that
both N and P cannot be selected independent of the sample time, T. Also, it is imperative that N be
equal to the open loop process settling time in samples to avoid truncation error in the predicted
process variable profile.
Step 4 requires the specification of a control horizon, M. Recommended values of M are such
that M × T is larger than the time required for the slowest open loop response to reach 60% of the
steady state. A convenient way to select M is to compute an integer value using τp and T as shown in
Appendix C. Selecting M > 1 can be very useful to the practitioner since this provides advance
knowledge of the impending controller output moves.
Step 5 involves computation of a move suppression coefficient, λ. With M = 1, the need for a
move suppression is eliminated and λ is set equal to zero. However, if M > 1, a positive move
suppression coefficient of appropriate magnitude is essential to suppress otherwise aggressive control
action.
As with all controllers, it may be necessary to perform final on-line tuning. The best single
tuning parameter for performance adjustment is the move suppression coefficient, λ. Increasing λ
produces smaller controller output move sizes and slower process variable response.

23.5 Example Implementation

The design and tuning of DMC are demonstrated on a perfect DMC controller. A perfect DMC
controller employs the identical models in the DMC step response model as is used for the process
simulation. Be aware that in real-world applications, most DMC step response models will differ
from the true process behavior. Hence, the DMC controller capabilities shown here should be
considered as the best possible performance.
The ideal process simulation is a second order plus dead time model (SOPDT) overdamped
linear model of the form:
Process Gain, KP = 2.0
First Time Constant, τP1 = 20.0 time units
Second Time Constant, τP2 = 10.0 time units
Dead Time, θP = 10.0 time units

The first step in tuning is to fit a FOPDT model to process data. For this example, Design
Tools yields the model parameters for the ideal process as Kp = 2.0 PV units/CO unit, τp = 24.1 time
units, θp = 13.8 time units. Based on the criteria presented in Step 2, a sample time of 7.0 time units is
selected. For the FOPDT parameters estimated and the value of T selected above, the prediction
horizon, P, and the model horizon, N, are computed in Step 3 to be 19, the open loop settling time of

253
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
the ideal process in samples. Next, a control horizon, M, of 5 is computed in Step 4. Finally, in Step
5, an appropriate move suppression coefficient of 23.9 is computed for the ideal process.
With the tuning parameters selected, the unit step response coefficients are determined for the
ideal process for a sample time of 7.0 time units. In the initialization of the DMC controller, the step
response coefficients are generated for you based on the values of T, P and N. The model used to
generate the step response coefficients is entered as a transfer function on the design menu. The user
wants to use the transfer function that best represents the process. In this case, the transfer function
will be a SOPDT overdamped model. The model parameters listed above are entered into the design
menu. With the design complete, DMC is implemented and its performance tested for the ideal
process.
The closed loop performance achieved for the ideal process, with DMC tuned using the
tuning strategy in the appendix, is shown in Fig. 23.3. The top half of this plot shows the process
variable response for a step change in set point. The lower half of the plot shows the controller output
moves made by DMC to achieve the desired set point tracking.
Desirable closed loop performance is defined as the process variable response to a step
change in set point exhibiting a modest peak overshoot (say, less than 10%). Also, the corresponding
controller output move sizes should not exceed 2 to 3 times the final change in controller output. As
shown in Fig. 23.3, such a performance is achieved.

Figure 23. 3 - DMC performance for the ideal process

Effect of Tuning Parameters on DMC Performance

The significance of the tuning strategy can be illustrated in a sensitivity study where the effect of each
tuning parameter on closed loop performance is individually explored. For comparison, the DMC
tuning and design demonstrated in Fig. 23.3 is labeled as the ‘base case’ in Figures 23.4 through
23.10.
From left to right, Fig. 23.4 shows the impact of reducing the model horizon, N, from 38 to
19 (base case) and then to 10 and finally to 5. With the reduced value of N, the now truncated step
response model in DMC is able to predict the effect of controller output moves for only 5 sample
times (plus the dead time) after a set point change. Beyond 5 sample times, even though the past

254
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
controller output moves continue to impact the process variable, their effect is not realized in what
amounts to an incorrect computation from the truncated model profile. As in this case, such truncation
generally results in unpredictable and undesirable controller performance.
From left to right, Fig. 23.5 shows the impact of reducing the prediction horizon, P, from 19
to 8 and finally to 3. For small values of P the process variable response shows a longer rise time.
Note that this change occurs abruptly as the value of P becomes very small rather than occurring
gradually as the value of P decreases. This performance change results because: i) the controller
output moves computed by DMC seek to eliminate the predicted error over a fewer number of future
sampling intervals, ii) the inability of DMC with a small P to predict the process behavior far enough
ahead to realize that the past moves will bring the process variable to the new set point and beyond,
and iii) the presence of a rather large move suppression coefficient, appropriate for the base case (P =
19), but resulting in very small moves for the case with the reduced prediction horizon (P = 3).

N = 38 N = 19 N = 10 N=5
(base case)

Figure 23.4 – Effect of Varying Only the Model Horizon, N, from the Base Case

As stated previously, the appropriate value for the move suppression coefficient is related to
the prediction horizon, P. From left to right, Fig. 23.6 shows the impact of reducing the move
suppression coefficient from 192 to 23.9 (base case) and then to 3 and finally to 0.375 for constant
values of sample time (T = 7.0 time units), model horizon (N = 19), prediction horizon (P = 19), and
control horizon (M = 5). As shown below, increasing the move suppression coefficient for a constant
prediction horizon can dramatically increase the rise time, resulting in a slower response and poor
control. This illustrates the impact of the move suppression coefficient upon the nature of the
response.

255
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 P = 19 P=8 P=3
(base Case)

Figure 23.5 – Effect of Varying Only the Prediction Horizon, P, from the Base Case

As the move suppression coefficient decreases, the controller output moves become
increasingly aggressive resulting in a process variable response that oscillates with some overshoot.
Decreasing the move suppression coefficient reduces the penalty on the move sizes made by DMC
resulting in more aggressive moves. Mathematically, this can also be explained as the ineffectiveness
of a small move suppression coefficient in conditioning the inherently ill-conditioned system matrix,
A T A , in the DMC control law (Eq. 23.7).
The model and prediction horizons, N and P, respectively, should have the same value.
Figures 23.4 through 23.5 demonstrate the individual effects of changing N and P away from one
another, but the effect of changing both parameters together is also very interesting. From left to
right, Fig. 23.7 illustrates the impact of reducing both N and P from 38 to 19 (base case) to 10 to 5
and finally to 3 for constant values of sample time (T = 7.0 time units), control horizon (M = 5), and
move suppression coefficient (λ = 23.9). Notice how the response becomes more oscillatory as N and
P decrease until becoming overdamped and sluggish once P equals 3. Therefore, the coupled impact
of a low prediction horizon (P = 3) and a large move suppression coefficient (λ = 23.9) overtakes the
gradual effect of decreasing both the model and prediction horizons.
From left to right, Fig. 23.8 illustrates the effect of reducing the control horizon, M, from 10
to 5 (base case) to 3 and finally to 1 with the sample time (T = 7.0 time units), model and prediction
horizons (N = P = 19), and move suppression coefficient (λ = 23.9) held constant. The difference in
performance between the base case and the other cases is rather insignificant, especially in the
presence of a significant move suppression coefficient.

256
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 λ = 192 λ = 23.9 λ=3 λ = 0.375
(base case)

Figure 23.6 – Effect of Varying Only the Move Suppression Coefficient, λ, from the Base Case

N = 19
N = 38 P = 19 N = 10 N=5 N=3
P = 38 (base case) P = 10 P=5 P=3

Figure 23.7 – Effect of Varying the Model and Prediction Horizons Together from the Base Case

For the base case with its large control horizon (M = 5), the faster process variable response,
slight increase in overshoot and increased activity of controller output moves is due to the added
degrees of freedom from the extra moves computed at every sample time. With a larger number of
moves computed at every sampling interval, DMC can afford to make more aggressive first moves
that are compensated for by the extra moves available. Since only the first move is actually
implemented while the others are discarded for a new set of M moves computed at the subsequent
sample time, the result is a more aggressive response.

257
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 M = 10 M=5 M=3 M=1
(base case)

Figure 23.8 – Effect of Varying Only the Control Horizon, M, from the Base Case

From left to right, Fig. 23.9 demonstrates the impact of changing the sample time, T, from 14
to 7 (base case) to 3.5 to 1.75 time units. Note that these changes occurred without changing any of
the other DMC parameters from their base case values (as calculated based for T = 7 time units).
Notice how the response grows more oscillatory as T decreases.
Similarly, Fig. 23.10 also demonstrates the impact of changing T from 14 to 7 (base case) to
3.5 to 1.75 time units. However, note that the other DMC tuning parameters were changed along
with the sample time, using the applicable tuning rules found in the appendix. In Table 23.1 below,
you will find the DMC tuning parameters used for each step, from left to right, in Fig. 23.10.

Example T (time units) N P M λ

(a) 14 9 9 2 5.66
(b) 7 19 19 5 23.9
(c) 3.5 39 39 11 97.34
(d) 1.75 78 78 22 366.7

Table 23.1 – Changes in DMC Tuning Parameters with Sample Time, T

It is important to mention here that when the sample time is altered, the unit step response
model internal to the DMC architecture should also be appropriately updated to eliminate plant-model
mismatch. LOOP-PRO calculates the step response model from a linear model of the process and T,
both input by the user. The linear model used does not change with T, but the step response model
computed from it by LOOP-PRO does. However, the step response model should change whenever T
changes if the step response model is input directly into the controller without such a calculation.
Specifically, the updated model should be formed from unit step response coefficients, ai, collected at
the new sampling interval, T.

258
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 T = 14 T=7 T = 3.5 T = 1.75
(base case)

Figure 23.9 – Effect of Varying Only the Sample Time, T, from the Base Case

(a) T = 14 (b) T = 7 (base case) (c) T = 3.5 (d) T = 1.75

Figure 23.10 – Effect of Varying the DMC Tuning Parameters with T from the Base Case

Notice how the response obtained in Fig. 23.9 shows a highly erratic set of controller output
moves accompanied with an irregular, underdamped process variable response while the responses in
Fig. 23.10 are quite similar. The cause for this difference in behavior can be noticed from the tuning
strategy found in Appendix C, where the computation of all of the tuning parameters depends on the
choice of the sample time. Hence, changing the sample time without updating all of the other DMC
tuning parameters adversely affects closed loop performance. Figures 23.9 and 23.10 above
demonstrate the importance of changing N, P, M, and λ along with any changes in the sample time, T.

259
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
Effect of Constraints on DMC Performance
The significance of constraints can be illustrated using the following example. For both step tests
shown below, the tuning parameters are given as T = 7.0 time units, P = N = 19, M = 5, and λ = 23.9,
and the constraints are:
45 ≤ ŷ ≤ 60
−1.0 ≤ ∆u ≤ 1.0
50.0 ≤ u ≤ 52.5

The first plot in Fig. 23.11 shows the performance achieved by DMC tuned for the case when
T = 7, P = N = 19, M = 5, and λ = 23.9 without the above constraints. The second plot represents the
performance when constraints listed above are included in the controller design. Notice that when
constraints are included the move sizes computed by DMC for the controller output are not greater
than 0.5, and the controller output is constrained at values of 50% and 52.5%. When constraints are
included, the process variable response is significantly slower than it was without constraints.

Base Case (Without Constraints) Base Case (With Constraints)

Figure 23.11 – Effect of Constraints on the Base Case

23.6 Chapter Nomenclature

ai = ith unit step response coefficient
A = dynamic matrix
d = prediction error
e = predicted error
e = vector of predicted errors
I = index
I = identity matrix
j = index of sampling intervals
k = discrete dead time
Kp = process gain
M = control horizon (number of controller output moves)
n = current sampling interval

260
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 N = model horizon (process settling time in samples)
P = prediction horizon
T = sampling interval
u = controller output
y = process variable
yss = initial steady state of process variable
ŷ = predicted process variable
ysp = process variable set point

Greek Symbols
∆ui = change in controller output at the ith sampling interval
θp = effective dead time of process
λ = move suppression coefficient (controller output weight)
τp = overall process time constant

261
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
23.7 Tuning Strategy for Single Loop DMC

1. Approximate the manipulated-to-measured-process-variable dynamics with a first order plus dead

time (FOPDT) model:

-θ s
p
dy( t ) y( s ) K p e
τp + y( t ) = K p u( t − θ p ) or =
dt u( s ) τ p s +1

2. Select the sampling interval as close as possible to:

T = 0.1τ p or T = 0.5 θ p whichever is larger

3. Compute the prediction horizon, P, and the model horizon, N, as the process settling time in
samples (rounded to the next integer):

5τ p τp
P=N = +k where: k = +1
T T

4. Select the control horizon, M, as the time in samples required for the open loop response to reach
60% of the steady state:
τp
M= +k
T

5. Compute the move suppression coefficient:

⎧0 M =1
⎪
λ = ⎨ M ⎛ 3.5 τ p ( M −1) ⎞ 2
⎪ 10 ⎜⎜ T + 2 − 2
⎟⎟ K p M >1
⎩ ⎝ ⎠

6. Implement DMC using the traditional step response matrix of the actual process and the above
parameters.

262
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 25. Non-Self Regulating (Integrating) Processes

25.1 Open Loop Behavior of Integrating Processes

Up to this point we have explored only self regulating processes. Self regulating processes naturally
seek a steady state operating level if the manipulated and disturbance variables are held constant for a
sufficient period of time. The left plot of Fig. 25.1 shows the ideal open loop behavior of a self
regulating process. Here the controller output (CO) and process variable (PV) are initially at steady
state. The CO is stepped up and back from this steady state while the disturbances (not shown) remain
quiet. The PV responds to the step, but ultimately returns to its original operating level. This is a key
characteristic of a self regulating process.
Some processes are non-self regulating and these are commonly referred to as integrating
processes. Non-self regulating (integrating) processes are very challenging to control. After studying
this chapter, you may come to realize that some of the level, temperature, pressure, pH and other
loops you work with are integrating in nature.
The plot to the right in Fig. 25.1 shows the ideal open loop behavior of an integrating
process. As shown, the PV steadies at a different operating level when the CO returns to its original
value after the step. If the CO had not been stepped back to its original value (not shown), the PV
would not settle at a new steady state, but rather, would continue to drift upward to extreme and
possibly dangerous levels. Changes in any of the disturbance variables can also cause the PV to drift.
Hence, integrating processes are rarely operated in open loop for very long.

Self Regulating Behavior Integrating Behavior

g g
PV

PV returns to original level PV does not return to original

when CO returns level when CO returns
CO

40 60 80 100 120 40 60 80 100 120

Time (time units) Time (time units)

Figure 25.1 – Open loop behavior of self regulating (left) and integrating (right) processes

While the methodology for tuning a PID controller is the same for both self regulating and
integrating processes, the specifics are quite different. As we have learned in previous chapters, the
general controller tuning method for all processes is comprised of these steps:
- Collect dynamic process test data near the design level of operation
- Fit a linear dynamic model to the process test data
- Use the parameters from the model fit in correlations to obtain controller tuning values
- Test the tuning online and fine tune if necessary

263
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Control Station’s Pumped Tank Case Study is an integrating process. As shown in Fig. 25.2,
the measured process variable is liquid level. To maintain level, the controller output manipulates
flow rate out of the bottom of the tank by adjusting a throttling valve at the discharge of a constant
pressure pump. The disturbance variable is the flow rate of a secondary feed to the tank.

Figure 25.2 – Pumped Tank process

When the Pumped Tank controller output signal is increased from 70% to 80% as shown in
Fig 25.2, the discharge flow rate out of the bottom of the tank increases. The total feed flow rate to
the tank remains unchanged, however, and since flow out becomes greater than flow in, the tank level
begins to fall. As the situation persists, liquid level continues to fall until the tank is drained.
Conversely, if the controller output were to be decreased enough to cause flow rate out to be
less than flow rate in, the tank level would rise until full. If this were a real process, the tank would
overflow and spill, creating safety and profitability issues. As we will learn in the next sections, this
non-self regulating behavior makes the control of integrating processes quite challenging.

25.2 The FOPDT Integrating Model

The familiar first order plus dead time dynamic model does a good job of describing self regulating
process behavior but it is not able to reasonably model the behavior of integrating processes. For
controller tuning, integrating process behavior is best described with the FOPDT Integrating model
form:
dy (t )
= K P* u ( t − θ P ) (25.1)
dt

264
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Individual values for the familiar process gain, K P , and process time constant, τ P , cannot be
identified for the FOPDT Integrating model. Instead, an integrator gain, KP*, is defined that has units
of the ratio of the process gain to the process time constant, or:
KP
K P* = where KP* [=] y(t)/[u(t) * time] (25.2)
τP
Analogous to our FOPDT investigations, the FOPDT Integrating model parameters KP* and θP in Eq.
25.1 can be determined both graphically and using Control Station’s Design Tools.

25.3 Modeling Data From Integrating Processes

The graphical method of fitting a FOPDT Integrating model to process data requires a data set that
includes at least two constant values of controller output, u1 and u2. Both u1 and u2 must be held
constant long enough such that the slope of the measured process variable response trend can be
visually identified in the data.
The FOPDT Integrating model describes the process behavior at each value of constant
controller output u1 and u2 as:
dy (t )
= K P* u1 ( t − θ P ) (25.3)
dt 1
and
dy (t )
= K P* u 2 ( t − θ P ) (25.4)
dt 2

Subtracting Eq. 25.3 from Eq. 25.4 and solving for KP* yields:

dy (t ) dy (t )
−
dt 2 dt 1 slope2 − slope1
K P* = = (25.5)
u2 − u1 u2 − u1

The dead time, θP, is then estimated from the plot using the same method described in Chapter 3.

Example 1: Graphical Estimation of Model Parameters for Pumped Tank Process

Figure 25.3 shows open loop data collected from the Pumped Tank case study of Fig. 25.2. The
controller output is stepped from 68% up to 80%, causing the liquid level to fall. The controller
output is then stepped from 80% down to 65%, causing an upward slope in the liquid level.
The slope of each segment is calculated as the change in liquid level divided by the change in
time. From the data in Fig 25.3 we compute:

dy (t ) ∆PV1 3.1 − 4.1 m

slope1 = = = = − 0.25 (25.6)
dt 1 ∆t1 20 − 16 min
and
dy (t ) ∆PV2 3.4 − 3.0 m
slope 2 = = = = 0.10 (25.7)
dt 2 ∆t 2 30 − 26 min

265
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Process: Pumped Tank Cont.: Manual Mode
4.5
Process Variable
4.0 (30, 3.4)
(16, 4.1)
3.5
(26, 3.0)

3.0
(20, 3.1)
80
Controller Output

u1 = 80%
76

72
u2 = 66%
68

12 14 16 18 20 22 24 26 28 30 32 34
Time (mins)

Figure 25.3 – Open loop data for Pumped Tank processes

Using the two slopes computed in Eq. 25.6 and 25.7 along with their respective controller output
values in Eq. 25.3 yields the integrator gain, KP*, for the pumped tank process:

slope 2 − slope1 0.10 − (−0.25) m

K P* = = = − 0.025
u2 − u1 66 − 80 % ⋅ min

The process dead time is estimated using the same method outlined in Chapter 3. That is, θP is
computed as the difference in time from when the controller output signal was stepped (tUstep) and
the time when the measured process variable starts showing a clear response to that change
(tYstart). From the plot we estimate:

θP = tYstart − tUstep = 22 − 21 = 1.0 min

Substituting these model parameters into Eq. 25.1 yields the FOPDT Integrating dynamic model
describing the pumped tank process behavior at this particular level of operation:

dy (t )
= −0.025 u ( t − 1.0 )
dt
Õ Õ Õ

Design Tools offers two tools that automate the FOPDT Integrating model fit. The choice of
which tool to use depends on the nature of the process data you have collected. These should be
either:
- process data that starts at an initial steady state, or
- process data that has two periods of constant controller output

A process is considered to be at steady state if both the controller output signal and measured
process variable are substantially constant for a period of time before the collection of dynamic data

266
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
begins. For integrating processes, this situation might occur if the process is already under feedback
control and the disturbances are quiet. With data that starts at steady state, Design Tools can conduct
an automated model parameter search like those we have conducted numerous times throughout the
book using the FOPDT model.
If the process data does not start at steady state, as might be expected with open loop data,
then the data must include at least two periods where the controller output signal is held constant long
enough such that the slope of the measured process variable response is clearly evident in the data.
For such data, Design Tools offers assistance in performing the graphical analysis like that conducted
in example 1.

Example 2: Using Design Tools’ Graphical Analysis to Determine Model Parameters

The use of the graphical analysis tool to fit an FOPDT Integrating model for the same open loop
data used in Example 1 is shown in Fig. 25.4. After selecting the FOPDT Integrating model in
Design Tools and then specifying that the process data has two periods of constant controller
output, a plot is presented along with adjustable nodes, slopes and tie lines.
Two controller output (CO) nodes are displayed that should be clicked and dragged to match
the two values of constant controller output expected in the data. Both CO nodes have tie lines to
identify their associated measured process variable (PV) slope bar. Each slope bar has two end
point nodes. As shown, these should be clicked and dragged so that each bar approximates the
corresponding sloped segment on the graph.
With the six nodes (two CO and four PV) properly positioned, the FOPDT Integrating model
parameters KP* and θP result. The PV slope bars may need to be tweaked up or down until the
computed model overlays the data to your satisfaction.

PV nodes

tie line
CO node

Figure 25.4 – Using the graphical analysis tool to fit a FOPDT Integrating dynamic model

As shown in Fig 25.4, with the nodes properly positioned over the process data, the graphical
analysis tool yields FOPDT Integrating model parameters that are satisfactorily close to those
computed in example 1:

267
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 m
K P* = − 0.023 θ P = 1.2 min
% ⋅ min

Õ Õ Õ
Example 3: Using Design Tools’ Automated FOPDT Integrating Model Fit
When process data starts at steady state, such as that shown in Fig 25.5 for the pumped tank process,
then the most convenient modeling tool is the automated model fit like we have used in previous
chapters. The data in Fig 25.5 was collected with the disturbance flow rate quiet (unchanging) and the
process under P-Only control. A set point step initiated the dynamic event and the data below was
collected during the closed-loop response.

FOPDT Integrator fit

closed loop data

Figure 25.5 – Using the Graphical Integrator tool to determine model parameters

The Design Tools automated fit yields the FOPDT Integrating model parameters:

m
K P* = − 0.024 θ P = 1.0 min
% ⋅ min

These are again consistent with the parameter values computed in the previous two examples.

Õ Õ Õ

25.4 Designing and Implementing Controllers for Integrating Processes

Using the controller tuning guide for integrating processes found in Appendix D.2, we can determine
the parameters for any type of controller we wish to design.

268
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Example 5: Use IMC Tuning Correlations to Design a PI Controller for an Unstable Process
We start with the model parameters obtained by Design Tools in Example 2:

m
K P* = −0.023 θ P = 1.2 min
% ⋅ min

We first find the value of τC from the dead time, using Standard Tuning (see Appendix D.2):

τ C = θ P 10 = 1.2 10 = 3.8 min

Next, we can substitute KP*, θP, and τC into Eq. 25.6 to obtain the controller tuning values:

1 2τC + θ P 1 ⎡ 2(3.83) + 1.2 ⎤ %

KC = = ⎢ ⎥ = −15.3
K *p (θ P + τC ) 2 −0.023 ⎣ (1.2 + 3.8) ⎦
2 m

τ I = 2τC + θ P = 2(3.8) + 1.2 = 8.9 min

Because Design Tools also uses the IMC-derived correlations to determine the tuning parameters,
these values should match the ones given by the program.
Õ Õ Õ

Interaction of Tuning Parameters

The two PI tuning parameters (KC and τI ) interact with each other, meaning that if one is changed, the
response of the other may also be different. Figure 25.5 shows the base case performance, where the
actual controller gain and reset time computed by the tuning correlations are used. The effect on
controller performance when KC is changed is then presented.

Figure 25.5 – step response performance while varying KC for integrating process

While the first step response shows performance you may like to improve, the subsequent steps show
that performance degrades further by adjusting controller gain. Integrating process are indeed a control
challenge.

269
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
Appendix A: Derivation of IMC Tuning Correlations
A.1 Self Regulating Processes
For a general discussion of the derivation methodology and the details of the PI tuning correlation
derivation, please see Chap. 17, “Deriving PID Controller Tuning Correlations.”

A.1.a Ideal PID Control

Start with the general FOPDT process model:

K P e −θ P s
GP* ( s ) =
τ Ps +1

Substitute the 1/1 Padé approximation for e−θ P s :

1 − 0.5θ P s
e −θ P s ≈
1 + 0.5θ P s

K P (1 − 0.5θ P s )
so GP* ( s ) =
(τ P s + 1)(1 + 0.5θ P s )
Factor GP* ( s ) into invertible terms, GP* − ( s ) , and noninvertible terms, GP* + ( s) . Recall that an invertible
term will not yield positive poles (positive roots of the denominator of the transfer function indicating
unstable behavior) when taken in the reciprocal:

GP* ( s ) = GP* + ( s)GP* − ( s )

so GP* + ( s ) = (1 − 0.5θ P s )

KP
and GP* − ( s ) =
(τ P s + 1)(1 + 0.5θ P s )

Now express the IMC controller, GC* ( s ) , in terms of GP* − ( s ) and a first-order filter term, F(s):

1
GC* ( s ) = F (s)
GP* − ( s )

The filter term can be expressed in terms of a closed-loop time constant, τ C :

1
F ( s) =
τC s +1

Substitute F(s) and GP* − ( s ) to express the IMC controller as:

270
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 ⎛ (τ s + 1)(1 + 0.5θ P s ) ⎞ ⎛ 1 ⎞ (τ P s + 1)(1 + 0.5θ P s )
GC* ( s ) = ⎜⎜ P ⎟⎟⎜ τ s + 1 ⎟ =
⎝ KP ⎠⎝ C ⎠ K P (τ C s + 1)

Relate this IMC controller, GC* ( s ) , to a classical feedback controller, GC ( s ) :

GC* ( s )
GC ( s ) =
1 − GP* ( s )GC* ( s )

Substitute the above equations for GC* ( s ) and GP* ( s ) and simplify:

(τ P s + 1)(1 + 0.5θ P s )
K P (τ C s + 1)
GC ( s ) =
⎛ (τ P s + 1)(1 + 0.5θ P s ) ⎞ ⎛ K (1 − 0.5θ s ) ⎞
P P
1− ⎜ ⎟⎜ ⎟
⎜ K P (τ C s + 1) ⎟ ⎜ (τ P s + 1)(1 + 0.5θ P s ) ⎟
⎝ ⎠⎝ ⎠

(τ P s + 1)(1 + 0.5θ P s )
K P (τ C s + 1)
=
1−
(1 − 0.5θ P s )
(τ C s + 1)

=
(τ P s + 1)(1 + 0.5θ P s )
K P (τ C s + 1 − 1 + 0.5θ P s )

=
(τ P s + 1)(1 + 0.5θ P s )
K P (τ C s + 0.5θ P s )

=
(τ P s + 1)(1 + 0.5θ P s )
K P (τ C + 0.5θ P ) s

Factor this into a form we can compare with the ideal PID controller form:

τ P s + 1 + 0.5θ Pτ P s 2 + 0.5θ P s
GC ( s ) =
K P (τ C + 0.5θ P ) s

(τ P + 0.5θ P ) s + 1 + 0.5θ Pτ P s 2
=
K P (τ C + 0.5θ P ) s

271
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 GC ( s ) =
(τ P + 0.5θ P ) ⎛1 + 1
+
0.5θ Pτ P ⎞
⎜⎜ s⎟
K P (τ C + 0.5θ P ) ⎝ (τ P + 0.5θ P ) s (τ P + 0.5θ P ) ⎟⎠

Compare this to the classical feedback form for an ideal PID controller:

⎛ 1 ⎞
GC ( s ) PID Ideal = KC ⎜1 + +τ Ds ⎟
⎝ τIs ⎠

we obtain the following controller tuning correlations:

(τ P + 0.5θ P ) τ Pθ P
KC = , τ I = (τ P + 0.5θ P ) , and τ D =
K P (τ C + 0.5θ P ) 2 (τ P + 0.5θ P )

A.1.b Interacting PID Control

Start with the general FOPDT process model:

K P e −θ P s
GP* ( s ) =
τ Ps +1

Substitute the 1/1 Padé approximation for e−θ P s :

1 − 0.5θ P s
e −θ P s ≈
1 + 0.5θ P s

K P (1 − 0.5θ P s )
so GP* ( s ) =
(τ P s + 1)(1 + 0.5θ P s )
Factor GP* ( s ) into invertible terms, GP* − ( s ) , and noninvertible terms, GP* + ( s ) . Recall that an invertible
term will not yield positive poles (positive roots of the denominator of the transfer function indicating
unstable behavior) when taken in the reciprocal:

GP* ( s ) = GP* + ( s )GP* − ( s )

so GP* + ( s ) = (1 − 0.5θ P s )

KP
and GP* − ( s ) =
(τ P s + 1)(1 + 0.5θ P s )

Now express the IMC controller, GC* ( s ) , in terms of GP* − ( s ) and a first-order filter term, F(s):

1
GC* ( s ) = F (s)
GP* − ( s )

272
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
The filter term can be expressed in terms of a closed-loop time constant, τ C :

1
F ( s) =
τC s +1

Substitute F(s) and GP* − ( s ) to express the IMC controller as:

⎛ (τ s + 1)(1 + 0.5θ P s ) ⎞ ⎛ 1 ⎞ (τ P s + 1)(1 + 0.5θ P s )

GC* ( s ) = ⎜⎜ P ⎟ τ s +1⎟ =
⎟⎜
⎝ KP ⎠⎝ C ⎠ K P (τ C s + 1)

Relate this IMC controller, GC* ( s) , to the classical feedback controller, GC ( s) :

GC* ( s )
GC ( s ) =
1 − GP* ( s )GC* ( s )

Substitute the above equations for GC* ( s ) and GP* ( s ) and simplify:

(τ P s + 1)(1 + 0.5θ P s )
K P (τ C s + 1)
GC ( s ) =
⎛ (τ P s + 1)(1 + 0.5θ P s ) ⎞ ⎛ K (1 − 0.5θ s ) ⎞
P P
1− ⎜ ⎟⎜ ⎟
⎜ K P (τ C s + 1) ⎟ ⎜ (τ P s + 1)(1 + 0.5θ P s ) ⎟
⎝ ⎠⎝ ⎠

(τ P s + 1)(1 + 0.5θ P s )
K P (τ C s + 1)
=
1−
(1 − 0.5θ P s )
(τ C s + 1)

=
(τ P s + 1)(1 + 0.5θ P s )
K P (τ C s + 1 − 1 + 0.5θ P s )

=
(τ P s + 1)(1 + 0.5θ P s )
K P (τ C s + 0.5θ P s )

=
(τ P s + 1)(1 + 0.5θ P s )
K P (τ C + 0.5θ P ) s

Factor this into a form we can compare with the interacting PID form:

273
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 τP ⎛ 1 ⎞
GC = ⎜1 + ⎟ ( 0.5θ P s + 1)
K P (τ C + 0.5θ P ) ⎝ τ P s ⎠

Compare this to the classical feedback form for an interacting PID controller,

⎛ 1 ⎞
GC ( s ) PID Interact = KC ⎜ 1 + ⎟ (τ D s + 1)
⎝ τIs ⎠

we obtain the following controller tuning correlations:

τP
KC = , τ I = τ P , and τ D = 0.5θ P
K P (τ C + 0.5θ P )

A.1.c Ideal PID with Filter Control

Start with the general FOPDT process model:

K P e −θ P s
GP* ( s ) =
τ Ps +1

Factor GP* ( s ) into invertible terms, GP* − ( s ) , and noninvertible terms, GP* + ( s ) . Recall that an invertible
term will not yield positive poles (positive roots of the denominator of the transfer function indicating
unstable behavior) when taken in the reciprocal:

GP* ( s ) = GP* + ( s)GP* − ( s )

so GP* + ( s) = e−θ P s

KP
and GP* − ( s ) =
τ Ps +1

Express the IMC controller model, GC* ( s ) , in terms of GP* − ( s) and a first-order filter term, F(s):

1
GC* ( s ) = F (s)
GP* − ( s )

The filter term can be expressed in terms of a closed-loop time constant, τ C :

1
F ( s) =
τC s +1

Substitute F(s) and GP* − ( s ) to express the IMC controller as:

274
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 ⎛ τ s +1⎞⎛ 1 ⎞ τ Ps +1
GC* ( s ) = ⎜ P ⎟⎜ ⎟=
⎝ K P ⎠⎝ τ C s + 1 ⎠ K P (τ C s + 1)

Relate this IMC controller, GC* ( s) , to a classical feedback controller, GC ( s ) :

GC* ( s )
GC ( s ) =
1 − GP* ( s )GC* ( s )

Substitute the above equations for GC* ( s ) and GP* ( s ) and simplify:

(τ P s + 1)
K P (τ C s + 1)
GC ( s ) =
⎛ K e−θ P s ⎞⎛ (τ P s + 1) ⎞
1− ⎜ P
⎜ (τ s + 1) ⎟⎜
⎟ ⎜ K (τ s + 1) ⎟⎟
⎝ P ⎠⎝ P C ⎠

(τ P s + 1)
K P (τ C s + 1)
=
e−θ P s
1−
(τ C s + 1)
τ Ps +1
=
(
K P τ C s + 1 − e −θ P s )
Substitute the 1/1 Padé approximation for e−θ P s :

1 − 0.5θ P s
e−θ P s ≈
1 + 0.5θ P s

τ Ps +1
so GC ( s ) =
⎛ 1 − 0.5θ P s ⎞
K P ⎜τ C s + 1 − ⎟
⎝ 1 + 0.5θ P s ⎠

(τ P s + 1)(1 + 0.5θ P s )
=
(
K P τ C s + 0.5θ Pτ C s 2 + 1 + 0.5θ P s − 1 + 0.5θ P s )
(τ P s + 1)(1 + 0.5θ P s )
=
K P (τ C s + 0.5θ Pτ C s 2 + (τ C + θ P ) s )

Factor into a form we can compare to the ideal PID with filter form:

275
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 GC ( s ) =
(1 + 0.5θ P s + τ P s + 0.5θ Pτ P s
2
)
⎛ τ Cθ P ⎞
K P (τ C + θ P ) s ⎜ s + 1⎟
⎜ 2 (τ + θ ) ⎟
⎝ C P ⎠

⎡(τ P + 0.5θ P ) s + 1 + 0.5θ Pτ P s 2 ⎤

= ⎣ ⎦
⎛ τ Cθ P ⎞
K P (τ C + θ P ) s ⎜ s + 1⎟
⎜ 2 (τ + θ ) ⎟
⎝ C P ⎠

GC ( s ) =
(τ P + 0.5θ P ) ⎛1 + 1
+
τ Pθ P ⎞ 1
⎜⎜ s⎟
K P (τ C + θ P ) ⎝ (τ P + 0.5θ P ) s 2 (τ P + 0.5θ P ) ⎟⎠ ⎛ τ Cθ P ⎞
⎜⎜ s + 1⎟
⎝ 2 (τ C + θ P )
⎟
⎠

Compare this to the classical feedback form for an PID with filter controller:

⎛ 1 ⎞⎛ 1 ⎞
GC ( s ) PID Ideal, Filter = KC ⎜ 1 + + τ D s ⎟⎜ ⎟
⎝ τIs ⎠⎝ ατ D s + 1 ⎠

we obtain the following controller tuning correlations:

KC =
(τ P + 0.5θ P ) , τ I = τ P + 0.5θ P , τ D =
τ Pθ P
, and α =
τ C (τ P + 0.5θ P )
K P (τ C + θ P ) 2 (τ P + 0.5θ P ) τ P (τ C + θ P )

A.1.d Interacting PID with Filter Control

Start with the general FOPDT process model:

K P e −θ P s
GP* ( s ) =
τ Ps +1

Factor GP* ( s ) into invertible terms, GP* − ( s ) , and noninvertible terms, GP* + ( s ) . Recall that an invertible
term will not yield positive poles (positive roots of the denominator of the transfer function indicating
unstable behavior) when taken in the reciprocal:

GP* ( s ) = GP* + ( s)GP* − ( s )

so GP* + ( s ) = e −θ P s

KP
and GP* − ( s ) =
τ Ps +1

276
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
Now express the IMC controller, GC* ( s ) , in terms of GP* − ( s ) and a first-order filter term, F(s):

1
GC* ( s ) = F (s)
GP* − ( s )

The filter term can be expressed in terms of a closed-loop time constant, τ C :

1
F ( s) =
τC s +1

Substitute F(s) and GP* − ( s ) to express the controller model as:

⎛ τ s +1⎞⎛ 1 ⎞ τ Ps +1
GC* ( s ) = ⎜ P ⎟⎜ ⎟=
⎝ K P ⎠⎝ τ C s + 1 ⎠ K P (τ C s + 1)

Relate this IMC controller, GC* ( s ) , to the classical feedback controller, GC ( s ) :

GC* ( s )
GC ( s ) =
1 − GP* ( s )GC* ( s )

Substitute equations for GC* ( s ) and GP* ( s ) and simplify:

(τ P s + 1)
K P (τ C s + 1)
GC ( s ) =
⎛ K e −θ P s ⎞⎛ (τ P s + 1) ⎞
1− ⎜ P ⎟⎜ ⎟⎟
⎝ (τ P s + 1) ⎠ ⎝ K P (τ C s + 1) ⎠
⎜ ⎟⎜

(τ P s + 1)
K P (τ C s + 1)
=
e −θ P s
1−
(τ C s + 1)
τ Ps +1
=
(
K P τ C s + 1 − e −θ P s )
Substitute the 1/1 Padé approximation for e−θ P s :

1 − 0.5θ P s
e −θ P s ≈
1 + 0.5θ P s

277
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 τ Ps +1
so GC ( s ) =
⎛ 1 − 0.5θ P s ⎞
K P ⎜τ C s + 1 − ⎟
⎝ 1 + 0.5θ P s ⎠

(τ P s + 1)(1 + 0.5θ P s )
=
(
K P τ C s + 0.5θ Pτ C s 2 + 1 + 0.5θ P s − 1 + 0.5θ P s )
(τ P s + 1)(1 + 0.5θ P s )
=
K P (τ C s + 0.5θ Pτ C s 2 + (τ C + θ P ) s )

Now factor this in such a way that we can compare it to the interacting PID with filter form:

GC ( s ) =
(τ P s + 1)(1 + 0.5θ P s )
⎛ τ Cθ P ⎞
K P (τ C + θ P ) s ⎜ s + 1⎟
⎜ 2 (τ + θ ) ⎟
⎝ C P ⎠

⎛ 1 ⎞
τ P ⎜1 + ⎟ (1 + 0.5θ P s )
⎝ τ Ps ⎠
=
⎛ τ Cθ P ⎞
K P (τ C + θ P ) ⎜ s + 1⎟
⎜ 2 (τ + θ ) ⎟
⎝ C P ⎠

⎛ ⎞
⎛ ⎞ ⎜ ⎟
τP 1 ⎜ 0.5θ P s + 1 ⎟
= ⎜ 1 + ⎟
K P (τ C + θ P ) ⎝ τ P s ⎠ ⎜ τ Cθ P ⎟
⎜ 2 (τ + θ ) s + 1 ⎟
⎝ C P ⎠

Comparing this with the classical feedback model for interacting PID with filter control,

⎛ 1 ⎞⎛ τ D s + 1 ⎞
GC ( s ) PID Interact, Filter = KC ⎜ 1 + ⎟⎜ ⎟
⎝ τ I s ⎠⎝ ατ D s + 1 ⎠

we obtain the following controller tuning parameters:

τP τC
KC = , τ I = τ P , τ D = 0.5θ P , and α =
K P (τ C + θ P ) (τ C + θ P )

Note: The derivation of PI tuning correlations for self-regulating processes can be found in Chap 17.

278
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
A.2 Non-Self Regulating Processes

A.2.a Ideal PID Control

Start with the general FOPDT process model for integrating processes:

K P* e −θ P s
GP* ( s ) =
s

Substitute the 1/1 Padé approximation for e−θ P s :

1 − 0.5θ P s
e −θ P s ≈
1 + 0.5θ P s

K P* (1 − 0.5θ P s )
so GP* ( s ) =
s (1 + 0.5θ P s )

Factor GP* ( s ) into invertible terms, GP* − ( s ) , and noninvertible terms, GP* + ( s ) . Recall that an invertible
term will not yield positive poles (positive roots of the denominator of the transfer function indicating
unstable behavior) when taken in the reciprocal:

GP* ( s ) = GP* + ( s)GP* − ( s )

so GP* + ( s ) = 1 − 0.5θ P s

K P*
and GP* − ( s ) =
s (1 + 0.5θ P )

Now express the IMC controller, GC* ( s ) , in terms of the invertible terms, GP* − ( s ) , and a first-order
filter term, F(s):

1
GC* ( s ) = F (s)
GP* − ( s )

The filter term can be expressed in terms of a closed-loop time constant, τ C , and the noninvertible
process model, GP* + ( s ) , as:

F (s) =
( 2τ C − GP* + ' (0) s + 1)
(τ C s + 1) 2

and since GP* + ' (0) =

( ) (0) = −0.5θ
d GP* +
P
ds

279
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
then F ( s) =
( 2τ C + 0.5θ P ) s + 1
(τ C s + 1)2
Substitute F(s) and GP* − ( s ) to express the IMC controller as:

s [1 + 0.5θ P s ] ⎡⎣1 + s ( 0.5θ P + 2τ C ) ⎤⎦

GC* ( s ) =
K P* (τ C s + 1)
2

Relate this IMC controller, GC* ( s ) , to a classical feedback controller, GC ( s ) :

GC* ( s )
GC ( s ) =
1 − GP* ( s )GC* ( s )

Substitute the above equations for GC* ( s ) and GP* ( s ) :

s [1 + 0.5θ P s ] ⎡⎣1 + s ( 0.5θ P + 2τ C ) ⎤⎦

GC ( s ) =
K P* ⎡⎢(τ C s + 1) − (1 − 0.5θ P s ) (1 + s ( 0.5θ P + 2τ C ) ) ⎤⎥
2
⎣ ⎦

Factor this into a form we can compare with the Ideal PID controller form:

s [1 + 0.5θ P s ] ⎡⎣1 + s ( 0.5θ P + 2τ C ) ⎤⎦

GC ( s ) =
( )
K P* ⎡1 + 2τ C s + τ C 2 s 2 − 1 + 0.5θ P s + 2τ C s − 0.5θ P s − 0.25θ P2 s 2 − θ Pτ C s 2 ⎤
⎣ ⎦

=
(
s 0.5θ P s + 1 + 0.25θ P2 s 2 + 0.5θ P s + τ Cθ P s 2 + 2τ C s )
K P* ⎡⎢ s 2 (τ C + 0.5θ P ) ⎥
2⎤
⎣ ⎦

(
s ⎡1 + ( 2τ C + θ P ) s + τ Cθ P + 0.25θ P2 s 2 ⎤
= ⎣ ⎦ )
* ⎡ 2 2⎤
K P ⎢ s (τ C + 0.5θ P ) ⎥
⎣ ⎦

1 ⎡
= * ⎢
1
⎥⎢ + C P
+
(
⎤ ⎡ 1 ( 2τ + θ ) s τ Cθ P + 0.25θ P2 s 2 ⎤
⎥ )
2 ⎢s ⎥
K P ⎢ (τ C + 0.5θ P ) ⎥ s s
⎣ ⎦ ⎢⎣ ⎥⎦

1 ⎡ 2τ C + θ P ⎤ ⎢
= * ⎢ ⎥
⎡
1+
1
+
(
τ Cθ P + 0.25θ P2 ⎤
s⎥
)
K P ⎢ (τ C + 0.5θ P )2 ⎥ ⎢ ( 2τ C + θ P ) s ( 2τ C + θ P ) ⎥
⎣ ⎦ ⎢⎣ ⎥⎦

Compare this to the classical feedback form for an ideal PID controller,

280
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 ⎛ 1 ⎞
GC ( s ) PID Ideal = KC ⎜1 + +τ Ds ⎟
⎝ τIs ⎠
we obtain the following controller tuning correlations:

KC =
1 2τ C + θ P
, τ I = 2τ C + θ P , and τ D =
(τ Cθ P + 0.25θ P2 )
K P* (τ C + 0.5θ P ) 2
( 2τ C + θ P )

A.2.b Interacting PID Control

Start with the general FOPDT process model for integrating processes:

K P* e −θ P s
GP* ( s ) =
s

Substitute the 1/1 Padé approximation for e−θ P s :

1 − 0.5θ P s
e −θ P s ≈
1 + 0.5θ P s

K P* (1 − 0.5θ P s )
so GP* ( s ) =
s (1 + 0.5θ P s )

Factor GP* ( s ) into invertible terms, GP* − ( s ) , and noninvertible terms, GP* + ( s ) . Recall that an invertible
term will not yield positive poles (positive roots of the denominator of the transfer function indicating
unstable behavior) when taken in the reciprocal:

GP* ( s ) = GP* + ( s )GP* − ( s )

so GP* + ( s ) = 1 − 0.5θ P s

K P*
and GP* − ( s ) =
s (1 + 0.5θ P )

Now express the IMC controller, GC* ( s ) , in terms of the invertible terms, GP* − ( s ) , and a first-order
filter term, F(s):

1
GC* ( s ) = F (s)
GP* − ( s )

The filter term can be expressed in terms of a closed-loop time constant, τ C , and the noninvertible
process model, GP* + ( s ) , as:

281
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 F (s) =
( 2τ C )
− GP* + ' (0) s + 1

(τ C s + 1) 2

and since GP* + ' (0) =

( ) ( 0) = −0.5θ
d GP* +
P
ds

then F ( s) =
( 2τ C + 0.5θ P ) s + 1
(τ C s + 1)2
Substitute F(s) and GP* − ( s ) to express the IMC controller as:

s [1 + 0.5θ P s ] ⎡⎣1 + s ( 0.5θ P + 2τ C ) ⎤⎦

GC* ( s ) =
K P* (τ C s + 1)
2

Relate this IMC controller, GC* ( s ) , to a classical feedback controller, GC ( s ) :

GC* ( s )
GC ( s ) =
1 − GP* ( s )GC* ( s )

Substitute the above equations for GC* ( s ) and GP* ( s ) :

s [1 + 0.5θ P s ] ⎡⎣1 + s ( 0.5θ P + 2τ C ) ⎤⎦

GC ( s ) =
K P* ⎡⎢(τ C s + 1) + (1 − 0.5θ P s ) (1 + s ( 0.5θ P + 2τ C ) ) ⎤⎥
2
⎣ ⎦

Factor this into a form we can compare with the Interacting PID controller form:

s ⎡⎣1 + s ( 0.5θ P + 2τ C ) ⎤⎦
GC ( s ) = [1 + 0.5θ P s ]
K P* ⎡1 + 2τ C s + s 2τ C 2 − (1 − 0.5θ P s ) (1 + 0.5θ P s + 2τ C s ) ⎤
⎣ ⎦

s ⎡⎣1 + s ( 0.5θ P + 2τ C ) ⎤⎦
= [1 + 0.5θ P s ]
(
K P* ⎡1 + 2τ C s + s 2τ C 2 − 1 + 2τ C s + 0.25θ P2 s 2 + θ Pτ C s 2 ⎤
⎣ ⎦)
s ⎡⎣1 + s ( 0.5θ P + 2τ C ) ⎤⎦
= [1 + 0.5θ P s ]
K P* ⎡ s 2τ C 2 + 0.25θ P2 s 2 + θ Pτ C s 2 ⎤
⎣ ⎦

s ⎡⎣1 + s ( 0.5θ P + 2τ C ) ⎤⎦
= [1 + 0.5θ P s ]
K P* s 2 ⎡τ C 2 + 0.25θ P2 + θ Pτ C ⎤
⎣ ⎦

282
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 1 + s ( 0.5θ P + 2τ C )
= [1 + 0.5θ P s ]
K P* s ⎡τ C 2 + 0.25θ P2 + θ Pτ C ⎤
⎣ ⎦

1 ⎡ ( 0.5θ P + 2τ C ) 1 ⎤
= ⎢ + ⎥ [1 + 0.5θ P s ]
K P* ⎢ (τ C + 0.5θ P )2 (τ C + 0.5θ P )2 s ⎥
⎣ ⎦

1 ( 0.5θ P + 2τ C ) ⎡1 + 1 ⎤
= ⎢ ⎥ [1 + 0.5θ P s ]
K P* (τ C + 0.5θ P )2 ⎢⎣ ( 2τ C + 0.5θ P ) s ⎥⎦

Compare this to the classical feedback form for an interacting PID controller,

⎛ 1 ⎞
GC ( s ) PID Interact = KC ⎜ 1 + ⎟ (τ D s + 1)
⎝ τIs ⎠

we obtain the following controller tuning correlations:

KC =
1 ( 2τ C + 0.5θ P ) , τ I = 2τ C + 0.5θ P , and τ D = 0.5θ P
K P* (τ C + 0.5θ P )2

A.2.c Ideal PID with Filter Control

Start with the general FOPDT process model for integrating processes:

K P* e−θ P s
GP* ( s ) =
s

Factor GP* ( s ) into invertible terms, GP* − ( s ) , and noninvertible terms, GP* + ( s) . Recall that an invertible
term will not yield positive poles (positive roots of the denominator of the transfer function indicating
unstable behavior) when taken in the reciprocal:

GP* ( s ) = GP* + ( s )GP* − ( s )

so GP* + ( s) = e−θ P s

K P*
and GP* − ( s ) =
s

Now express the IMC controller, GC* ( s ) , in terms of the invertible terms, GP* − ( s ) , and a first-order
filter term, F(s):
1
GC* ( s ) = * F (s)
GP − ( s )

283
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
The filter term can be expressed in terms of a closed-loop time constant, τ C , and the noninvertible
process model, GP* + ( s ) , as:

F (s) =
( 2τ C − GP* + ' (0) s + 1)
(τ C s + 1) 2

and since GP* + ' (0) =

( ) (0) = −θ
d GP* +
P
ds

then F (s) =
( 2τ C + θ P ) s + 1
(τ C s + 1)2
Substitute F(s) and GP* − ( s ) to express the IMC controller as:

s ⎡1 + s ( 2τ C + θ P ) ⎤⎦
GC* ( s ) = ⎣
K P* (τ C s + 1)
2

Relate this IMC controller, GC* ( s ) , to a classical feedback controller model, GC ( s ) :

GC* ( s )
GC ( s ) =
1 − GP* ( s )GC* ( s )

Substitute the above equations for GC* ( s ) and GP* ( s ) :

s ⎡⎣1 + s ( 2τ C + θ P ) ⎤⎦
GC ( s ) =
⎡ ⎛ s ⎡1 + s ( 2τ + θ ) ⎤ ⎞ ⎛ K * e −θ P s ⎞⎤
K P* (τ C s + 1) ⎢⎢1 − ⎜⎜ ⎣ * C 2P ⎦ ⎟⎟ ⎜⎜ P
2 ⎟⎥
⎟⎥
K P (τ C s + 1) s
⎣ ⎝ ⎠⎝ ⎠⎦

s ⎡⎣1 + s ( 2τ C + θ P ) ⎤⎦
GC ( s ) =
K P* ⎡⎢(τ C s + 1) − e −θ P s (1 + s ( 2τ C + θ P ) ) ⎤⎥
2
⎣ ⎦

Substitute the 1/1 Padé approximation for e−θ P s :

1 − 0.5θ P s
e −θ P s ≈
1 + 0.5θ P s

s ⎡⎣1 + s ( 2τ C + θ P ) ⎤⎦
so GC ( s ) =
⎡ 2 ⎛ 1 + 0.5θ P s ⎞ ⎤
K P* ⎢(τ C s + 1) − ⎜ ⎟ (1 + s ( 2τ C + θ P ) ) ⎥
⎢⎣ ⎝ 1 − 0.5θ P s ⎠ ⎥⎦

284
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
Factor this into a form we can compare with the ideal PID controller with filter controller form:

s ⎡⎣1 + s ( 2τ C + θ P ) ⎤⎦ [1 + 0.5θ P s ]
GC ( s ) =
K P* ⎡⎢(τ C s + 1) (1 + 0.5θ P s ) − (1 + s ( 2τ C + θ P ) ) (1 − 0.5θ P s ) ⎤⎥
2
⎣ ⎦

2τ C s 2 + θ P s 2 + s + τ Cθ P s 3 + 0.5θ P2 s 3 + 0.5θ P s 2
=
K P* ⎡τ C2 s 2 + 0.5θ Pτ C2 s 3 + 2τ Cθ P s 2 + 0.5θ P2 s 2 ⎤
⎣ ⎦

⎡ 1 ⎤
s 2 ⎢ 2τ C + θ P + + τ Cθ P s + 0.5θ P2 s + 0.5θ P ⎥
⎣ s ⎦
=
* 2 ⎡ 2 2 2⎤
K P s τ C + 0.5θ Pτ C s + 2τ Cθ P + 0.5θ P
⎣ ⎦

⎡ 1 0.5θ P2 + τ Cθ P ⎤
[ 2τ C + 1.5θ P ] ⎢1 + + s⎥
⎣⎢ ( 2τ C + 1.5θ P ) s ( 2τ C + 1.5θ P ) ⎦⎥
=
⎡ 0.5θ P2τ C ⎤
K P* ⎡τ C2 + 2τ Cθ P + 0.5θ P2 ⎤ ⎢ 2 s + 1⎥
⎣ ⎦ ⎢τ + 2τ θ + 0.5θ 2 ⎥⎦
⎣ C C P P

GC ( s ) =
( 2τ C + 1.5θ P ) ⎡ 1 0.5θ P2 + τ Cθ P ⎤ 1
⎢1 + + s⎥
(
K P* τ C2 + 2τ Cθ P + 0.5θ P2 ) ⎢⎣ ( 2τ C + 1.5θ P ) s ( 2τ C + 1.5θ P ) ⎥⎦ ⎡
⎢ 2
2
0.5θ Pτ C
2
⎤
s + 1⎥
⎣⎢τ C + 2τ Cθ P + 0.5θ P ⎦⎥

Compare this to the classical feedback model for an ideal PID controller with filter,

⎛ 1 ⎞⎛ 1 ⎞
GC ( s ) PID Ideal, Filter = KC ⎜ 1 + + τ D s ⎟⎜ ⎟
⎝ τIs ⎠⎝ ατ D s + 1 ⎠

we obtain the following controller tuning correlations:

( 2τ C + 1.5θ P ) 0.5θ P2 + τ Cθ P
KC = , τ I = 2τ C + 1.5θ P , τ D = ,
(
K P* τ C2 + 2τ Cθ P + 0.5θ P2 ) 2τ C + 1.5θ P

0.5τ C2 ( 2τ C + 1.5θ P )
α=
(τ 2
C )
+ 2τ Cθ P + 0.5θ P2 ( 0.5θ P + τ C )

A.2.d Interacting PID with Filter Control

Start with the general FOPDT process model for integrating processes:

K P* e −θ P s
GP* ( s ) =
s

285
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
Factor GP* ( s ) into invertible terms, GP* − ( s ) , and noninvertible terms, GP* + ( s) . Recall that an invertible
term will not yield positive poles (positive roots of the denominator of the transfer function indicating
unstable behavior) when taken in the reciprocal:

GP* ( s ) = GP* + ( s )GP* − ( s )

so GP* + ( s) = e−θ P s

K P*
and GP* − ( s ) =
s

Now express the IMC controller, GC* ( s ) , in terms of the invertible terms, GP* − ( s ) , and a first-order
filter term, F(s):
1
GC* ( s ) = * F (s)
GP − ( s )

The filter term can be expressed in terms of a closed-loop time constant, τ C , and the noninvertible
process model, GP* + ( s ) , as:

F (s) =
( 2τ C )
− GP* + ' (0) s + 1

(τ C s + 1) 2

and since GP* + ' (0) =

( ) (0) = −θ
d GP* +
P
ds

then F (s) =
( 2τ C + θ P ) s + 1
(τ C s + 1)2
Substitute F(s) and GP* − ( s ) to express the IMC controller as:

s ⎡1 + s ( 2τ C + θ P ) ⎤⎦
GC* ( s ) = ⎣
K P* (τ C s + 1)
2

Relate this IMC controller, GC* ( s ) , to a classical feedback controller, GC ( s ) :

GC* ( s )
GC ( s ) =
1 − GP* ( s )GC* ( s )

Substitute the above equations for GC* ( s) and GP* ( s ) :

286
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 s ⎡⎣1 + s ( 2τ C + θ P ) ⎤⎦
GC ( s ) =
K P* ⎡⎢(τ C s + 1) − e −θ P s (1 + s ( 2τ C + θ P ) ) ⎤⎥
2
⎣ ⎦

Substitute the 1/1 Padé approximation for e−θ P s :

1 − 0.5θ P s
e −θ P s ≈
1 + 0.5θ P s

s ⎡⎣1 + s ( 2τ C + θ P ) ⎤⎦
so GC ( s ) =
⎡ 2 ⎛ 1 + 0.5θ P s ⎞ ⎤
K P* ⎢(τ C s + 1) − ⎜ ⎟ (1 + s ( 2τ C + θ P ) ) ⎥
⎣⎢ ⎝ 1 − 0.5θ P s ⎠ ⎦⎥

Factor this into a form we can compare with the interacting PID controller with filter controller form:

s ⎡⎣1 + s ( 2τ C + θ P ) ⎤⎦ [1 + 0.5θ P s ]
GC ( s ) =
K P* ⎡⎢(τ C s + 1) (1 + 0.5θ P s ) − (1 + s ( 2τ C + θ P ) ) (1 − 0.5θ P s ) ⎤⎥
2
⎣ ⎦

s ⎡⎣1 + s ( 2τ C + θ P ) ⎤⎦ [1 + 0.5θ P s ]
=
K P* ⎡ s 2τ C2 + 2θ Pτ C s 2 + 0.5θ P s 3τ C2 + 0.5θ P2 s 2 ⎤
⎣ ⎦

⎡⎣1 + s ( 2τ C + θ P ) ⎤⎦ [1 + 0.5θ P s ]
=
K P* ⎡ sτ C2 + 2θ Pτ C s + 0.5θ P s 2τ C2 + 0.5θ P2 s ⎤
⎣ ⎦

⎡⎣1 + s ( 2τ C + θ P ) ⎤⎦ [1 + 0.5θ P s ]
=
(
K P* ⎡ 0.5θ P s 2τ C2 + s τ C2 + 2θ Pτ C + 0.5θ P2 ⎤
⎣ ⎦ )
⎡ ⎤
⎢ ⎥
⎤⎢ ⎥
1 ⎢ (1 + s ( 2τ C + θ P ) )
⎡
= * ⎥⎢ (1 + 0.5θ P s ) ⎥

⎣⎢ (
K P ⎢ s τ C2 + 2θ Pτ C + 0.5θ P2 ) ⎥ ⎢⎛
⎦⎥ ⎢ ⎜ 0.5θ Pτ C2
⎞ ⎥
⎥
s ⎟ + 1⎥
⎢⎜ 2
( )2 ⎟
⎢⎣ ⎜⎝ τ C + 2θ Pτ C + 0.5θ P ⎟⎠ ⎥⎦

⎡ ⎤
⎡ ⎤⎡ ⎢ ⎥
GC ( s ) = ⎢
( 2τ C + θ P ) ⎥ ⎢1 + 1 ⎤⎢ 0.5θ P s + 1 ⎥
⎥⎢
(
⎢ K * τ 2 + 2θ τ + 0.5θ 2
⎣⎢ P C P C P ) ⎥ ⎢ ( 2τ C + θ P ) s ⎥ ⎛
⎦⎥ ⎣ ⎦⎢
⎜⎜ 2
0.5θ Pτ C2 ⎞ ⎥
⎟⎟ s + 1 ⎥
⎢ τ + 2θ τ + 0.5θ 2 ⎥
⎣⎝ C P C P ⎠ ⎦

Compare this to the classical feedback process model for an interacting PID controller with filter,

287
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 ⎛ 1 ⎞⎛ τ D s + 1 ⎞
GC ( s ) PID Interact, Filter = KC ⎜ 1 + ⎟⎜ ⎟
⎝ τ I s ⎠⎝ ατ D s + 1 ⎠

we obtain the following controller tuning correlations:

2τ C + θ P τ C2
KC = , τ I = 2τ C + θ P , τ D = 0.5θ P , and α =
(
K P* τ C2 + 2τ Cθ P + 0.5θ P2 ) τ C2 + 2τ Cθ P + 0.5θ P2

288
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
Appendix B: Table of Laplace Transforms

Time Domain Laplace Domain

f (t ) F (s )

a f (t ) a F (s )

δ(t) (impulse) 1

a
a (step)
s
a
at (ramp) s2
n!
t n
s +1
n

ω
sin(ω t ) s +ω2
2

s
cos(ω t ) s +ω2
2

1
at
e− s+a

t n −1e − at 1
( n − 1)! ( s + a)n

ω
sinh(ω t ) s −ω2
2

s
cosh(ω t ) s −ω2
2

d f (t )
dt sF ( s ) − f (0)

d 2 f (t ) df(t)
s 2 F ( s ) − s f (0) −
dt 2 dt t = 0
t 1
∫ 0 f (t ) dt s
F (s)

f (t − θ ) e −θ s F (s)

289
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Appendix C: DMC Controller Tuning Guides
C.1 DMC Tuning Guide for Self Regulating (Stable) Processes
Fit a first order plus dead time (FOPDT) dynamic model to process data. “Process” is defined to include all dynamic
information from the output signal of the controller through the measured response signal of the process variable.

Generate process data by forcing the measured process variable with a change in the controller output signal. For
accurate results:
- the process must be at steady state before forcing a dynamic response; the first data point in the file must equal that steady
state value
- the data collection sample rate should be ten times per time constant or faster (T ≤ 0.1 τ P )
- the controller output should force the measured process variable to move at least ten times the noise band
Use Design Tools to fit a FOPDT dynamic model to the process data set. A FOPDT model has the form:
dy (t ) Y (s ) K P e −θ P s
Time Domain: τ P + y (t ) = KP u(t − θ P ) Laplace Domain: =
dt U (s ) τP s + 1
where: y(t) = measured process variable signal also: T = DMC sample time; units of time
u(t) = controller output signal N = DMC model horizon; samples
KP = process gain; units of y(t)/u(t) P = DMC prediction horizon; samples
τP = process time constant; units of time M = DMC control horizon; samples
θ = process dead time; units of time λ = DMC move suppression coefficient; unitless
P

These correlations provide a starting point for tuning. Final tuning may require online trial and error. “Best” tuning is
defined by you and your knowledge of the capabilities of the process, desires of management, goals of production, and
impact on other processes.

DMC Tuning

1. Approximate the manipulated-to-measured-process-variable dynamics with a first order plus dead time (FOPDT)
model shown above.
2. Select the sample time as close as possible to:
T = 0.1 τ p or T = 0.5 θ p whichever is larger

3. Compute the prediction horizon, P, and the model horizon, N, as the process settling time in samples (rounded to
the next integer):
⎛ 5τ p ⎞ ⎛θp ⎞
P = N = Int ⎜ ⎟ + k where k =Int ⎜ ⎟ + 1
⎝ T ⎠ ⎝T ⎠
4. Select the control horizon, M, as the time in samples required for the open loop response to reach 60% of the
steady state:
⎛τ p ⎞
M = Int ⎜ ⎟ +k
⎝T ⎠
5. Compute the move suppression coefficient:
⎧0 M =1
⎪
λ = ⎨ M ⎛ 3.5 τ p ( M − 1) ⎞ 2
⎪ 10 ⎜ T + 2 − 2 ⎟ K p M >1
⎩ ⎝ ⎠
6. Implement DMC using the traditional step response matrix of the actual process and the above parameters.

290
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
C.2 DMC Tuning Guide for Integrating (Non-Self Regulating) Processes
Fitt a FOPDT Integrating dynamic model to process data. “Process” is defined to include all dynamic information from
the output signal of the controller through the measured response signal of the process variable.
Integrating processes are unstable so the process should already be in closed loop. If not, stabilize the process with a
simple P-Only controller and generate process data with a set point step. For accurate results:
- the process must be at steady state before forcing a dynamic response; the first data point recorded must equal that steady
state value
- the controller output move from the set point step should force the process variable to move at least ten times the noise
band
Use Design Tools to fit a FOPDT Integrating dynamic model to the process data set. A FOPDT Integrating model has the form:
dy (t ) Y ( s ) K P* e − θ P s
Time Domain: = K P* u(t − θ P ) Laplace Domain: =
dt U (s) s
where: y(t) = measured process variable signal also: T = DMC sample time; units of time
u(t) = controller output signal N = DMC model horizon; samples
K*P = integral gain; units of y(t)/(u(t)·time) P = DMC prediction horizon; samples
θ = process dead time; units of time M = DMC control horizon; samples
P
λ = DMC move suppression coefficient; unitless

These correlations provide a starting point for tuning. Final tuning may require online trial and error. “Best” tuning is
defined by you and your knowledge of the capabilities of the process, desires of management, goals of production, and
impact on other processes.

DMC Tuning

1. Approximate the manipulated-to-measured-process-variable dynamics with a first order plus dead time
integrating (FOPDT Integrating) model shown above.
2. Select the sample time as close as possible to:
T ≤ 0.5θ p

3. The closed-loop time constant for a FOPDT integrating model can be approximated as:
τ CL = θ p 10
4. Compute the prediction horizon, P, and the model horizon, N, as the process settling time in samples (rounded
to the next integer):
⎛ 5τ ⎞ ⎛θp ⎞
P = N = Int ⎜ CL ⎟ + k where k = Int ⎜ ⎟ + 1
⎝ T ⎠ ⎝T ⎠
5. Select the control horizon, M, as the time in samples required for the open loop response to reach 60% of the
steady state:
⎛τ ⎞
M = Int ⎜ CL ⎟+k
⎝ T ⎠
6. Compute the move suppression coefficient:
⎧0 M=1
⎪
λ=⎨ 1 ⎛ M 2 ( P − k + 1)3 ⎞
⎪ ⎜⎜ − 0.08M 3 ( P − k + 1)2 ⎟ ⎡ K *pT ⎤ 2 M>1
⎟ ⎣ ⎦
⎩10 M ⎝ 3 ⎠
7. Implement DMC using the traditional step response matrix of the actual process and the above parameters.

291
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Appendix D: PID Controller Tuning Guides
D.1 PID Tuning Guide for Self Regulating (Stable) Processes
Begin by fitting a first order plus dead time (FOPDT) dynamic model to process data. “Process” is defined to include all dynamic
information from the output signal of the controller through the measured response signal of the process variable.
Generate process data by forcing the measured process variable with a change in the controller output signal. For accurate results:
- the process must be at steady state before forcing a dynamic response; the first data point recorded must equal that steady state value
- the data collection sample rate should be ten times per time constant or faster (T ≤ 0.1 τ P )
- the controller output should force the measured process variable to move at least ten times the noise band
Use Design Tools to fit a FOPDT dynamic model to the process data set. A FOPDT model has the form:
dy (t ) Y (s) K P e −θ P s
Time Domain: τ P + y (t ) = KP u(t − θ P ) Laplace Domain: =
dt U ( s) τP s + 1
where: y(t) = measured process variable signal also:
u(t) = controller output signal KC = controller gain; units of u(t)/y(t)
KP = process gain; units of y(t)/u(t) τ I = controller reset time; units of time
τP = process time constant; units of time τ D = controller derivative time; units of time
θP = process dead time; units of time α = derivative filter constant; unitless
Values of KP, τ P and θ P that describe the dynamic behavior of your process are important because:
- they are used in correlations (listed below) to compute initial PID controller tuning values KC , τ I , τ D and α
- the sign of KP indicates the action of the controller (+KP → reverse acting; − KP → direct acting)
- the size of τ P indicates the maximum desirable loop sample time (be sure sample time T ≤ 0.1 τP )
- the ratio θ P / τ P indicates whether a Smith predictor would show benefit (useful when θ P ≥ τ P )
- the model itself is used in feed forward, Smith predictor, decoupling and other model-based controllers
These correlations provide an excellent start for tuning. Final tuning may require online trial and error. “Best” tuning is defined by you
and your knowledge of the capabilities of the process, desires of management, goals of production, and impact on other processes.
IMC (lambda) Tuning
* This is an ITAE
Aggressive Tuning: τ C is the larger of 0.1 τ P or 0.8 θ P correlation as no
Moderate Tuning: τ C is the larger of 1.0 τ P or 8.0 θ P P-Only IMC exists
Conservative Tuning: τ C is the larger of 10 τ P or 80 θ P

KC τI τD α

P-Only* 0.2
KC = (τ p / θ p ) 1.22
Kp
1 τP
PI τP
K P (θ P + τ C )

1 ⎛ τ P + 0.5 θ P ⎞ τP θP
PID Ideal ⎜ ⎟ τ P + 0.5 θ P
KP ⎜ τ + 0.5 θ ⎟ 2τ P + θ P
⎝ C P ⎠
1 ⎛ τP ⎞
PID Interacting ⎜ ⎟ τP 0.5 θ P
KP ⎜ τ + 0 .5 θ ⎟
⎝ C P ⎠
1 ⎛ τ P + 0.5 θ P ⎞ τP θP τ C (τ P + 0.5 θ P )
PID Ideal w/filter ⎜ ⎟ τ P + 0.5 θ P
KP ⎜ τ +θ
⎝ C P
⎟
⎠ 2τ P + θ P τ P (τ C + θ P )

1 ⎛ τP ⎞ τC
PID Interacting w/filter ⎜ ⎟ τP 0.5 θ P
KP ⎜τ +θ ⎟ τC + θ P
⎝ C P ⎠

292
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
D.2 PID Tuning Guide for Integrating (Non-Self Regulating) Processes
Begin by fitting a first order plus dead time integrating (FOPDT Integrating) dynamic model to process data. “Process” is defined to
include all dynamic information from the output signal of the controller through the measured response signal of the process variable.
Integrating processes are unstable so the process should already be in closed loop. If not, stabilize the process with a simple P-Only
controller and generate process data with a set point step. For accurate results:
- the process must be at steady state before forcing a dynamic response; the first data point recorded must equal that steady state value
- the controller output movement from the set point step should force the process variable to move at least ten times the noise band
Use Design Tools to fit a FOPDT Integrating dynamic model to the process data set. A FOPDT Integrating model has the form:
dy (t ) Y ( s ) K P* e − θ P s
Time Domain: = K P* u(t − θ P ) Laplace Domain: =
dt U (s) s
where: also:
y(t) = measured process variable signal KC = controller gain; units of (u(t)·time)/y(t)
u(t) = controller output signal τ I = controller reset time; units of time
KP* = integrator gain; units of y(t)/(u(t)·time) τ D = controller derivative time; units of time
θP = process dead time; units of time α = derivative filter constant; unitless
Values of K P* and θ P that describe the dynamic behavior of your process are important because:
- they are used in correlations (listed below) to compute initial PID controller tuning values KC , τ I , τ D and α
- the sign of KP* indicates the action of the controller (+KP* → reverse acting; − KP* → direct acting)

These correlations provide an excellent start for tuning. Final tuning may require online trial and error. “Best” tuning is defined by you
and your knowledge of the capabilities of the process, desires of management, goals of production, and impact on other processes.

IMC (lambda) Tuning * This is a Ziegler-Nichols

Standard Tuning: τ C = θ P 10 process reaction curve
(PRC) correlation as no
Conservative Tuning: τ C = 5θ P 10 P-Only IMC exists

KC τI τD α

1
P-Only* K *p θ P

1 2τ C + θ P
PI 2τ C + θ P
K *p (θ P + τ C ) 2

1 ⎛⎜ 2τ C + θ P ⎞
⎟ 0.25θ 2P + τ C θ P
PID Ideal 2τ C + θ P
K *p ⎜⎝ (τ C + 0.5 θ P ) 2 ⎟
⎠ 2τ C + θ P

1 ⎛⎜ 2τ C + 0.5 θ P ⎞
⎟
PID Interact 2τ C + 0.5θ P 0.5θ P
K *p ⎜⎝ (τ C + 0.5 θ P )2 ⎟
⎠

1 ⎛⎜ ⎞ 0.5θ2P + τC θP
2τC + 1.5θ p 2
⎟ (0.5τC )(2τC + 1.5θ P )
PID Ideal 2τ C + 1.5θ P
* ⎜ 2 2 ⎟ 2τC + 1.5θP 2
w/filter K P ⎝ τC + 2τC θ p + 0.5θ P ⎠ (τC + 2τC θ P + 0.5θ2P )(0.5θP + τC )

2
1 ⎛⎜ 2τC + θ P ⎞
⎟
τC
PID Interact 2τ C + θ P 0.5θ P
* ⎜ 2 2 ⎟ 2
w/filter K P ⎝ τC + 2τC θ P + 0.5θ P ⎠ τC + 2τC θ P + 0.5θ 2P

293
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 Index
cascade jacketed reactor, 19, peak overshoot ratio, 84
187 rise time, 84
A block diagram, 188 settling time, 84
anti-reset windup, 78 distillation column, 22, 209 improvement using Smith
apparent dead time, 25, 26, 69, decoupling control loops, predictors, 244
238 218 peek overshoot ratio, 238
definition, 30 interacting control loops, cruise control, 13, 39, 42
estimation from step test data, 214 Custom Process. See Control
31 furnace air/fuel ratio, 19 Station
automatic control, 8 gravity drained tanks, 15, 26,
examples, 12 28, 31, 41, 42, 44, 48, 67,
automatic mode, 11, 25 76, 84, 91, 96, 103 D
heat exchanger, 16, 27, 29, 32, decay ratio. See controller
43, 46, 49, 56, 59, 79 performance
B jacketed reactor, 18, 49, 184 decouplers, 22, 64, 210, 232
feed forward disturbance as feed forward elements, 218
balance
rejection, 202 construction of, 212
energy
vs. cascade jacketed reactor, for eliminating chatter, 222
well-stirred tank, 104
185 for improved set point tracking,
mass, 99
multi-tank process, 21 221
draining tank, 100
pumped tank, 17 implementation of, 211
non-interacting draining
characteristic equation. See derivative kick, 88
tanks in series, 103
ordinary differential equations derivative time. See tuning,
species (component)
(ODEs) controller
well-stirred tank, 106
closed loop time constant. See design level of operation, 24, 42,
block diagrams, 11, 155
internal model control (IMC) 43, 46
cascade architecture, 180
tuning correlations Design Tools. See Control Station
cascade jacketed reactor
complementary solution. See deviation variables, 113, 155
architecture, 188
ordinary differential equations direct action, 26, 44
closed loop, 158
(ODEs) distillation column. See Case
control loop
complex conjugates Studies
general, 12
poles as, 150 disturbance
distillation column process,
complex s plane, 133 definition, 10
210, 212
conserved variables, 99 disturbance rejection, 43, 46, 76,
feed forward controller with
control objective 79, 183, 187, 191, 194
feedback trim, 195
definition, 10 doublet test, 26, 53, 56
level-to-flow cascade
Control Station, 9, 24, 64 dynamic matrix control (DMC).
architecture, 184
Custom Process, 51, 64, 67, See model predictive control
multiplier block, 156
223, 224, 241 (MPC)
simplified, 161
Design Tools, 15, 43, 51, 52, dynamic process
Smith predictor architecture,
55, 56, 57, 58, 59, 60, 67, behavior
240
75, 76, 77, 90, 95, 186, 202, definition, 24
summer block, 156
215, 216, 219, 241, 243, 254 empirical modeling from data,
bumpless transfer, 48, 75
controller bias, 42, 75 99
definition, 39, 43 modeling for controller tuning
C controller design, 37, 55 and design, 24, 26
using closed loop data, 59 theoretical models derived
cascade control controller error, 42 from first principles, 99
controller design, 183, 187 definition, 10, 39
controller tuning, 184 controller gain. See tuning,
for improved disturbance controller F
rejection, 180, 183 controller output, 26, 39
primary loop control, 190 feed forward control, 26, 56, 64
definition, 10, 25
secondary loop control, 189 feed forward model
controller performance, 83
cascade jacketed reactor. See Case controller design, 198
criteria for evaluation, 83, 84
Studies feed forward element, 199
decay ratio, 84
Case Studies, 9, 15, 35, 36 feedback trim, 194

294
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 for improved disturbance derivations for non-self quadratic (QDMC), 251
rejection, 194, 196, 198, 203 regulating processes, 279 sample time, 250, 252, 253
model limits derivations for self single-input-single-output
dead time difference, 200 regulating processes, 270 (SISO) process, 253
highest model order, 200 internal model control (IMC) tuning single-loop controller
model order ratio, 201 correlations, 43, 56, 75, 90, 95 design, 248
process lead, 201 closed loop time constant, 75, tuning strategy, 262
static controller, 206 90, 95 multi-tank process. See Case
theory, 198 PI control, 76 Studies
feedback control loop, 37 PID control, 90 multivariable process control, 209,
final control element with filter, 95 223
definition, 10 standard tuning vs. cross-loop disturbance, 211
examples, 12 conservative tuning, 75, 90 decoupling cross-loop control
first order plus dead time inverse process effect of overall process
(FOPDT) dynamic model, 24, FOPDT fit, 63 time constant on, 235
25, 43, 56, 64, 65 effect of process dead time
controller output driven vs. on, 236
disturbance driven, 62 J effect of process gain on,
limitations, 32 jacketed reactor. See Case Studies 232
parameters. See steady state loop interaction, 224
process gain, overall process effect of overall process
time constant, apparent dead L time constant on, 228
time effect of process dead time
lambda tuning correlations, 56, 75
vs. SOPDT model, 66 on, 230
Laplace transform
vs. SOPDT w/L model, 71 effect of process gain on,
definition, 133
first order plus dead time with 224
properties, 135
integrator dynamic model, 56 multi input multi output
table, 289
flash drum process, 181, 196 (MIMO), 209, 223, 224
lead time, 70
furnace air/fuel ratio. See Case
definition, 70
Studies
limit of stability, 165 N
linearization, 110
G for functions of one variable, noise band, 54
111 nonlinear behavior of real
good engineering practice for for functions of two variables, processes, 33, 42
derivations, 99 112 non-self regulating (integrating)
gravity drained tanks. See Case loop sample time, 26 processes, 17
Studies

M O
H offset, 45, 74
manual mode, 11, 24
heat exchanger. See Case Studies measurement noise, 54, 89, 93 on/off control, 38, 39
hydrostatic head, 101, 102 measurement sensor ordinary differential equations
definition, 11 linear vs. nonlinear, 110
examples, 12 ordinary differential equations
I model predictive control (MPC), (ODEs)
integral of time-weighted absolute 239, 248 Laplace domain, 133, 144
error (ITAE) tuning correlations, dynamic matrix control moving from Laplace domain
43, 76 (DMC), 248 to time domain, 141
PI control control horizon, 250, 253 moving from time domain to
for disturbance rejection, 76 controller performance, 254, Laplace domain, 138
for set point tracking, 76 260 time domain, 117
P-Only control controller tuning, 253 second order
for disturbance rejection, 43 model horizon, 250, 253 characteristic equation,
for set point tracking, 43 move suppression 144
integrating factor, 117 coefficient, 253 second order underdamped
intermediate value control, 39, 41 prediction horizon, 249, form, 126
internal model control (IMC) 252, 253 solving first order, 117
tuning correlations process model, 251 solving second order, 121

295
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved
 characteristic equation, algorithms, 42 second order plus dead time with
121 controller design, 44 lead time (SOPDT w/L)
solving second-order process variable dynamic model, 65, 70
complementary solution, manipulated overdamped, 56
121 definition, 10 self regulating process
particular solution, 121 measured, 26, 42 definition, 65
overall process time constant, 25, definition, 10, 25 servo control. See set point
26, 67 primary, 181 tracking
63.2% of Process Step secondary, 181, 183 set point
Response rule, 119 profit motive, 8 definition, 10
definition, 28 proportional band, 48 set point tracking, 43, 44, 76, 91,
estimation from step test data, pseudo-random binary sequence 96, 192, 207
28 (PRBS) test, 26, 53 settling time. See controller
pulse test, 26, 52 performance
pumped tank. See Case Studies single loop control problems
P cascade solution, 183
Padé approximation, 161 feed forward solution, 197
particular solution. See ordinary
R Smith predictors, 26, 64, 239
differential equations (ODEs) regulatory control. See disturbance control algorithm, 239
peak overshoot ratio. See rejection steady state process gain, 25, 26,
controller performance relative gain, 224 42, 44, 67
peak time. See controller equal to 1, 227, 233 definition, 26
performance from 0.5-1, 227, 232 step test, 26, 52
perturbation variables. See from 0-0.5, 226, 232 sum of squared errors (SSE), 56
deviation variables greater than 1, 228, 233
PI control, 72 loop interaction, 225 T
algorithms, 72 much greater than 1, 228, 234
continuous form vs. discrete negative, 225, 232 time varying bevahior of real
form, 81 reset rate, 81 processes, 33
controller design, 75 reset time. See tuning, controller transfer functions
controller tuning map, 80 reset windup, 82 closed loop
oscillatory behavior, 75, 92 reverse action, 26, 44 process variable to
PID control rise time. See controller disturbance, 160
algorithms, 39, 86 performance process variable to set point,
derivative on measurement, root locus, 148, 162 160
88 roots of the characteristic combination, 155
derivative term, 40, 87, 89 equation, 121, 127, 144 controller, 146
ideal (non-interacting) form, poles (roots), 148
86, 90 process, 144
ideal (non-interacting) with S tuning, controller
derivative filter form, 95, safety, 8, 20, 21, 33 parameters, 11
96 second order plus dead time controller gain, 39, 42, 43,
integral term, 40, 87 (SOPDT) dynamic model, 65 44
interacting form, 86, 90 overdamped, 56 derivative time, 39, 86
interacting with derivative underdamped, 56 reset time, 39, 72
filter form, 95, 96 vs. FOPDT model, 66, 67, 69 tuning guide, 292
proportional term, 40, 87 second order plus dead time with
controller design, 90 lead time (SOPDT w/L)
with filter, 95 feed forward element, 200
U
derivative filter, 94 underdamped process
P-Only control, 41, 43 FOPDT fit, 62

296
Practical Process Control® by Douglas J. Cooper
Copyright © 2005 by Control Station, Inc.
All Rights Reserved