Control Engineering Series 90

Application of Dimensional Analysis in

Systems Modeling and Control Design

Application of Dimensional Analysis in
Systems Modeling and Control Design Application of
Dimensional
Dimensional analysis is an engineering tool that is widely applied to numerous engineering
problems, but has only recently been applied to control theory and problems such as
Pedro Balaguer is a lecturer
at the Department of Industrial
Analysis in Systems
Modeling and
identification and model reduction, robust control, adaptive control, and PID control. Systems Engineering and
Design, Universitat Jaume I de
Application of Dimensional Analysis in Systems Modeling and Control Design provides an Castelló, Spain. His research
introduction to the fundamentals of dimensional analysis for control engineers, and shows interests include the application
how they can exploit the benefits of the technique to theoretical and practical control of dimensional analysis to
problems. Topics covered include dimensional analysis and dimensional similarity, dynamical control problems, PID control,

Control Design
systems dimensionless representation, dimensionless systems identification and model order supervision of iterative and
reduction, homogeneity of PID tuning rules, dimensionless PID tuning rules comparison, adaptive control systems,
and energy cost and energy
dimensional analysis control fundamentals, control of dimensionally similar systems, and
consumption optimization.
adaptive control in the presence of input saturation.
Balaguer

Pedro Balaguer
The Institution of Engineering and Technology
www.theiet.org
978-1-84919-621-5

Application of Dimensional Analysis.indd 1 17/06/2013 14:04:01

 IET CONTROL ENGINEERING SERIES 90

Application of
Dimensional
Analysis in Systems
Modeling and
Control Design
Other volumes in this series:
Volume 8 A History of Control Engineering, 1800–1930 S. Bennett
Volume 18 Applied Control Theory, 2nd Edition J.R. Leigh
Volume 20 Design of Modern Control Systems D.J. Bell, P.A. Cook and N. Munro (Editors)
Volume 28 Robots and Automated Manufacture J. Billingsley (Editor)
Volume 33 Temperature Measurement and Control J.R. Leigh
Volume 34 Singular Perturbation Methodology in Control Systems D.S. Naidu
Volume 35 Implementation of Self-tuning Controllers K. Warwick (Editor)
Volume 37 Industrial Digital Control Systems, 2nd Edition K. Warwick and D. Rees (Editors)
Volume 39 Continuous Time Controller Design R. Balasubramanian
Volume 40 Deterministic Control of Uncertain Systems A.S.I. Zinober (Editor)
Volume 41 Computer Control of Real-time Processes S. Bennett and G.S. Virk (Editors)
Volume 42 Digital Signal Processing: principles, devices and applications N.B. Jones and J.D.McK. Watson
(Editors)
Volume 44 Knowledge-based Systems for Industrial Control J. McGhee, M.J. Grimble and A. Mowforth
(Editors)
Volume 47 A History of Control Engineering, 1930–1956 S. Bennett
Volume 49 Polynomial Methods in Optimal Control and Filtering K.J. Hunt (Editor)
Volume 50 Programming Industrial Control Systems Using IEC 1131-3 R.W. Lewis
Volume 51 Advanced Robotics and Intelligent Machines J.O. Gray and D.G. Caldwell (Editors)
Volume 52 Adaptive Prediction and Predictive Control P.P. Kanjilal
Volume 53 Neural Network Applications in Control G.W. Irwin, K. Warwick and K.J. Hunt (Editors)
Volume 54 Control Engineering Solutions: a practical approach P. Albertos, R. Strietzel and N. Mort (Editors)
Volume 55 Genetic Algorithms in Engineering Systems A.M.S. Zalzala and P.J. Fleming (Editors)
Volume 56 Symbolic Methods in Control System Analysis and Design N. Munro (Editor)
Volume 57 Flight Control Systems R.W. Pratt (Editor)
Volume 58 Power-plant Control and Instrumentation: the control of boilers and HRSG systems
D. Lindsley
Volume 59 Modelling Control Systems Using IEC 61499 R. Lewis
Volume 60 People in Control: human factors in control room design J. Noyes and M. Bransby (Editors)
Volume 61 Nonlinear Predictive Control: theory and practice B. Kouvaritakis and M. Cannon (Editors)
Volume 62 Active Sound and Vibration Control M.O. Tokhi and S.M. Veres
Volume 63 Stepping Motors, 4th Edition P.P. Acarnley
Volume 64 Control Theory, 2nd Edition J.R. Leigh
Volume 65 Modelling and Parameter Estimation of Dynamic Systems J.R. Raol, G. Girija and J. Singh
Volume 66 Variable Structure Systems: from principles to implementation A. Sabanovic, L. Fridman and
S. Spurgeon (Editors)
Volume 67 Motion Vision: design of compact motion sensing solution for autonomous systems
J. Kolodko and L. Vlacic
Volume 68 Flexible Robot Manipulators: modelling, simulation and control M.O. Tokhi and A.K.M. Azad
(Editors)
Volume 69 Advances in Unmanned Marine Vehicles G. Roberts and R. Sutton (Editors)
Volume 70 Intelligent Control Systems Using Computational Intelligence Techniques A. Ruano (Editor)
Volume 71 Advances in Cognitive Systems S. Nefti and J. Gray (Editors)
Volume 72 Control Theory: a guided tour, 3rd Edition J.R. Leigh
Volume 73 Adaptive Sampling with Mobile WSN K. Sreenath, M.F. Mysorewala, D.O. Popa and F.L. Lewis
Volume 74 Eigenstructure Control Algorithms: applications to aircraft/rotorcraft handling qualities
design S. Srinathkumar
Volume 75 Advanced Control for Constrained Processes and Systems F. Garelli, R.J. Mantz and H. De
Battista
Volume 76 Developments in Control Theory towards Glocal Control L. Qiu, J. Chen, T. Iwasaki and H. Fujioka
(Editors)
Volume 77 Further Advances in Unmanned Marine Vehicles G.N. Roberts and R. Sutton (Editors)
Volume 78 Frequency-Domain Control Design for High-Performance Systems J. O’Brien
Volume 81 Optimal Adaptive Control and Differential Games by Reinforcement Learning Principles
D. Vrabie, K. Vamvoudakis and F. Lewis
Volume 88 Distributed Control and Filtering for Industrial Systems M. Mahmoud
Volume 91 An Introduction to Fractional Control D. Valério and J. Costa
 Application of
Dimensional
Analysis in Systems
Modeling and
Control Design
Pedro Balaguer

The Institution of Engineering and Technology

Published by The Institution of Engineering and Technology, London, United Kingdom

The Institution of Engineering and Technology is registered as a Charity in

England & Wales (no. 211014) and Scotland (no. SC038698).

© The Institution of Engineering and Technology 2013

First published 2013

This publication is copyright under the Berne Convention and the Universal Copyright
Convention. All rights reserved. Apart from any fair dealing for the purposes of research
or private study, or criticism or review, as permitted under the Copyright, Designs and
Patents Act 1988, this publication may be reproduced, stored or transmitted, in any
form or by any means, only with the prior permission in writing of the publishers, or in
the case of reprographic reproduction in accordance with the terms of licences issued
by the Copyright Licensing Agency. Enquiries concerning reproduction outside those
terms should be sent to the publisher at the undermentioned address:

The Institution of Engineering and Technology

Michael Faraday House
Six Hills Way, Stevenage
Herts, SG1 2AY, United Kingdom

www.theiet.org

While the author and publisher believe that the information and guidance given
in this work are correct, all parties must rely upon their own skill and judgement when
making use of them. Neither the author nor publisher assumes any liability to
anyone for any loss or damage caused by any error or omission in the work, whether
such an error or omission is the result of negligence or any other cause. Any and all
such liability is disclaimed.

The moral rights of the author to be identified as author of this work have been
asserted by him in accordance with the Copyright, Designs and Patents Act 1988.

British Library Cataloguing in Publication Data

A catalogue record for this product is available from the British Library

ISBN 978-1-84919-621-5 (hardback)

ISBN 978-1-84919-622-2 (PDF)

Typeset in India by MPS Limited

Printed in the UK by CPI Group (UK) Ltd, Croydon
 Contents

Preface ix

1 Introduction 1
1.1 What is dimensional analysis? 1
1.2 What is dimensional similarity? 2
1.3 Application of dimensional analysis to science in general 3
1.3.1 Structure of physical relations 3
1.3.2 Dimensionless representation 5
1.3.3 Dimensional similarity 5
1.4 Application of dimensional analysis to control problems 6
1.4.1 Identification and model validation 8
1.4.2 Control theory 8
1.4.3 Control engineering 8
1.5 Book contents 9

2 Dimensional analysis and dimensional similarity 11

2.1 Physical quantities, units, and dimensions 11
2.1.1 Physical quantity 11
2.1.2 Units 11
2.1.3 Dimensions: fundamental and derived 11
2.1.4 Arithmetic of dimensions 12
2.2 Systems of units: dependence and independence of dimensions 13
2.2.1 System of units 13
2.2.2 Monomial power law 14
2.2.3 Dependent and independent dimensions 14
2.3 Buckingham pi theorem 14
2.4 Matrix approach for finding the dimensionless numbers 16
2.4.1 The dimensional matrix 16
2.4.2 The dimensional set 17
2.5 Dimensional similarity 18
2.5.1 Scale factors 20
2.5.2 Model law 21
2.6 Exercises 21
References 22
vi Application of dimensional analysis in systems modeling and control design

3 Dynamical systems: dimensionless representation 23

3.1 Introduction 23
3.2 Transfer function dimensionless representation 23
3.2.1 Transfer function parameters dimensions 24
3.2.2 Transfer function parameters with independent
dimensions 24
3.2.3 Transfer function dimensionless numbers 25
3.2.4 Dimensionless transfer function 26
3.3 State space dimensionless representation 27
3.3.1 Interpretation of the state space dimensionless
transformation 30
3.4 Comparison between transfer function and state space
dimensionless representation 32
3.5 Discrete time models dimensionless representation 35
3.5.1 Discrete time transfer function dimensionless
representation 35
3.5.2 Discrete time state space model dimensionless
representation 36
3.6 Exercises 38
References 38

4 Dynamical systems: dimensional similarity 41

4.1 Introduction 41
4.2 Continuous time dynamical systems similarity 41
4.2.1 Transfer function dimensional similarity 41
4.2.2 State space dimensional similarity 44
4.3 Discrete time dynamical system similarity 46
4.3.1 Discrete time transfer function similarity 46
4.3.2 Sampled-data transfer function similarity 46
4.3.3 Discrete state space similarity 47
4.4 Exercises 48
References 49

5 Dimensionless systems identification and model order reduction 51

5.1 Introduction 51
5.2 General procedure 51
5.3 Example 1: Second order inverse response model identification 52
5.3.1 Problem statement 52
5.3.2 Dimensionless representation of second order inverse
response model 53
5.3.3 Identification procedure 54
5.3.4 Application examples 61
5.4 Example 2: Reduced effective transfer function reduction for
PID decentralized control 64
 Contents vii

5.4.1 Problem statement 64

5.4.2 Dimensionless representation of the reduced effective
transfer function 67
5.4.3 Inverse response analysis 68
5.4.4 Reduced order model: general case 72
5.4.5 Reduced order model: particular cases 73
5.4.6 Application examples 75
References 77

6 Homogeneity of PID tuning rules 79

6.1 Introduction 79
6.2 Homogeneous PID tuning rules 80
6.2.1 Dimensionless controller parameters 80
6.2.2 Homogeneous tuning rules characterization 80
6.2.3 Dimensionless controller representation with
homogeneous tuning rules 83
6.3 Closed loop transfer functions 83
6.3.1 Loop transfer function 83
6.3.2 Dimensionless closed loop transfer functions 85
6.4 Optimality of homogeneous tuning rules 87
6.4.1 Weighting factors 88
6.5 Homogeneous and nonhomogeneous tuning rules 89
References 94

7 Dimensionless PID tuning rules comparison 97

7.1 Introduction 97
7.2 Elements of the comparative framework 98
7.3 Dimensionless comparative framework 99
7.4 Dimensionless elements 100
7.4.1 Loop transfer function 100
7.4.2 Dimensionless closed loop transfer functions 100
7.4.3 Dimensionless integral errors 100
7.4.4 Indexes 102
7.5 Application example 102
7.5.1 PID tuning rules dimensionless characterization 103
7.5.2 Dimensionless sensitivity bandwidth comparison 103
7.5.3 Dimensionless sensitivity peak comparison 104
7.5.4 Dimensionless integral absolute error 104
7.5.5 Dimensionless control action variation 106
7.6 PID tuning rules selection 106
References 108

8 Control of dimensionally similar systems 109

8.1 Introduction 109
viii Application of dimensional analysis in systems modeling and control design

8.2 Control of dimensionally similar systems 110

8.3 Complete similarity 111
8.3.1 Continuous time control 111
8.3.2 Discrete time control 111
8.4 Partial similarity 113
8.5 Experimental case study 115
References 119

9 Adaptive systems 121

9.1 Introduction 121
9.2 Actuator limitations and dimensionally similar model reference 122
9.2.1 Control effort 122
9.2.2 Similar model reference adaptive control 123
9.3 SMRAC for first order plants 126
9.4 SMRAC for arbitrary order plants 129
9.4.1 SMRAC control scheme 129
9.4.2 SMRAC stability analysis 131
9.4.3 SMRAC operation modes 133
9.5 Application example 134
References 137

Index 139
 Preface

Purpose of the book

This book is about applying dimensional analysis concepts to control theory and
control engineering problems. We find two fundamental book objectives: (i) to intro-
duce the fundamentals of dimensional analysis to control engineers and researchers
and (ii) to show how to exploit the benefits that dimensional analysis can provide to
control theoretic and practical problems.
The rationale for writing the book lies in the novelty of the topic for the control
community and its applicability to control engineering problems. As a result, this book
is intended to familiarize the reader with the basic concepts, principles, and methods
of dimensional analysis and its applications to modeling and control design. Since
the book is introductory in nature, we have considered control theory applications
at senior undergraduate level or beginning graduate students. In particular we have
focused on modeling and PID process control. More specific chapters on particular
fields on control systems design, such as robust control of dimensionally similar
systems and adaptive control, require previous knowledge on the fundamentals of
those fields.
Prerequisites
The book is elementary. A background in undergraduate mathematics and physics,
in particular linear algebra, ordinary calculus, and physic fundamentals, is sufficient
as a prerequisite for the first half of the book. Applications of modeling and PID
process control may also be followed by undergraduate students with a process con-
trol related minor. The book should therefore be accessible to a wide spectrum of
students. Control engineers may also follow the great majority of chapters related to
process control.
Content and arrangement
The content of the book has been arranged in four major blocks. Dimensional analysis
fundamentals, covered by Chapters 1 and 2, motivates the use of dimensional analysis
and introduces its basic concepts. Next, we develop the Application of dimensional
analysis and dimensional similarity to systems and control, in Chapters 3 and 4,
where fundamental results are obtained for the dynamical system dimensionless rep-
resentation and the characterization of dimensionally similar systems. Dimensional
analysis and process control is covered in Chapters 5–7. Dimensional analysis is
applied to system identification and model order reduction, the introduction of the
homogeneity concept of PID tuning rules, and the development of a framework for
homogeneous PID tuning rules comparison. Finally, New applications of dimensional
x Application of dimensional analysis in systems modeling and control design

analysis to advanced control systems, covered by Chapters 8 and 9, presents the poten-
tial research applications of dimensional analysis to advanced control systems such
as robust control of dimensionally similar systems and adaptive control.
Acknowledgments
This book has been the result of the research work of many people. First of all, we
are in debt to Dr. Andrew G. Alleyne, Dr. Sean Brennan, and Dr. Haftay Haily for
their initiative and development in the application of the dimensional analysis to the
control field.
In particular, I would like to thank Víctor Alfaro for his hospitality, and Orlando
Arrieta and Julio Romero for their friendship. Chapter 5 would not have been pos-
sible without them. I am also in debt to my former colleagues Asier Ibeas, Carles
Pedret, and Salvador Alcántara. Their knowledge and wisdom have made possible
Chapter 6. Finally I would like to thank Alejandro Claramonte for his intense work
and dedication in some parts of Chapter 8.
The author wishes to thank Dr. Nigel Hollingworth for his interest, help, and
support, making possible this work, and to Ting Ting Liu for her help and patience
with the book development. I would like to thank also Dr. Reza Katebi for his guidance,
wise counseling, and friendship.
Finally, I would like to thank my wife Maria for her patience while writing
this book. I dedicate this book to her and to my daughter Mireia.
 Chapter 1
Introduction

1.1 What is dimensional analysis?

Scientific theories are inventions of the human mind that explain and predict phe-
nomena. The development of scientific theories uses mathematics as a fundamental
tool to provide a formal framework for deductive reasoning. Scientific theories that
explain physical relations deal with physical dimensions. In this case, the mathe-
matical framework is enriched with another axiom, derived from a very simple but
fundamental idea: physical laws do not depend on the arbitrariness in the choice of
basic units of measurements, that is physical laws are homogeneous.
A mathematical equation is valid if both sides of the equality sign are equal.
A physical equation also requires that both sides of the equality sign have the same
physical dimensions, that is the equation must be homogeneous. As a result, a physical
equation, for validity, requires not only mathematical equality but also dimensional
homogeneity.
In the realm of mathematics the equation

x2 = x (1.1)

is perfectly valid and has the solutions x = 0 and x = 1. However, consider now
that equation (1.1) has a physical meaning. For instance variable x is the distance
between two points, thus it has length dimension, that is [x] = L. In this case, equa-
tion (1.1) is not a dimensionally homogeneous equation because the left hand side
has area dimension, [x2 ] = L2 , whereas the right hand side has length dimension,
[x] = L. We may conclude that physical equations that are not dimensionally homo-
geneous are not physically valid because in case of non-homogeneity the physical
magnitudes would be dependent on the choice of the measurements units.
The origins of dimensional analysis may be tracked back to several centuries,
although its foundations were established two centuries ago and it was the last cen-
tury that witnessed the application of dimensional analysis to a wide variety of physical
problems. One of the difficulties found in developing the dimensional analysis was
the existence of derived quantities, which were not established until the works of
Newton and Leibniz around the beginning of the eighteenth century.
At the end of the nineteenth century the advances of dimensional analysis were
applied to fluid dynamic problems, which had a difficult analytic solution, such as
the works of Helmholtz on hydrodynamics and Rayleigh on the theory of sound.
2 Application of dimensional analysis in systems modeling and control design

Finally, the fundamental theorem of dimensional analysis was established inde-

pendently by Federmann, Riabouchinsky, and Buckingham at the beginning of the
twentieth century. This fundamental theorem, known as the Buckingham pi theorem,
states that a physical relation can be reformulated in a dimensionless form, in such
a way that the number of dimensionless parameters is less than or equal to the num-
ber of dimensional parameters. The major implication of the theorem is that for any
physical relation there exists a compressed representation because it has the fewest
required parameters. Furthermore, these parameters are dimensionless.
Summing up, dimensional analysis is based on the homogeneity property of
physical equations. Homogeneity is a further constraint on mathematical equations
that is a necessary condition for correctness of physical equations. The fundamental
result of dimensional analysis is the Buckingham pi theorem, which establishes the
existence of dimensionless representation of physical equations. The dimensionless
representation is useful in several ways, one of them is to establish dimensional
similarity between systems.

1.2 What is dimensional similarity?

The use of scaled models is prevalent in engineering applications because they
allow performance of experimental tests without requiring the physical system
(i.e. prototype) to be constructed. The use of models is beneficial because it allows
experimental data to be obtained that would be very difficult to obtain analytically or
computationally, and, at the same time, provides safety and economic advantages.
However, to extrapolate model experimental data to prototype in a meaningful
way, the model must be built dimensionally similar with respect to the prototype.
One trait of dimensional similarity is geometrical similarity, as can be seen in
Figure 1.1. However, there are other forms of dimensional similarity such as kine-
matic similarity or thermal similarity.
The existence of dimensional similarity is also related to the existence of scaling
factors between model and prototype. In fact it is the existence of the scaling factors
that permits the data extrapolation from model to prototype. A scaling factor with
respect to a particular physical variable is the quotient of the magnitudes of that
variable for the prototype and for the model. It is a common misconception to talk
about the model scaling factor as if it was unique. From the scaling factor definition

Figure 1.1 Model and prototype showing geometrical similarity

 Introduction 3

it follows that there are as many scaling factors as physical parameters involved in
the model experimentation.
The fundamental Buckingham pi theorem of dimensional analysis establishes
that any physical relation can be written in dimensionless form by using dimension-
less parameters. The dimensionless representation of physical systems establishes
a formal baseline for dimensional similarity characterization: two systems, model
and prototype, are dimensionally similar if both systems have the same dimension-
less representation. In this way the dimensionless numbers are the fingerprints of
dimensional similarity.
The use of models has also a long history. It began in the sixteenth century with
the use of models for investigating the strength of materials for buildings and the
construction of machines. At the end of the nineteenth century dimensional sim-
ilarity was systematically applied to solve hydrodynamic problems, giving rise to
several famous dimensionless numbers such as the Froude number, the Reynolds
number, and the Prandtl number, among others. In fact, the fluid dynamic sci-
ence uses models extensively because the problems faced have a difficult analytical
solution and are computationally very demanding. However, the use of models
is pervasive in any area of engineering such as mechanics, thermodynamics, and
chemistry.

1.3 Application of dimensional analysis to science in general

Once the concepts of homogeneity, dimensional analysis, and dimensional similarity
are established, we introduce the applications of dimensional analysis to science in
general. These are:
1. Structure of physical relations: The homogeneity property may be helpful to find
the structure of physical relations.
2. Dimensionless representation: The Buckingham pi theorem provides a com-
pact representation of physical relations through dimensionless numbers. This is
beneficial because:
● It helps to understand the phenomenon under study, not only the depending
quantities but also their relations.
● The results can be presented in a more compact way with fewer variables.
3. Dimensional similarity: If we can establish dimensional similarity among several
objects of study then:
● The experimental results can be generalized.
● The number of experiments can be minimized.

1.3.1 Structure of physical relations

Physical relations are derived through experimentation and/or analytical derivation
from first principles. However, in some cases, the physical relation can be derived
without any prior knowledge or experimentation, by only applying the homogene-
ity property to the physical parameters involved in the phenomenon under study.
4 Application of dimensional analysis in systems modeling and control design

r
vi(t) v0(t)

Figure 1.2 Circuit showing the capacitor charge through a resistance. vi (t) is the
voltage input, v0 is the capacitor charge, c is the capacitor capacitance,
and r is the electrical resistance

A classical example is the derivation of the period tp of a single pendulum of

length l, in the presence of gravitational acceleration
 g. By looking for a homogene-
ous equation it is possible to find that tp = k l/g, with k an unknown constant.
We show the procedure of finding the structure of a physical relation by means
of an electrical problem. Consider the simple circuit shown in Figure 1.2, which
shows the charge of a capacitor through a resistance. Let’s consider that the time
needed to charge capacitor tc is a function of the capacitor capacitance c and the
resistance r. Mathematically we express tc = f (r, c) and we want to find the form of
the function f (·, ·). The units of the involved parameters are [tc ] = T , [c] = AT /V ,
and [r] = V /A, where T is time, V is voltage, and A is electrical current.
Recalling that the product rc has time dimension, that is [rc] = T , it is easy to
derive the following dimensionless number using the three involved parameters:
tc
= (1.2)
RC
 must be constant under any arbitrary dimensions used to measure time, voltage,
and current, due to its dimensionless nature. As a result, (1.2) can be written as

tc = rc (1.3)

with  a dimensionless constant. It follows that the time needed to charge a capac-
itor is proportional to the capacitance multiplied by the resistance. Therefore, we
have established that (i) the function f (·, ·) does not depend on two parameters but
just on one parameter rc, and (ii) the dependence is linear through an unknown
constant . It is important to stress that we have arrived at the conclusion by just
applying the homogeneity principle and no physical knowledge has been applied
beyond the units of the parameters involved.
In the preceding discussion we have not considered the voltage input vi nor the
initial capacitor charge v0 . Moreover, the capacitor charge time tc does not explic-
itly define the final charge vf to be reached. However, despite these not considered
factors, if they are held constant, it is true that the capacitor charge time is given by
tc = rc, and the value of the dimensionless constant  is given by  = g(vi , v0 , vf ),
with g an unknown function.
 Introduction 5

1.3.2 Dimensionless representation

The existence of a dimensionless representation allows for the understanding of the
phenomenon under study beyond the isolated enumeration of physical parameters
involved. Following the previous electrical example shown in Figure 1.2, we ask on
the evolution of capacitor voltage v0 , when the voltage input is a step of value vi .
We envisage that mathematically v0 (t) = f (t, vi , r, c), that is the voltage output v0 is
a function of time t, the amplitude of step vi , the capacitance c, and the resistance r.
Of course v0 (t) can be analytically obtained by solving a differential equation, but we
want to obtain some insight by mere dimensional evaluation.
First we form the following dimensionless numbers: v0 (t)/vi and t/rc. In this
way the dimensional expression v0 (t) = f (t, vi , r, c) can be expressed in dimension-
less form as
 
v0 (t) t
=φ (1.4)
vi rc

with φ an unknown function. It follows that v0 (t) depends on just one parameter,
the dimensionless time t/rc. However, the dependence of v0 (t) with respect to vi is
just linear as easily seen by rewriting previous expression as
 
t
v0 (t) = vi φ (1.5)
rc

The amount of information provided in (1.5) by mere dimensionless representa-

tion is of great significance because the parameters’ dependence on v0 (t) has been
reduced from 4 to 2, and one of them shows a linear relationship.
Another application of the dimensionless representation is to present exper-
imental results in a more compact way. Consider for instance that a series of
experiments is performed to investigate the form of the function v0 (t) = f (t, vi , r, c).
In this way eight combinatorial experiments are performed with the following val-
ues: r = {1, 2} Ohms, c = {1, 3} Farads, and vi = {1, 2} Volts. The experimental
results can be seen in Figure 1.3(a).
Now, if we take the same eight experiments but represent them in dimension-
less form by (i) scaling the voltage output v0 (t)/vi and (ii) scaling the time axis by
t/rc, we obtain Figure 1.3(b). It must be stressed that Figure 1.3(b) is the represen-
tation of the eight experiments, which all collapse in a single graph, showing the
same experimental results in just one graph.

1.3.3 Dimensional similarity

If dimensional similarity can be established between physical systems, the experi-
mental results can be generalized. For instance the capacitor charge v0 (t) follows the
dimensional equation

v0 (t) = (1 − e−t/rc )vi (1.6)

6 Application of dimensional analysis in systems modeling and control design

(a) Dimensional experiments

2
Voltage (v )

0
0 2 4 6 8 10
Time (s)

(b) Dimensionless representation of experiments

1
Gain v0 /vi

0.5

0
0 2 4 6 8 10
Dimensionless time t/RC

Figure 1.3 Dimensional and dimensionless representations of experiments. (a) The

eight experimental results for the capacitor charge, for r = {1, 2}
Ohms, c = {1, 3} Farads, and vi = {1, 2} Volts. (b) The same
experimental results as a function of dimensionless variables. In
dimensionless form the eight experiments provide the same result

which can be written in dimensionless form as

v0 (t)/vi = 1 − e−t/rc (1.7)

Dimensionless equation (1.7) depends only on one dimensionless number t/rc.

Recalling that two systems are dimensionally similar when they have the same dimen-
sionless numbers, it can be deduced that for any values of r and c, a time scale that
makes the systems dimensionally similar can be calculated. Consequently, the exper-
imental results can be generalized. In this way, the number of experiments can be
minimized. In fact, in the previous capacitor voltage experiment, only one experi-
ment is required and the capacitor voltage evolution can be easily obtained for any
value of vi , r, and c by amplitude and time scaling, as previously shown.

1.4 Application of dimensional analysis to control problems

The objective of control engineering is to make a physical system to behave in a
specified way, defined normally by some performance measure. The types of systems
 Introduction 7

to be controlled are enormous and each day new applications to control are envisaged,
designed, and applied. For example control applications range from classical fields
such as mechanical systems, electronic devices, and chemical processes, to newer
fields such as economics, biosciences, and quantum mechanics, to name just a few.
Dimensional analysis and dimensional similarity can be applied to systems and
control areas to provide solutions to known problems and to develop new approaches.
In order to structure the potential applications of dimensional analysis and dimen-
sional similarity to systems and control, the control problem is characterized by
means of three elements, the physical system to be controlled, the model, and the
controller.
The relations among the physical system, the model, and the controller are
shown in Figure 1.4. The physical system and the model are related by modeling,
identification, validation, and model reduction issues. Once a model is available,
the model and the controller are related by control theory, of which one of the main
objectives is to design proper controllers on the basis of mathematical models. Finally,
the controller and the physical system are related by control engineering, aimed
toward correct implementation and validation of control algorithms.
The conceptualization of control problems as presented in Figure 1.4 is use-
ful to classify the current contributions of dimensional analysis to control problems
and to envisage further applications. In this book we consider the following applica-
tions of dimensional analysis to control theory:

1. Identification and model validation:

● Simple identification algorithms can be designed exploiting the model
dimensionless representation.
● The dimensionless representation is also useful in model order reduction for
multivariable systems.

Control

Control Control
theory engineering

Mathematical Physical
model system

Identification
& modelling

Figure 1.4 Elements and relations of control systems

8 Application of dimensional analysis in systems modeling and control design

2. Control theory:
● The homogeneity property of proportional integral derivative (PID) tuning
rules is shown to be a fundamental property that has influence on the type
of functions used for deriving PID tuning rules.
● The use of dimensionally similar reference models in model reference
adaptive control with input constraints allows for the guarantee of stability
while maintaining a dimensionally similar behavior at the output.
3. Control engineering:
● The PID tuning rules comparison and selection are simplified by the homo-
geneity property of tuning rules and the application of a dimensionless
framework.
● The controller extrapolation from laboratory model experiments to proto-
types can be readily performed by scaling.

1.4.1 Identification and model validation

The dimensionless representation of dynamical systems, both in the transfer func-
tion approach and in the state space representation, is presented in Chapter 3. Once
a dimensionless representation is obtained, it is straightforward to characterize the
dimensional similarity as shown in Chapter 4.
The application to system identification of the dimensionless representation per-
mits to obtain simpler identification algorithms by exploiting limiting cases in the
dimensionless numbers. An example is presented in Chapter 5, where a very simple
graphical algorithm for identification of second order inverse response systems is
presented. Other existing approaches in the literature need to perform integration or
to solve a set of nonlinear equations. In the same way, the model order can be reduced
by applying dimensional analysis. An application example of model order reduction
of a multivariable system for decentralized control is also presented.

1.4.2 Control theory

In Chapter 6 the fundamental property of dimensional homogeneity is applied to
PID tuning rules. Questions related to optimality and closed loop dimensional simi-
larity are investigated. It follows that although the PID tuning rules homogeneity is
not mandatory for correctness, it has implications that make homogeneity a natural
choice.
Adaptive control also beneficiates from dimensional analysis. In Chapter 9 it
is shown how the design of model reference adaptive controllers with a reference
model that is no longer constant but changes in a dimensionally similar way yields a
suitable control scheme in the presence of input saturation.

1.4.3 Control engineering

Dimensional analysis also has direct application to control engineering. The homo-
geneity of tuning rules has an important practical application in the derivation of
 Introduction 9

procedures for PID tuning rules comparison as presented in Chapter 7. In fact,

the existence of a huge set for PID tuning rules makes it difficult to choose which
one is the most suited for the problem at hand. In the case of homogeneous tuning
rules, it follows that the comparison may be performed in a dimensionless way, thus
exploiting the benefits of parameter reduction and simplification of the limit cases.
Another application of dimensional similarity to control engineering is the
use of laboratory-controlled models to design controllers that, by proper scaling,
may be applied to prototypes. Chapter 8 presents the fundamental results for both,
complete similarity and partial similarity. Partial similarity is of interest because to
achieve complete similarity is costly and, sometimes, not possible. In this case it is
shown how robustness properties are considered, enabling the use of experimental
results to design controllers for prototypes.

1.5 Book contents

Chapter 2 The fundamental concepts, theorems, and results of dimensional analy-
sis and dimensional similarity are introduced in this chapter. Although
the chapter is self-contained, bibliographical references are presented
to guide on further developments and applications of dimensional anal-
ysis to engineering problems. First the fundamental definitions of
physical dimension, systems of units, and independence of dimen-
sions are introduced. The fundamental Buckingham pi theorem of
dimensional analysis is presented from a scalar viewpoint and from
an algebraic approach. In this way results can be easily obtained for
both the transfer function approach and the state space representa-
tion. Finally, the concepts and subtleties of dimensional similarity are
explained.
Chapter 3 The main objective of this chapter is to obtain dimensionless repre-
sentation of models given in transfer function form and state space
form. The dimensionless form is advantageous because on the one hand
the dimensionless representation has the minimum number of param-
eters required to define the dynamical model and, on the other hand,
several distinct dimensional models may have the same dimensionless
representation. As a result, being the dimensionless representation of a
compressed representation of dynamic systems, it may be used as a mea-
sure of system complexity. The dimensionless representation of dynam-
ical systems is the beginning for many of the applications of dimensional
analysis and similarity to systems and control. In fact the dimen-
sionless representation characterizes the similarity between dynamical
systems.
Chapter 4 The dimensionless representation defines the concept of dimensional
similarity between dynamical systems. The set of dimensional models
that have the same dimensionless representation is the set of dimen-
sionally similar models. In this way, the dimensionless representation
10 Application of dimensional analysis in systems modeling and control design

characterizes the dimensionally similar models as the models that have

the same dimensionless representation. The certification of dimen-
sional similarity between systems allows the extrapolation of experi-
mental results from model to prototype, and the application on novel
adaptive control schemes in the presence of input saturation.
Chapter 5 The use of dimensionless models can be exploited in the identification
and model order reduction procedures. The dimensionless representa-
tion makes it possible to point out limiting cases that may be exploited
by identification algorithms and to reduce the order of models. A receipt
is given in order to exploit the limiting cases, and it is shown how it can
be used in the identification of second order inverse response systems
and in the model reduction for decentralized control of multivariable
processes.
Chapter 6 Dimensional analysis raises the question of the dimensional homo-
geneity of PID tuning rules. Whereas physical equations must be
homogeneous for correctness, the PID tuning rules are just an algo-
rithm that relates system parameters with controller parameters. The
chapter presents the properties of homogeneous tuning rules and their
implications on the process control. It follows that the homogeneity
property of the tuning rules implies the mapping of dimensionally sim-
ilar open loop systems into dimensionally similar closed loop systems,
allowing for instance the development of frameworks for PID tuning
rules comparison.
Chapter 7 Following with PID tuning rules, if they are homogeneous, it is pos-
sible to derive a comparative framework based on dimensionless num-
bers. It has the benefit of using the minimum required set of parameters
for PID comparison, the dimensionless numbers, and it is possible to
exploit the limiting values. The chapter presents a recipe to derive a
suitable comparative framework and one example is presented.
Chapter 8 In the case of dimensional similarity experimental results from models
can be generalized to the prototype. In this chapter the extrapolation of
controllers designed on the basis of models to be used on prototypes
is investigated. The case of complete similarity and the case of partial
similarity are considered, where robust issues are considered for safer
controller extrapolation.
Chapter 9 In this chapter we show how dimensional analysis can be used to design
adaptive schemes. A novel model reference adaptive scheme is pre-
sented. The novelty of the approach lies on the fact that the reference
model is no longer constant but dimensionally similar. In this way,
although there are input constraints, the behavior is always similar to
the original model. It implies that the dimensionless model properties
are kept constant.
 Chapter 2
Dimensional analysis and dimensional similarity

2.1 Physical quantities, units, and dimensions

2.1.1 Physical quantity

A physical quantity is a physical property or quality that can be quantified by mea-
surement. A physical quantity is defined by the product of a numerical value and a
unit of measurement, that is

Physical quantity = (Numerical value) × (Unit) (2.1)

For instance, when we refer to the fact that the length of a cable is 5 meters, we are
stating the magnitude of a physical quantity.

2.1.2 Units
In order to obtain a physical quantity it is necessary to perform a measurement. The
measurement is the comparison of the physical quantity with a standard, the unit of
measurement. Following the cable length example, we could also say that the cable
length is 5000 millimeters, the only difference being the units used to describe the
physical quantity. Thus, in both cases we are referring to the same physical quantity
(i.e. the cable length is the same), but the difference with the numerical values is due
to the difference in the units used for measurement (e.g. meters vs millimeters).

2.1.3 Dimensions: fundamental and derived

The dimension is defined as a qualitative description of a sensory perception of a
physical property or quality. Previous examples of physical quantities deal with
the same dimension, the length dimension. We write the dimension of a physical
quantity φ between square brackets, that is [φ]. For example the quantity lc that
describes the length of the cable has length dimensions, that is [lc ] = L.
The dimensions can be classified into fundamental and derived. The funda-
mental dimensions are the ones arbitrarily chosen for the study of a physical
phenomenon. For instance consider a mechanical phenomenon. The fundamental
dimensions of the international system of units (SI) for the mechanical phenomenon
are length L, mass M , and time T . Furthermore, the fundamental units are meter m,
kilogram kg, and second s.
12 Application of dimensional analysis in systems modeling and control design

Derived quantities are obtained from fundamental ones according to physi-

cal laws. Correct derived quantities are obtained by multiplication and division of
fundamental ones, being addition and subtraction forbidden. For example the velocity
dimension is obtained by dividing the fundamental length dimension by the funda-
mental time dimension, that is [v] = LT −1 . The corresponding derived unit is meters
per second ms−1 .

2.1.4 Arithmetic of dimensions

In this section we review the fundamentals of the arithmetic of dimensions. The results
are quite straightforward and are presented without proof. The reader is referred to
[2, p. 95] for a deeper treatment.

Product
The product of the dimensions of a set of n variables is the dimension of the product
of these sets of n variables, that is if v1 , v2 , . . . , vn is a set of n variables, then

[v1 ][v2 ], . . . , [vn ] = [v1 , v2 , . . . , vn ] (2.2)

Power
The dimension of a power of a variable is the power of the dimension of that variable,
that is

[vn ] = [v]n (2.3)

Quotient
The dimension of the quotient of nth-order differentials of variable v1 with respect to
variable v2 is the quotient of dimensions of v1 and v2n , that is
 
d n v1 [v1 ]
= n (2.4)
dv2n [v2 ]

Integral
The dimension of a multiple integral is the product of the dimensions of the integrand
and the dimensions of all the independent variables. That is, if
  
I= ... y(v1 , v2 , . . . , vn )dv1 dv2 , . . . , dvn (2.5)

then

[I ] = [ y(v1 , v2 , . . . , vn )][v1 ][v2 ], . . . , [vn ] (2.6)

 Dimensional analysis and dimensional similarity 13

2.2 Systems of units: dependence and independence

of dimensions

2.2.1 System of units

A system of units consists of all the fundamental and derived dimensions, together
with the units chosen for their measurement, for instance the centimeter-gram-second
(CGS) system or the international system (SI). Table 2.1 presents the fundamental
dimensions and their measurement units used in the SI. In Table 2.2 some of the
SI-derived units are also presented.
Note that with reference to the dimensions length, mass, and time, both CGS and
SI share the same physical properties, although the measurement units are different.
A set of units that differ only in the measurement units but have the same dimensions
is called a class of systems of units.

Table 2.1 Fundamental dimensions and units of SI

Physical quantity Symbol Name of SI unit Abbreviation

Length L Meter m
Mass M Kilogram kg
Time T Second s
Electric current A Ampere A
Thermodynamic temperature  Kelvin K
Amount of substance Ml Mole mol
Luminous intensity I Candela cd

Table 2.2 Derived SI units

Unit Symbol Abbreviation

◦
Degree celsius C K
Hertz Hz s−1
Newton N m · kg · s−2
Pascal Pa N · m−2
Joule J N·m
Watt W J · s−1
Volt V W · A−1
Coulomb C s·A
Ohm  V · A−1
Siemens S A · V−1
Farad F C · V−1
Weber Wb V·s
Henry H Wb · A−1
Tesla T Wb · m−2
14 Application of dimensional analysis in systems modeling and control design

2.2.2 Monomial power law

The dimensions of a physical quantity are always a power law monomial. In this way,
if we consider length, mass, and time dimensions, any derived dimension is always
of the form:

[Quantity] = La M b T c (2.7)

where a, b, and c are dimensionless coefficients. The result can be proven by using
two systems of units of the same class, as shown in Reference 1 or by approximation
theory [2, p. 133]. The result follows by the fact that all systems of units within a
given class are equivalent.

2.2.3 Dependent and independent dimensions

Given a set of quantities a1 , . . . , ak , these quantities are said to have independent
dimensions if none of these quantities have dimensions that can be represented in
terms of a product of powers of the dimensions of the remaining quantities. That is
there are no p, . . . , q, u, . . . r, such that

[aj ] = [a1 ]p , . . . , [aj−1 ]q [aj+1 ]u , . . . , [ak ]r , j = 1, . . . , k (2.8)

For example density ([ρ] = ML−3 ), velocity ([v] = LT −1 ), and force ([ f ] = MLT −2 )
have independent dimensions. In fact there are no x and y that accomplish the
following equations:

[ρ] = [v]x [ f ]y (2.9)

−2 −3 x −1 y
LMT = (ML ) (LT ) (2.10)

because the set of equations

−3x + y = 1 (2.11)
x =1 (2.12)
y =2 (2.13)

has no solution. The concept of independence of dimensions is crucial for the

fundamental result of dimensional analysis, the Buckingham pi theorem.

2.3 Buckingham pi theorem

Consider a physical relation with Nv parameters involved, where the governed

parameter a is a function of Nv−1 governing parameters, that is

a = f (a1 , . . . , aNd , aNd+1 , . . . , aNv−1 ) (2.14)

 Dimensional analysis and dimensional similarity 15

The Buckingham pi theorem [1] states that if there is a set of Nd param-

eters (a1 , . . . , aNd ), which have independent dimensions then there are Nv − Nd
dimensionless parameters. One dimensionless number  is related to the governed
parameter a and j j ∈ [1, Nv − Nd − 1] are related to the Nv − Nd − 1 governing
parameters, in such a way that the relation (2.14) can be stated as

 = (1 , . . . , Nv −Nd −1 ) (2.15)

with  the dimensionless number related with the governed parameter defined as

a
= p (2.16)
a1 , . . . , arNd

where the exponents p, . . . , r are chosen such that the parameter  is dimension-
less. Moreover, each one of the dimensionless numbers related with the govern-
ing parameters is given by

aNd +j
j = p r (2.17)
a1j , . . . , aNjd

for j ∈ [1, Nv − Nd − 1], where the exponents pj , . . . , rj are chosen such that the
parameter j is dimensionless. As a result there are Nv − Nd − 1 dimension-
less numbers of the governing parameters (i.e. j j ∈ [1, Nv − Nd − 1]), plus one
dimensionless number related with the governed parameter a (i.e. ). It then
follows that the function f can be written in terms of a function of a smaller number
of variables in the following form:

p
f (a1 , . . . , aNv−1 ) = a1 , . . . , arNd (1 , . . . , Nv −Nd −1 ) (2.18)

Example 2.1. Consider the capacitor charge through a resistance as presented

in Figure 1.2. The capacitor voltage v0 is described by the dimensional function
v0 (t) = f (t, vi , r, c), where vi is the input voltage, r is the resistance, c is the capac-
itance, and t is time. The dimensions of the involved variables are [v0 ] = [vi ] = V ,
[t] = T , [r] = V /A, and [c] = AT /V , where V is voltage dimension, T is time
dimension, and A is electrical current dimension. The number of variables is Nv = 5,
whereas the number of variables with independent dimensions is Nd = 3, as a result
there are Nv − Nd = 2 dimensionless numbers. A possible subset of the governing
parameters with independent dimension is {vi , r, c}. As a result the dimensionless
governed parameter is given by

v0
= (2.19)
vi
16 Application of dimensional analysis in systems modeling and control design

and the dimensionless number of the governing parameter is

t
1 = (2.20)
rc
so the capacitor charge can be written in dimensionless form as
 
t
v0 = vi  (2.21)
rc

2.4 Matrix approach for finding the dimensionless numbers

In this section we present a structured approach in order to obtain the dimensionless
numbers of any dimensional relation. The approach is based on matrix computa-
tions, so it is very convenient for computer implementation.

2.4.1 The dimensional matrix

The dimensional matrix [2, p. 134] is a matrix representation of the dimensions of
the involved variables. The rows of the dimensional matrix are the dimensions, and
the columns are the variables. The entries of the dimensional matrix are the expo-
nents to which each particular dimension must be raised in the particular variable. For
instance, given the dimensional matrix presented in Table 2.3, the dimension of
variable v5 is d15 d22 d31 .

Table 2.3 Dimensional matrix

v1 v2 v3 v4 v5
d1 1 2 3 4 5
d2 2 4 3 0 2
d3 3 4 3 2 1

The dimensional matrix has Nd rows and Nv columns, where Nd is the number
of dimensions required to describe the problem and Nv is the number of relevant
variables necessary to describe the problem. It is assumed that they are not irrelevant
variables. Furthermore, it is also assumed that Nv ≥ Nd .
The dimensional matrix is partitioned into two submatrices, the M matrix and
the N matrix. The M matrix is square, of order Nd xNd , and non-singular. It is formed
by the rightmost Nd columns of the dimensional matrix. The N matrix is formed by
the first Nv − Nd columns of the dimensional matrix. Note that the partition may be
non-unique, although the fundamental requirement is that M matrix must be full rank.
For instance, one possible partition of the dimensional matrix given in Table 2.3 is
presented in Table 2.4. Note that the M matrix is full rank.
Conceptually, the variables related to the M matrix form the set of parame-
ters with independent dimensions introduced in the Buckingham pi theorem. The
 Dimensional analysis and dimensional similarity 17

Table 2.4 Partition of the

dimensional matrix

1 2 3 4 5
2 2 3 0 2
3 4 3 2 1
N matrix M matrix

full rank requirement of M matrix reflects the dimension independence of the Nd

variables. The variables related to the N matrix include the rest of variables.

2.4.2 The dimensional set

The dimensional set is a matrix representation that yields the dimensionless group
number in a compact and economic way. The dimensional set is formed by four
matrices, as presented in Table 2.5. N and M matrices are the dimensional matrix
previously defined. P matrix is the identity, of size (Nv − Nd )x(Nv − Nd ). The O
matrix is given by O = −P(M −1 N )T , and it provides the exponents to which the
Nv − Nd dimensions must be raised to obtain the dimensionless numbers.

Table 2.5 Dimensional set

Physical variables

Dimensions N matrix M matrix

Dimensionless var. P matrix O matrix

Example 2.2. Consider again the capacitor charge through a resistance as presented
in Figure 1.2. The dimensional matrix is given in Table 2.6.

Table 2.6 Capacitor charge circuit

dimensional matrix

v0 t vi r c
V 1 0 1 1 −1
A 0 0 0 −1 1
T 0 1 0 0 1

The dimensional matrix is partitioned as presented in Table 2.7, by choosing vi ,

r, and c as variables with independent dimensions.
18 Application of dimensional analysis in systems modeling and control design

Table 2.7 N and M matrices for the

capacitor charge circuit
dimensional matrix

v0 t vi r c
V 1 0 1 1 −1
A 0 0 0 −1 1
T 0 1 0 0 1

We recall that the partition is not unique. In fact v0 could be used instead of vi in the
M matrix. Even the set {t, v0 , r} defines an M matrix that is of full dimension.
The resulting dimensional set is shown in Table 2.8. The dimensional numbers
are given by matrices P and O. From the exponents of each variable we have

v0
= (2.22)
vi

and
t
1 = (2.23)
rc

that is the same result as in the previous example.

2.5 Dimensional similarity

The Buckingham pi theorem states that the behavior of a physical system may be
described in dimensionless form and is completely defined by a set of dimensionless
numbers. In this way, if two systems have the same dimensionless numbers, their
behavior must be correlated in some way, and the systems are dimensionally sim-
ilar. Whenever two systems are dimensionally similar, it is possible to extrapolate
the results in one system to the other. This is the fundamental idea of dimensional
modeling.

Table 2.8 Capacitor charge circuit

dimensional set

v0 t vi r c
V 1 0 1 1 −1
A 0 0 0 −1 1
T 0 1 0 0 1

 1 0 −1 0 0
1 0 1 0 −1 −1
 Dimensional analysis and dimensional similarity 19

In order to define dimensional similarity we consider two systems, the prototype

and the model, expressed by means of dimensionless numbers as presented in the
preceding section, that is

p = (p1 , . . . , p(Nv −Nd −1) ) (2.24)

m = (m1 , . . . , m(Nv −Nd −1) ) (2.25)

Conceptually the prototype can be understood as the system to be designed while the
model is a facility to be used for investigation and to extract information to be used
in the prototype.

Definition 2.1 (Szirtes and Rózsa [2], p. 473). Two systems (prototype and model)
are dimensionally similar if all the dimensionless variables describing the systems are
identical in construction and in magnitude, that is

p = m
p1 = m1
.. (2.26)
.
p(Nv −Nd −1) = m(Nv −Nd −1)

However, previous definition is redundant as Theorem 2.1 shows.

Theorem 2.1 (Szirtes and Rózsa [2], p. 473). If in two systems (prototype and model)
all the corresponding pairs of dimensionless variables containing only governing
variables are identical in construction and in magnitude, that is

p1 = m1
..
. (2.27)
p(Nv −Nd −1) = m(Nv −Nd −1)

then the dimensionless variables forming the remaining pair (containing the governed
variable, that is ({p = m }) are also identical.

The result easily follows from (2.24). The dimensional similarity definition between
systems is given by the following theorem.

Theorem 2.2. Assume that there is only one governed variable. Furthermore, the
governed variable appears in only one of the dimensionless variables, that is the
governed variable appears in the N matrix of the dimensional set. Then, if in two
systems (prototype and model) all the corresponding pairs of dimensionless variables
20 Application of dimensional analysis in systems modeling and control design

containing only governing variables are identical in construction and magnitude,

that is

p1 = m1
..
. (2.28)
p(Nv −Nd −1) = m(Nv −Nd −1)

then the prototype and model are dimensionally similar.

Finally, we define the concept of partial similarity.

Definition 2.2. If in two systems (prototype and model) some dimensionless num-
bers related to the governing variables are not equal, then the systems are partially
dimensionally similar.

Example 2.3. In the previous capacitor charge example, the dimensionless relation
obtained was

 = (1 ) (2.29)

with  = v0 /vi the governed variable dimensionless number and 1 = t/rc the gov-
erning variable dimensionless number. First note that there is only one governed
variable (i.e. v0 ) and that the governed variable appears in only one of the dimen-
sionless variables (i.e.  = v0 /vi ). In this case both assumptions are fulfilled and
two systems with identical 1 dimensionless numbers are dimensionally similar,
that is p1 = m1 with p1 = tp /rp cp and m1 = tm /rm cm .

2.5.1 Scale factors

Definition 2.3 (Szirtes and Rózsa [2], p. 479). The scale factor with respect to a
particular physical variable is the quotient of the magnitudes of that variable for the
prototype and its model.

It is important to note that there is not a unique scale factor between a model and a
prototype, but for each physical variable there is one scaling factor.

Example 2.4. Consider the variables involved in the 1 = t/rc dimensionless

number of the capacitor charge example. The dimensional variables involved are
t, r, and c. The scaling factors related with the previous dimensional variables are
the time scaling factor St = tp /tm , the resistance scaling factor Sr = rp /rm , and the
capacitance scaling factor Sc = cp /cm .
 Dimensional analysis and dimensional similarity 21

2.5.2 Model law

A model law is a set of relations among scaling factors that are derived from
dimensional similarity constraints.

Example 2.5. In the capacitor charge example, for dimensional similarity between
model and prototype, it must follow that p1 = m1 , that is

tp tm
= (2.30)
rp c p rm c m

which can be rearranged as

tp rp cp
= (2.31)
tm r m cm

and by definition of scaling factors it yields

St = Sr Sc (2.32)

which is the model law that must follow any capacitor charge circuit in order to be
dimensionally similar.

2.6 Exercises
1. Given a right triangle and considering the area a as the governed variable and the
hypotenuse h and the hypotenuse angle α as the governing variables:
a) What is the physical dimension and units of the hypotenuse angle α?
b) Can α appear in the M matrix?
c) Find one dimensional set and derive the dimensionless numbers.
d) Derive the dimensionless relation.
2. In the ceramic tile glazing process, the tile is covered by a uniform layer of glazing
material. The tile passes under a continuous cascade of liquid glaze. The variable
to be controlled is the quantity (mass) of glaze deposited per tile. It is assumed
that the glaze mass deposited ml with units [ml ] = kg/m depends on the glaze
cascade flow q, [q] = m3 /s, the tile velocity v, [v] = m/s, and the glaze density ρ,
[ρ] = kg/m3 .
a) Classify the variables as governing or governed variables.
b) Derive a dimensionless relation among the previous variables by direct
application of the Buckingham pi theorem.
c) Find one dimensional set and derive the dimensionless numbers that define
the ceramic tile glazing process.
d) Obtain the model law between a model and a prototype for dimensional
similarity.
22 Application of dimensional analysis in systems modeling and control design

3. In the previous ceramic tile glazing process, it is important to stress the dif-
ference between the glaze cascade flow q and the deposited flow in the tile qd .
It is assumed that the difference depends on the length of the tile lt and the
characteristic length of the cascade lc .
a) Construct a dimensional set if the governed variable is q.
b) If a model must be constructed, find the model law.
c) For dimensional similarity, is there any constraint between the cascade flow q
and the tile length lt ?

References
[1] G.I. Barenblatt. Dimensional Analysis. Gordon and Breach Science Publishers,
1987
[2] T. Szirtes and P. Rózsa. Applied Dimensional Analysis and Modeling.
Butterworth-Heinemann, 2007
 Chapter 3
Dynamical systems: dimensionless representation

3.1 Introduction
The dimensionless representation minimizes the number of parameters necessary
for the system description, and may provide further knowledge of the parameters
influence on the dynamical response. Furthermore, as proposed in Reference 7, the
number of dimensionless numbers is a measure of the complexity of the problem at
hand. Henceforth, the dimensionless representation may provide a system complex-
ity characterization that is different from mere system order. In fact, a small number
of dimensionless numbers imply that the dynamical system can be characterized
with a small number of experiments. For instance, in fluid dynamics problems dom-
inated by gravitational forces, only the Froud number is of relevance, so Euler and
Reynolds dimensionless numbers are not considered. On the contrary, when the sys-
tem dynamics depends on a large amount of relevant dimensionless numbers, the
characterization of the dynamical system requires a great number of experiments.
For example consider the plasma confinement problem. In this case all the hydro-
dynamic, electromagnetic, and nuclear dimensionless numbers are relevant in the
plasma confinement, making the dynamical behavior very complex [7].
Another property of the dimensionless representation is that it is the fingerprint
of dimensional similarity. In this way two dimensionally similar systems have the
same dimensionless representation. Consequently, the analysis and design results
over dimensionless models are applicable not only to one dynamical model, but also
to the infinite family of all dimensionally similar models.
In this chapter we show how to obtain the dimensionless form of dynamical
models represented in transfer function and in state space. The similarities and impli-
cations of dimensionless models are stated and compared with other dynamical system
transformations such as coordinate transformation and signal scaling. Finally, we
perform the dimensionless transformation of discrete time systems, showing its dif-
ferences with respect to continuous time models. The dimensionless representation of
dynamical systems forms the backbone for the applications of dimensional analysis
to control theory presented in the rest of the book.

3.2 Transfer function dimensionless representation

The transfer function dimensionless representation is based on direct application of
the Buckingham pi theorem. First, it is worth to remember that physical parameters
24 Application of dimensional analysis in systems modeling and control design

such as mass, length, capacitance, etc., are dumped into transfer function parameters
like gain, time constants, and time delay.
Let’s consider the application of the Buckingham pi theorem to a transfer func-
tion to find a dimensionless representation. The Buckingham pi theorem requires
first, to find a set of transfer function parameters with independent dimensions,
second, to find the dimensionless numbers, and finally to express the transfer function
in dimensionless form.

3.2.1 Transfer function parameters dimensions

Prior to find a set of transfer function parameters with independent dimensions, the
physical dimensions of transfer function parameters must be determined [1]. The
dimensions of the transfer function parameters are given by the following lemma.

Lemma 3.1. Consider the generic transfer function G(s) in pole-zero form
m
(τcj s + 1)
e−hs
j=1
G(s) = k n (3.1)
sl i=1 (τpi s + 1)

with transfer function parameters θG = k, h, τp1 , . . . , τpn , τc1 , . . . , τcm , l. The relative
system order is n + l − m ≥ 0. K is the gain dimension and T the time dimen-
sion. The dimensions of the transfer function parameters are [k] = K/(Tl ), [τpi ] =
[τcj ] = T for i = 1, . . . , n and j = 1, . . . , m, [h] = T, and [l] = 1, where 1 stands for
dimensionless quantity.
 
Proof. First we show that ni=1 (τpi s + 1) is dimensionless, thus [ ni=1 (τpi s + 1)]
= 1. The dimension of τpi for i = 1, . . . , n is time dimension [τpi ] = T. On the
other hand, the dimension of the s operator is [s] = T−1 , as can be seen either
in the time domain, where the s operator is transformed to operator d/dt, or in
the frequency domain, where s is transformed to jω. Consequently,each product
τpi s yields a dimensionless number. The same reasoning applies to mj=1 (τcj s + 1)
and e−hs . The parameter l is also dimensionless because exponents of transcen-
dental functions must be dimensionless for homogeneity [6, p. 104]. Consequently,
[G(s)] = [ks−l ] or equivalently, [k] = [G(s)][(sl )] that yields [k] = K/(Tl ).
As a result, only two physical dimensions are required to label the physical dimen-
sions of transfer function parameters, the gain dimension K and the time dimension T ,
thus Nd = 2.

3.2.2 Transfer function parameters with independent dimensions

Once the fundamental dimensional set is established and the dimensions of the trans-
fer function parameters determined, we can construct the dimensional matrix of
transfer function parameters, as presented in Table 3.1. It can be seen that all the
parameters can be divided into three sets, the parameters with gain and/or time
 Dynamical systems: dimensionless representation 25

dimension {G, k}, the parameters with time dimension only {τp1 , . . . , τpn , τc1 , . . . ,
τcm , h, s}, and the dimensionless parameter {l}.

Table 3.1 Dimensional matrix of transfer function parameters

G k τp1 … τpn τc1 … τcm h s l

K 1 1 0 … 0 0 … 0 0 0 0
T 0 −l 1 … 1 1 … 1 1 −1 0

By choosing k and τp1 as the set of parameters with independent dimensions,

the dimensional matrix can be partitioned as shown in Table 3.2. It can be seen that
k and τp1 are independent because matrix M has full rank, even if l = 0.
Note that choosing k and τp1 as parameters with independent dimension is arbi-
trary. However, the parameter l, which is already dimensionless, cannot be a column
of the M matrix, because in this case matrix M would be singular. In the same way,
any two parameters that depend on the time dimension have always a singular M
matrix, thus they have not independent dimensions. Finally, it is possible to consider
G and k as a set of parameters with independent dimensions because the M matrix
would be in this case
 
1 0
M= (3.2)
0 −l

and for l  = 0 it has full rank. However, l is the number of transfer function integra-
tors, and in the case of no integrators l = 0, hence the M matrix becomes singular.
Consequently in the general case it is not possible to choose the pair G and k as
parameters with independent dimensions.

Table 3.2 Partition of the dimensional matrix of transfer

function parameters

G τp2 ... τpn τc1 ... τcm h s l k τp1

K 1 0 ... 0 0 ... 0 0 0 0 1 0
T 0 1 ... 1 1 ... 1 1 −1 0 −l 1

3.2.3 Transfer function dimensionless numbers

We have established that the number of independent dimensions is 2, that is Nd = 2,
and a convenient set of transfer function parameters with independent dimensions is
{k, τp1 }. The next step, following the Buckingham pi theorem, is to obtain the Nv − Nd
dimensionless numbers. Taking as starting point the dimensional matrix partition of
Table 3.2, we construct the dimensional set presented in Table 3.3, with matrix P an
identity matrix of size (Nv − Nd )x(Nv − Nd ), and matrix O = −P(M −1 N )T .
26 Application of dimensional analysis in systems modeling and control design

Each row of the P and O matrices provide a dimensionless number i , i =

1, . . . , Nv − 2, as can be seen in Table 3.2. For instance 1 is given by
G(s)
1 = l
(3.3)
kτp1

Note that, by P matrix being the identity, there is one dimensionless number
for each one of the transfer function parameters with dependent dimensions {G,
τp2 , . . . , τpn , τc1 , . . . , τcm , h, s, l}. As a result, for a generic transfer function with inte-
grators and time delay, the quantity of dimensionless numbers is equal to Nv − Nd =
n + m + 4, where n is the number of poles and m the number of zeroes. The reduction
of transfer function dimensional parameters Nd is limited to Nd = 2.
The dimensionless transfer function parameters can be interpreted as follows:
● The dimensionless number 1 = G(s)
l
kτp1
is the dimensionless transfer function.
● The dimensionless numbers i = τpi /τp1 , i = 2, . . . , n and n+j = τcj /τp1 ,
j = 1, . . . , m, are the dimensionless time constants.
● The dimensionless number Nv −4 = h/τp1 is the dimensionless time delay.
● The dimensionless number Nv −3 = sτp1 is the dimensionless Laplace variable.
● The dimensionless number Nv −2 = l is the number of integrators, which is
already dimensionless.

Table 3.3 Transfer function dimensional set

G τp2 ... τpn τc1 ... τcm h s l k τp1

K 1 0 0 0 0 0 0 0 1 0
T 0 1 1 1 1 1 −1 0 −l 1

1 1 0 0 0 0 0 0 0 −1 −l
2 0 1 0 0 0 0 0 0 0 −1
.. .. ..
. . .
n 0 0 1 0 0 0 0 0 0 −1
n+1 0 0 0 1 0 0 0 0 0 −1
.. .. ..
. . .
n+m 0 0 0 0 1 0 0 0 0 −1
Nv −4 0 0 0 0 0 1 0 0 0 −1
Nv −3 0 0 0 0 0 0 1 0 0 1
Nv −2 0 0 0 0 0 0 0 1 0 0

3.2.4 Dimensionless transfer function

Once the transfer function dimensionless numbers are determined we provide a
straightforward methodology to represent a transfer function in a dimensionless form.
The procedure is as follows:
 Dynamical systems: dimensionless representation 27

(i) Choose the transfer function dominant time constant τp1 .

(ii) Divide each transfer function parameter with time dimension by τp1 .
(iii) Perform the following variable substitution s → s̄ := τp1 s. We obtain G(s̄).
l
(iv) Divide the resulting transfer function G(s̄) by kτp1 to obtain the dimensionless
transfer function Ḡ(s̄).
Steps (ii) and (iii) leave the dimensional transfer function unaltered, that is G(s) =
G(s̄). G(s̄) has dimension equal to [G(s̄)] = KTl , while transfer function Ḡ(s̄) is
dimensionless, that is [Ḡ(s̄)] = 1.

Example 3.1. Consider the second order system with time delay and τ1 ≥ τ2

ke−hs
G(s) = (3.4)
(τ1 s + 1)(τ2 s + 1)

The quantity of variables is Nv = 5, the quantity of dimensions is Nd = 2, and the

expected quantity of dimensionless numbers is Nv − Nd = 3. First, following step (i)
we choose the dominant time constant τ1 as normalising variable. After steps (ii)
and (iii) we have

ke−h̄s̄
G(s̄) = (3.5)
(s̄ + 1)(τ¯2 s̄ + 1)

with h̄ = τh1 , τ¯2 = τ2

τ1
, and s̄ = τ1 s. Finally, step (iv) provides the dimensionless
transfer function

G(s̄) e−h̄s̄
Ḡ(s̄) = = (3.6)
k (s̄ + 1)(τ¯2 s̄ + 1)

3.3 State space dimensionless representation

The development of modern control theory had its roots in the state space represen-
tation of dynamical systems and the concept of system state space [3]. The standard
linear time invariant (LTI) state space representation is given by

ẋ(t) = Ax(t) + Bu(t) (3.7)

y(t) = Cx(t) + Du(t) (3.8)

with state vector x(t) ∈ Rn , input vector u(t) ∈ Rp , output vector y(t) ∈ Rq , dynamic
matrix A ∈ Rn×n , input matrix B ∈ Rn×p , output matrix C ∈ Rq×n , and matrix
D ∈ Rq×p .
The first step, following Buckingham pi theorem, would be to determine the
dimensions of matrices A, B, C, and D of the state space representation. Although
in Reference 4 the dimensions of matrix A are determined, the determination of
28 Application of dimensional analysis in systems modeling and control design

the matrices dimensions is not as straightforward as in the transfer function case

because the matrix entries are combinations of physical parameters.
The approach taken to find the dimensionless representation of the state space
representation ((3.7)–(3.8)) is based on system transformation, as presented in
Reference 2. In this way, first we construct the following dimensionless groups:

x(t) = Nx x̄(t̄) (3.9)

u(t) = Nu ū(t̄) (3.10)
y(t) = Ny ȳ(t̄) (3.11)
t = Nt t̄ (3.12)

where Nx ∈ Rn×n , Nu ∈ Rp×p , Ny ∈ Rq×q , Nt ∈ R, and x̄(t̄) is the dimensionless

state space, ū(t̄) the dimensionless input, ȳ(t̄) the dimensionless output, and t̄ the
dimensionless time.
To obtain the dimensionless representation first note that

dx(t) d
= (Nx x̄(t̄))
dt dt
d x̄(t̄)
= Nx
dt
d x̄(t̄) d t̄
= Nx
d t̄ dt
˙
= Nx x̄Nt−1 (3.13)

and by direct substitution of ((3.9)–(3.13)) into ((3.7)–(3.8)), we have

˙ = ANx x̄(t) + BNu ū(t)

Nx Nt−1 x̄(t) (3.14)
Ny ȳ(t) = CNx x̄(t) + DNu ū(t) (3.15)

which can be rewritten in dimensionless form as

˙ t̄) = Āx̄(t̄) + B̄ū(t̄)

x̄( (3.16)
ȳ(t̄) = C̄ x̄(t̄) + D̄ū(t̄) (3.17)

with dimensionless matrices

Ā = Nt Nx−1 ANx
B̄ = Nt Nx−1 BNu
C̄ = Ny−1 CNx
D̄ = Ny−1 DNu
 Dynamical systems: dimensionless representation 29

Example 3.2. Consider the state space representation of an inverted pendulum

given by
⎡ ⎤ ⎡ ⎤
  c mgl − k   1
θ̈ − ⎦ θ̇ + ⎣ i ⎦ u
= ⎣ i i (3.18)
θ̇ θ
1 0 0
 

θ̇
y = 01 (3.19)
θ

with θ the angular position, θ̇ the angular velocity, u the input torque, g the gravity
acceleration, l the pendulum length, k the angular spring stiffness, c the friction
constant, m the pendulum mass, and i the pendulum inertia.
From the dimensional set of Table 3.4, we can obtain the dimensionless sig-
nals as
Table 3.4 Inverted pendulum dimensional set

θ θ̇ u t g l c i m k
M 0 0 1 0 0 0 1 1 1 1
L 0 0 2 0 1 1 2 2 0 2
T 0 −1 −2 1 −2 0 −1 0 0 −2

θ̄ 1 0 0 0 0 0 0 0 0 0
θ̄˙ 0 1 0 0 0 0 0 1/2 0 −1/2
ū 0 0 1 0 0 0 0 0 0 −1
t̄ 0 0 0 1 0 0 0 −1/2 0 1/2
ḡ 0 0 0 0 1 0 0 1/2 1/2 −1
l̄ 0 0 0 0 0 1 0 −1/2 1/2 0
c̄ 0 0 0 0 0 0 1 −1/2 0 −1/2

θ̄ = θ

i
θ̄˙ = θ̇
k
1
ū = u
k

k
t̄ = t
i
The resulting dimensionless matrices are thus easily obtained as
⎡ ⎤
k
0⎦
Nx = ⎣ i
0 1

Nu = k
30 Application of dimensional analysis in systems modeling and control design

Ny = 1
 
i
Nt =
k

That finally yields the dimensionless representation of the state space inverted
pendulum
    ˙  
θ̄¨ −c̄ ḡ l̄ − 1 θ̄ 1
= + ū (3.20)
θ̄˙ 1 0 θ̄ 0
 ˙

ȳ = 0 1 θ̄ (3.21)
θ̄

with
√
im
ḡ = g
k

m
l̄ = l
i

1
c̄ = c
ik

Note that all the dimensionless variables involved {θ̄ , θ̄˙ , ū, t̄, ḡ, l̄, c̄} are obtained as
the rows of the lower submatrices in the dimensional set (Table 3.4).

3.3.1 Interpretation of the state space dimensionless transformation

The state space model ((3.7)–(3.8)) is given in dimensionless form by the following
transformation:

Ā = Nt Nx−1 ANx (3.22)

B̄ = Nt Nx−1 BNu (3.23)
C̄ = Ny−1 CNx (3.24)
D̄ = Ny−1 DNu (3.25)

The transformations to obtain a dimensionless model can be interpreted in terms of

a coordinate transformation, a signal scaling, and a time scaling. The state space
vector x of a state space model can be transformed by a nonsingular matrix T , giv-
ing a new state space representation xT , that is x = TxT [3]. This transformation
is also known as similar transformation. However, to avoid confusion with dimen-
sional analysis similarity concepts, we will refer it as coordinate transformation.
 Dynamical systems: dimensionless representation 31

The coordinate transformation provides a new state space representation given by

matrices:

AT = T −1 AT (3.26)
−1
BT = T B (3.27)
CT = CT (3.28)
DT = D (3.29)

Another system transformation is the signal scaling that is motivated by a

normalization of signal values. In this way, the input and output signals are scaled as
y = Dy ys and u = Du us , with Dy and Du diagonal matrices that include the inverse of
the maximum expected value of y and u, respectively [5]. The state space matrices
after a signal scaling are given by

AS = A (3.30)
BS = BDu (3.31)
CS = Dy−1 C (3.32)
DS = Dy−1 DDu (3.33)

By comparison of (3.22)–(3.25) with (3.26)–(3.29) and (3.30)–(3.33), it can

be seen that the state space dimensionless transformation includes both a coor-
dinate transformation and a signal transformation. Moreover, the dimensionless
transformation also includes a time scaling by means of factor Nt . Note that neither
coordinate transformation nor signal scaling incorporates a timescaling.
Certain system properties are invariant under coordinate transformation and
signal scaling, for instance the characteristic polynomial, the system output, and
the system transfer function, among others. In Table 3.5 it can be seen how these
properties are invariant under coordinate transformation and signal scaling. How-
ever, under dimensionless transformation, due to time scaling, we see that the
characteristic polynomial is scaled by the time factor Nt , that is |Nt ||s̄I − A|, with
s̄ = s/Nt . In the same way, the output is scaled by the time variable and given by

Table 3.5 Characteristic polynomial (Char. Pol), System output (Output), and
transfer function (T.F.) for coordinate transformation, signal scaling, and
dimensionless transformation. Note that s̄ = s/Nt and t = Nt t̄

Coordinate transformation Signal scaling Dimensionless

transformation

Char. Pol. |sI − A| |sI − A| |Nt ||s̄I − A|

Output y(t) = Cx(t) + Du(t) y(t) = Cx(t) + Du(t) y(t̄) = Cx(t̄) + Du(t̄)
T.F. C(sI − A)−1 B + D C(sI − A)−1 B + D Nt2 (C(s̄I − A)−1 B + D)
32 Application of dimensional analysis in systems modeling and control design

y(t̄) = Cx(t̄) + Du(t̄), with t = Nt t̄. Finally, the transfer function is also time scaled
by Nt2 (C(s̄I − A)−1 B + D), with s̄ = s/Nt .

3.4 Comparison between transfer function and state space

dimensionless representation

The transfer function dimensionless representation is obtained from transfer function

parameters such as poles and zeros time constants, gain, time delay, and num-
ber of integrators. On the contrary, the state space dimensionless representation is
obtained from physical parameters such as capacitance and electrical resistance. The
physical parameters of state space representation are dumped into transfer function
parameters. In this way the dimensionless representation of transfer function loses
physical insight. However, the obtention of dimensionless representation of transfer
functions is straightforward and of great generality. In fact, the procedure for dimen-
sionless representation of transfer functions is not dependent on the physical problem
at hand.
The use of the transfer function representation or the state space representation
in control theory depends on the type of problems to be solved. For instance, model
reference adaptive control and process control are traditionally tackled under the
transfer function formalism, whereas modern control theory concepts as controllabil-
ity and observability arise in state space models. Their differences are shown by the
following example.

Example 3.3. Consider the mass-spring-damper dynamical system of Figure 3.1

with initial conditions x(0) = x0 , ẋ(0) = 0, input F(t), and output x(t). The output is
given by

1 sm + c
X (s) = F(s) + 2 x0 (3.34)
ms2 + cs + k ms + cs + k

that is represented in gain and time constants form as

k1 k2 (τ3 s + 1)
X (s) = F(s) + (3.35)
(τ1 s + 1)(τ2 s + 1) (τ1 s + 1)(τ2 s + 1)

f
c
m

Figure 3.1 Mass-spring-damper dynamical system

 Dynamical systems: dimensionless representation 33

with transfer function parameters

1
k1 = (3.36)
k
cx0
k2 = (3.37)
k
2m
τ1 =  (3.38)
(c + (c2 − 4km))
2m
τ2 =  (3.39)
(c − (c2 − 4km))
m
τ3 = (3.40)
c
The dimensionless transfer function is easily obtained as

X̄F (s) XF (s) 1

:= = (3.41)
F(s) k1 F(s) (s̄ + 1)(τ̄2 s̄ + 1)

with s̄ = τ1 s and τ̄2 = τ2 /τ1 . The response due to initial conditions in dimensionless
form is
Xx0 (s) (τ̄3 s̄ + 1)
X̄x0 (s) := = (3.42)
k2 (s̄ + 1)(τ̄2 s̄ + 1)

with τ̄3 = τ3 /τ1 . The dimensionless system output is given by

X̄ (s) = X̄F (s) + X̄x0 (s) (3.43)

that is a function of the following dimensionless parameters, s̄, τ̄2 , and τ̄3 . Note that
the dimensionless time is given as t̄ = t/τ1 .
A state space model of system in Figure 3.1 is given by
⎡ ⎤ ⎡ ⎤
  c k   1
ẍ ⎣ − − ⎦ ẋ
= m m ⎣
+ m⎦ f (3.44)
ẋ x
1 0 0
 

ẋ
y = 0 1 (3.45)
x

From the dimensional set of Table 3.6, we can obtain the dimensionless sig-
nals as

k
t̄ = t
m

ẋ m
x̄˙ =
x0 k
34 Application of dimensional analysis in systems modeling and control design

Table 3.6 Transfer function dimensional set

t ẋ x c f x0 k m
M 0 0 0 1 1 0 1 1
L 0 1 1 0 1 1 0 0
T 1 −1 0 −1 −2 0 −2 0

t̄ 1 0 0 0 0 0 1/2 −1/2
x̄˙ 0 1 0 0 0 −1 −1/2 1/2
x̄ 0 0 1 0 0 −1 0 0
c̄ 0 0 0 1 0 0 −1/2 −1/2
f¯ 0 0 0 0 1 −1 −1 0

x
x̄ =
x0

1
c̄ = c
mk
1
f¯ = f
x0 k

The resulting dimensionless matrices are

⎡  ⎤
k
Nx = ⎣x0 m 0⎦
0 x0

Nu = x0 k

Ny = x0
 
m
Nt =
k

That finally yields the dimensionless representation of the state space inverted
pendulum
      
x̄¨ −c̄ −1 x̄˙ 1
= + ū (3.46)
x̄˙ 1 0 x̄ 0
 

x̄˙
ȳ = 0 1 (3.47)
x̄

Summarizing, the transfer function dimensionless representation relates the

dimensional input and dimensional output divided by the transfer function gain
as can be seen in (3.41) and (3.42). Furthermore, the time scale is given by the
dominant time constant as t̄ = t/τ1 .
 Dynamical systems: dimensionless representation 35

In the state space dimensionless representation all signals are dimensionless and
the time scale depends on the physical parameters of the problem at hand. Note,
however, that just one dimensionless parameter is necessary for state space model c̄
while two parameters, τ̄2 and τ̄3 , are required for the transfer function approach.

3.5 Discrete time models dimensionless representation

Heretofore, we have considered only dimensionless representations of continuous
time models. However, discrete time models are prevalent in control system appli-
cations. Therefore, in this section we derive the dimensionless representation of
discrete time dynamical systems in the transfer function framework and in the state
space representation.

3.5.1 Discrete time transfer function dimensionless representation

The dimensions of the discrete time transfer function parameters are given by the
following lemma.

Lemma 3.2. Consider a generic discrete time transfer function G(z) in polynomial
form
m −j −q
j=0 (bj z )z
G(z) = k  n  (3.48)
1 + i=1 (ai z −i )

with q integer and transfer function parameters k, bj , and ai . The relative system
order is n + q − m ≥ 0. The dimensions of the transfer function parameters are
[k] = K, [aj ] = 1, and [bi ] = 1, where K is the gain dimension and 1 is a dimension-
less quantity.

Proof. Consider the difference equation of transfer function (3.48)

y(k) = −a1 y(k − 1) − · · · − an y(k − n)
+ k(b0 u(k − q) + · · · + bm u(k − m − q))
by dimensional homogeneity considerations [6], the dimensions of [ai y(k − i)] must
be equal to [ y(k)], then [ai y(k − i)] = Y . It then follows that [ai ] = 1. In a similar
way [kbj u(k − j − q)] = Y , being Y the output dimension. Note also that [u(k − j −
q)] = U , being U the input dimension, then [kbj ] = YU −1 = K. As a result, a possible
dimensional decomposition of the parameter kbj into a gain k and a parameter bj is
[k] = K and [bj ] = 1.
The number of transfer function variables is Nv = n + m + 3, and the num-
ber of variables with independent dimensions is Nd = 1. As a result there are n =
m + n + 2 dimensionless numbers, which are
● The dimensionless number G(z) = G(z)
k
is the dimensionless transfer function.
● The dimensionless number q = q.
36 Application of dimensional analysis in systems modeling and control design

● The dimensionless numbers bj = bj j ∈ [1, . . . , m].

● The dimensionless numbers ai = ai i ∈ [1, . . . , n].
The dimensionless discrete transfer function is given by
m −j −q
G(z) j=0 (bj z z )
=  n  (3.49)
k 1 + i=1 (ai z −i )

As a result, a discrete transfer function is made dimensionless by just dividing by k.

3.5.2 Discrete time state space model dimensionless representation

Consider the following discrete time state space model.

xk+1 = Ad xk + Bd uk (3.50)
yk = Cd xk + Dd uk (3.51)

In this case the dimensionless discrete state space model is given in a similar manner
as in the continuous time case. However, in the discrete time case, despite its name,
there is no time, just a sequence of samples. As a result, there is no time scaling
factor as the dimensions of xk+1 are equal to the dimensions of xk because there is no
time derivative, just a sequence index increase. The dimensionless discrete time state
space model is given by

A¯d = Nx−1 Ad Nx (3.52)

B¯d = Nx−1 Bd Nu (3.53)
C¯d = Ny−1 Cd Nx (3.54)
D̄d = Ny−1 Dd Nu (3.55)

with

xk = Nx x̄k (3.56)
uk = Nu ūk (3.57)
yk = Ny ȳk (3.58)

In the discrete time case, the dimensionless transformation is equal to a coordinate

transformation plus a signal scaling, but there is no time scaling.

Example 3.4. Consider the dynamics of a rabbit–wolf population, given by the

following state space model:
      
rk+1 1 + br −dr rk 0
= + u (3.59)
wk+1 bw 1 − d w wk 1
 

rk
y = 0 1 (3.60)
wk
 Dynamical systems: dimensionless representation 37

where rk is the rabbit population at instant k, wk is the wolf population at instant k,

br is the rabbit birth rate, bw is the wolf birth rate, dr is the rabbit death rate, and
dw is the wolf death rate. The control action u is the number of wolves set free at each
instant k. w0 is the initial number of wolves.

Table 3.7 Transfer function dimensional set

rk wk u br dw dr w0 bw
Qr 1 0 0 0 0 1 0 −1
Qw 0 1 1 0 0 −1 1 1

r̄k 1 0 0 0 0 0 −1 1
w̄k 0 1 0 0 0 0 −1 0
ū 0 0 1 0 0 0 −1 0
b̄r 0 0 0 1 0 0 0 0
d̄w 0 0 0 0 1 0 0 0
d̄r 0 0 0 0 0 1 0 1

First, from the dimensional set of Table 3.7, we can obtain the dimensionless
signals as
bw
r̄k = rk
w0
1
w̄k = wk
w0
1
ū = u
w0
b̄r = br
d̄w = dw
d̄r = dr bw

The resulting dimensionless matrices are

⎡  ⎤
w0
x 0⎦
Nx = ⎣ 0 bw
0 w0

Nu = w0

N y = w0

That finally yields the dimensionless representation

      
r̄k+1 1 + br −d̄r r̄k 0
= + ū (3.61)
w̄k+1 1 1 − dw w̄k 1
 

r̄k
ȳ = 0 1 (3.62)
w̄k
38 Application of dimensional analysis in systems modeling and control design

3.6 Exercises
1. Consider the following transfer functions commonly used in process control:
k
G1 (s) = ehs
(τ1 s + 1)
k
G2 (s) =
(τ1 s + 1)(τ2 s + 1)
k
G3 (s) = ehs
(τ1 s + 1)(τ2 s + 1)
k(τ 3s + 1)
G4 (s) =
(τ1 s + 1)(τ2 s + 1)
k(τ 3s + 1)
G5 (s) = ehs
(τ1 s + 1)(τ2 s + 1)

a)
Calculate the number of dimensionless number for each transfer function.
b)
Considering the number of dimensionless numbers a measure of complexity,
classify the transfer functions in terms of complexity.
c) Obtain the dimensionless representation of the transfer functions.
2. The state space model of an active band-pass filter is
⎡ −1 ⎤ ⎡ 1 ⎤
  0
dx ⎢ r 1 c1 ⎥
⎢ r1 c 1 ⎥
=⎢ ⎣
⎥ x +⎢
⎣ −1 ⎦ u
⎥ (3.63)
dt 1 −1 ⎦
r 1 c2 r 2 c 2 r1 c 2

y = 0 1 x (3.64)
with r1 , r2 electrical resistances, c1 , c2 capacitances, and the state x and u
voltages.
a) Obtain the dimensionless state space form.
b) Obtain the transfer function as G(s) = C(sI − A)−1 B.
c) Calculate the dimensionless transfer function representation.
d) Simulate the output for a step input for the dimensionless state space model
and for the dimensionless transfer function and compare the time scales of
each dimensionless model.

References
[1] P. Balaguer, A. Ibeas, C. Pedret, and S. Alcántara. ‘Controller parameters
dependence on model information through dimensional analysis’. In Decision
and Control, 2009 held jointly with the 2009 28th Chinese Control Confer-
ence. CDC/CCC 2009. Proceedings of the 48th IEEE Conference on, pages
1914–1919. IEEE, 2009
 Dynamical systems: dimensionless representation 39

[2] S.N. Brennan. On size and control: the use of dimensional analysis in con-
troller design. PhD thesis, Citeseer, 2002
[3] K. Ogata. Modern Control Engineering. Prentice Hall PTR, 2001
[4] H.J. Palanthandalam-Madapusi, D.S. Bernstein, and R. Venugopal. Dimen-
sional analysis of matrices state-space models and dimensionless units [lecture
notes]. Control Systems Magazine, IEEE, 27(6):100–109, 2007
[5] S. Skogestad and I. Postlethwaite. Multivariable Feedback Control. Analysis
and Design. Wiley, 1996
[6] T. Szirtes and P. Rózsa. Applied Dimensional Analysis and Modeling.
Butterworth-Heinemann, 2007
[7] B. West. ‘The tyranny of many dimensionless constants: a constraint on knowa-
bility’. Uncertainty and Surprise in Complex Systems, Springer, pages 87–92,
2005
 Chapter 4
Dynamical systems: dimensional similarity

4.1 Introduction
The dimensionless representation of dynamical systems presented in the previous
chapter is the fingerprint for establishing dimensional similarity among dynami-
cal systems. The concept of dimensional similarity is of great utility in model
experimentation as it permits to extrapolate the model experimental results to the pro-
totype. Furthermore, even the theoretical results obtained in dimensionless systems
can be easily extrapolated to dimensionally similar models [2].
In this chapter we present the conditions for establishing dimensional simi-
larity between systems represented in transfer function form and in state space
form. It is easier to establish dimensional similarity between transfer functions than
between state space systems representation. The transfer function framework only
requires two scaling factors between model and prototype, the gain scaling factor
and the time scaling factor. On the contrary, in the state space framework, there are
as many scaling factors as physical variables involved in the matrices representa-
tion, thus the scaling factors are problem dependent.
The second half of the chapter is devoted to dimensional similarity between
discrete time systems, considering also the transfer function and the state space
representation. With discrete time systems it is necessary to stress between purely
discrete time systems and the discrete time systems obtained from continuous time
systems by discretization. For purely discrete time systems there is no time, and
dimensional similarity is reduced to model equality up to a gain factor. For sam-
pled data systems, the time concept is recovered and for preserving continuous time
dimensional similarity the sampling times used must also be scaled properly.

4.2 Continuous time dynamical systems similarity

4.2.1 Transfer function dimensional similarity

The definition of transfer function dimensional similarity follows directly from its
dimensionless representation presented in Chapter 3.
42 Application of dimensional analysis in systems modeling and control design

Definition 4.1. Two transfer functions, the prototype Gp (sp ) and the model Gm (sm ),
are similar if the following dimensionless numbers are equal:
   
τpi τpi
= , i = 2, . . . , n, (4.1)
τp1 p τp1 m
   
τcj τcj
= , j = 1, . . . , m (4.2)
τp1 p τp1 m
   
h h
= (4.3)
τp1 p τp1 m
   
τp1 s p = τp1 s m (4.4)
(l)p = (l)m (4.5)

with τp1 the dominant time constant, (  )p stands for the prototype and (  )m stands
for the model.

The same prototype and model transfer function structure is a necessary but
not sufficient condition for transfer function similarity, because dimensionless num-
bers must be equal in construction and in magnitude. As a result, in the rest of the
chapter we assume that (l)p = (l)m , that is the numbers of integrators in prototype
and model are equal.
Recall that the scaling factor of a particular physical variable is defined as
the quotient of the prototype and the model physical magnitude. Given a proto-
type and a model transfer function we have, for instance, the gain scaling factor
Sk := (k)p /(k)m , associated with the transfer function gain, or, the time delay scaling
factor Sh := (h)p /(h)m , associated with the transfer function time delay. For all the
remaining parameters we could consider their corresponding scaling factor. How-
ever, we show next that when two transfer functions are similar, there are only two
scaling factors, the time scaling factor St and the gain scaling factor Sk . First we need
the following lemma.

Lemma 4.1. The scaling factor associated with any transfer function parameter
with time dimension is equal to St := (τp1 )p /(τp1 )m , where (τp1 )p and (τp1 )m are the
prototype and model dominant time constant, respectively.

Proof. First consider the time delay scaling factor Sh . From (4.3) it follows that

(h)p (τp1 )p
Sh := = (4.6)
(h)m (τp1 )m

The same reasoning can be applied to all the time constants presented in (4.1) and
(4.2). Finally the dimensionless number involving frequency is given by τp1 s. From
(4.4) it can be derived that
 Dynamical systems: dimensional similarity 43

(s)p (τp1 )m
Ss := = (4.7)
(s)m (τp1 )p

or in the time domain

tp (τp1 )p
St := = (4.8)
tm (τp1 )m

As a result the time scaling factor St is the same and equal to (τp1 )p /(τp1 )m for every
transfer function parameter with time dimension.

Recall now that if all dimensionless numbers related to governing variables

are equal, then the dimensionless numbers related to the governed variables of
prototype and model are also equal, that is
   
G(s) G(s)
l
= l
(4.9)
kp τp1 kp τp1
p m

which can be rewritten as

 l 
kp τp1 p
(G(s))p =  l  (G(s))m (4.10)
kp τp1 m

and by definition of the gain and time scaling factors it yields

(G(s))p = Sk Stl (G(s))m (4.11)

that provides the relation between dimensionally similar transfer functions in terms
of gain and time scaling factors. Furthermore, the transformations between the
model transfer function Gm (sm ) := (G(s))m , the prototype transfer function Gp (sp ) :=
(G(s))p and the dimensionless transfer function Ḡ(s̄) are summarized in Figure 4.1.
For instance, the prototype transfer function is obtained from the model transfer
function by just scaling with Sk Stl and the change of variable sm → St sp .

Example 4.1. As an example, consider two first-order plus dead-time (FOPDT)

transfer functions:

kp
Gp (sp ) = ehp sp (4.12)
(τp + 1)
km
Gm (sm ) = ehm sm (4.13)
(τm + 1)

where sp and sm are the Laplace variables of prototype and model, respectively.
Gp (sp ) and Gm (sm ) are dimensionally similar if hp /τp = hm /τm and τp sp = τm sm .
44 Application of dimensional analysis in systems modeling and control design

Sk Slt
sm → Stsp

Gm(sm) Gp(sp)

Sk−1St−l
sp → S −1
t sm

s̄ → (tp1)msm ¯s → (tp1)psp
sm → s̄
(tp1)m
km(tp1)ml kp(tp1)pl
1 1 sp → s̄
(tp1)p
km (t ) l 1 1
p1 m kp (t ) l
p1 p

¯ (s)
G ¯

Figure 4.1 Transformations between model Gm (sm ), prototype Gp (sp ) and

dimensionless transfer function Ḡ(s̄). The dimensionless form is
obtained by gain scaling and time scaling. The relation between model
and prototype is defined by means of two scaling factors, the gain
scaling factor Sk and the time scaling factor St

In what follows s̄ refers to either τp sp or τm sm . It follows that Gp (s̄)/kp = Gm (s̄)/km ,

or equivalently

Gp (s̄) = Sk Gm (s̄) (4.14)

with Sk = (k)p /(k)m , the gain scaling factor.

4.2.2 State space dimensional similarity

The dimensionless representation of dynamical systems in state space form is given as

˙ t̄) = Āx̄(t̄) + B̄ū(t̄)

x̄( (4.15)
ȳ(t̄) = C̄ x̄(t̄) + D̄ū(t̄) (4.16)

where the dimensionless matrices Ā, B̄, C̄, and D̄ are functions of dimensionless
numbers related to governing variables. In this way we can write

Ā = Ā(A1 , . . . , An )
B̄ = B̄(B1 , . . . , Bm )
C̄ = C̄(C1 , . . . , Cp )
D̄ = D̄(D1 , . . . , Dq )
 Dynamical systems: dimensional similarity 45

where there are n + m + p + q dimensionless numbers related to governing vari-

ables. In this case, there are as many scaling factors as governing variables, and
they are problem dependent. Dimensional similarity is established if

(Ā)p = (Ā)m (4.17)

(B̄)p = (B̄)m (4.18)
(C̄)p = (C̄)m (4.19)
(D̄)p = (D̄)m (4.20)

Example 4.2. Consider again the dimensionless representation of the state space
inverted pendulum




θ̄¨ −c̄ ḡ l̄ − 1 θ̄˙ 1
= + ū (4.21)
θ̄˙ 1 0 θ̄ 0


 θ̄˙
ȳ = 0 1 (4.22)
θ̄

with
√
im
ḡ = g
k

m
l̄ = l
i

1
c̄ = c
ik
We have one scaling factor per governing variable, that is a gravity acceleration
scaling factor Sg , a length scaling factor Sl , a damping scaling factor Sd , a mass
scaling factor Sm , an inertia scaling factor Si , and a stiffness scaling factor Sk .
In this case, for dimensional similarity it must follow that
     
1 1
c = c (4.23)
ik ik
p m
 √    √  
im m im m
g l = g l (4.24)
k i k i
p m

that is equivalent to the following model laws:


Sc = Si Sk (4.25)
Sk
Sl = (4.26)
Sm Sg
46 Application of dimensional analysis in systems modeling and control design

Summing up, to establish dimensional similarity between prototype and model

in the state space framework is more complicated than the transfer function case
because prototype and models are related by more than two scaling factors.

4.3 Discrete time dynamical system similarity

In this section we establish the similarity properties of discrete transfer functions and
sampled data transfer functions discretized from similar continuous time systems.

4.3.1 Discrete time transfer function similarity

Definition 4.2. Two discrete time transfer functions, the prototype Gp (z) and the
model Gm (z), are dimensionally similar if the following dimensionless numbers are
equal:
   
G(z) G(z)
= (4.27)
k p k m

(q)p = (q)m (4.28)

(bj )p = (bj )m , j = 1, . . . , m (4.29)
(ai )p = (ai )m , i = 1, . . . , n (4.30)

In this case dimensional similarity requires equality between the prototype

and model transfer function, except for a gain constant. As a result there is just
one scaling factor, the gain scaling factor Sk := (k)p /(k)m . It then follows that
under dimensional similarity model and prototype are related by

Gp (z) = Sk Gm (z) (4.31)

4.3.2 Sampled-data transfer function similarity

A sampled-data transfer function G(zSD ) is a discrete time transfer function
obtained by discretization of a continuous time transfer function G(s). Whereas
the discrete time transfer function variable z is related to the shift operator with
no time relation whatsoever, the sampled data transfer function variable zSD is
related with the continuous time variable by means of z = eTs s , where Ts is the
sampling time [1].

Definition 4.3. Two sampled data transfer functions Gp (zSD ) and Gm (zSD ), dis-
cretized with sampling time (Ts )p and (Ts )m , are dimensionally similar if (i) Gp (zSD )
and Gm (zSD ) are discrete time dimensionally similar and (ii) (esTs )p = (esTs )m .

Note that sampled data similarity implies discrete time similarity, but the converse
is not true. The requirements in the discretization process in order to yield similar
sampled-data transfer functions are derived next.
 Dynamical systems: dimensional similarity 47

Theorem 4.1. Consider two continuous time transfer functions Gp (s) and Gm (s) with
transfer function time scaling factor St , discretized using sampling times (Ts )p and
(Ts )m , with (Ts )p = St (Ts )m , yielding the sampled data transfer functions Gp (zSD )
and Gm (zSD ), respectively, then if Gp (s) and Gm (s) are continuous time dimensionally
similar, then Gp (zSD ) and Gm (zSD ) are sampled-data dimensionally similar.

Proof. By Definition 4.2, two discrete time transfer functions are dimensionally sim-
ilar if they have the same dimensionless numbers, that is (q)p = (q)m ; (bj )p = (bj )m ,
j = 0, . . . , m; and (ai )p = (ai )m , i = 1, . . . , n. The parameters bj and ai are obtained by
the discretization process , which can be, for instance, the zero order hold method
or Tustin method. The discretization depends on the continuous time transfer function
parameters and on the sampling time Ts , thus

ai = (τp1 , . . . , τpn , τc1 , . . . , τcm , h, l, Ts ) (4.32)

which has the following dimensionless form:

 
τp2 τpn τc1 τcm h l Ts
¯
ai =  ,..., , ,..., , , , (4.33)
τp1 τp1 τp1 τp1 τp1 τp1 τp1

If Gp (s) and Gm (s) are continuous time dimensionally similar and the sam-
pling time fulfils the condition (Ts )p = St (Ts )m , then all the dimensionless numbers
in (4.33) are equal for the prototype and the model, yielding (ai )m = (ai )p . The equal-
ity (bj )m = (bj )p is established by similar reasoning. The discrete time delay is related
with the continuous time delay and the sampling time by q = h/Ts . In order to have
h (T )
(q)p = (q)m it must follow that hmp = (Tss)mp , which is true by continuous time simi-
larity. Finally, for achieving sampled-data similarity it must also hold the equality
in the dimensionless number zSD , that is (zSD )m = (zSD )p . The equality follows by
recalling the relation existing between the s transform and the z transform, given
by z = esTs , and the assumption of continuous time similarity, which implies that if
es(Ts )p = es(Ts )m then (zSD )p = (zSD )m .

Example 4.3. Consider the two continuous time transfer functions presented in
the first column of Table 4.1. They are dimensionally similar with similarity scale
factors Sk = 10 and St = 0.8. Both transfer functions are discretized by ZOH with
sampling times (Ts )m = 0.1 and (Ts )p = 0.08, respectively. It can be seen in the
second column of Table 4.1 that they are sampled-data dimensionally similar.

4.3.3 Discrete state space similarity

The dimensionless representation of a discrete state space model is

x̄k+1 = Ād x̄k + B̄d ūk (4.34)

ȳk = C̄d x̄k + D̄d ūk (4.35)
48 Application of dimensional analysis in systems modeling and control design

Table 4.1 Example of sampled data transfer function similarity

Similar continuous time Similar discrete time

0.3s2 + 30.01s + 1 0.1617z 2 − 0.06392z − 0.09693

Gm
s3 + 3s2 + 3s + 1 z 3 − 2.715z 2 + 2.456z − 0.7408
0.192s2 + 24.01s + 1 0.1617z 2 − 0.06392z − 0.09693
Gp 10 10
0.512s3 + 1.92s2 + 2.4s + 1 z 3 − 2.715z 2 + 2.456z − 0.7408

In the same way as in the continuous time case, the dimensionless matrices Ād ,
B̄d , C̄d , and D̄d are functions of dimensionless numbers related to governing vari-
ables, thus

Ād = Ād (Ad 1 , . . . , Ad n )

B̄d = B̄d (Bd 1 , . . . , Bd m )
C̄d = C̄d (Cd 1 , . . . , Cd p )
D̄d = D̄d (Dd 1 , . . . , Dd q )

where there are n + m + p + q dimensionless numbers related with governing

variables. Dimensional similarity is established if the dimensionless matrices of
model and prototype are equal, that is

(Ād )p = (Ād )m (4.36)

(B̄d )p = (B̄d )m (4.37)
(C̄d )p = (C̄d )m (4.38)
(D̄d )p = (D̄d )m (4.39)

In this case, there are as many scaling factors as governing variables, and they are
also problem dependent.

4.4 Exercises
1. Given the following prototype and model transfer functions:
2(2s − 1)
Gm (s) = e1.75s (4.40)
(5s + 1)(3s + 1)(s + 1)
7(1.5s − 1)
Gp (s) = e1.5s (4.41)
(3.75s + 1)(2.25s + 1)(0.75s + 1)
a) Determine if they are dimensionally similar.
b) Modify the model time delay in order to achieve complete similarity.
 Dynamical systems: dimensional similarity 49

c)Find the relationship between sampling times that preserve dimensional

similarity.
d) Obtain the zero order hold discretization of continuous time transfer func-
tions with previous sampling times ratio and check the equality of discrete
time transfer functions up to a gain factor.
2. The state space representation of an inverted pendulum is
⎡ ⎤ ⎡ ⎤

 c mgl − k
 1
θ̈ − ⎦ θ̇ + ⎣ i ⎦ u
= ⎣ i i (4.42)
θ̇ θ
1 0 0


 θ̇
y = 0 1 (4.43)
θ
with θ the angular position, θ̇ the angular velocity, u the input torque, g the
gravity acceleration, l the pendulum length, k the angular spring stiffness, c the
friction constant, m the pendulum mass, and i the pendulum inertia.
a) Obtain a dimensionless description and derive the model laws.
b) Repeat the exercise in case the inertia is given by i = ml 2 .

References
[1] G.F. Franklin, J.D. Powell, and M.L. Workman. Digital Control of Dynamic
Systems. Prentice Hall, 1997
[2] M. Zlokarnik. Scale-up in Chemical Engineering. Wiley-VCH, 2002
 Chapter 5
Dimensionless systems identification and
model order reduction
with J. Romero, V. Alfaro and O. Arrieta

5.1 Introduction

The identification of models from experimental data is the first step in model-based
control design. Furthermore, when the model is of high order, it is normal to reduce
the model order for a simpler controller design. Low order identified models are the
fundamental elements for the design of practical control systems.
Although there are well-established techniques of general application for model
identification and model order reduction, such as minimum squares and balanced
model reduction, they are based on numerical calculations and give no parameter
value insight. Particularized algorithms for system identification and model reduc-
tion can provide further understanding on the problem at hand based on the relative
value among parameters.
The development of system identification and model reduction algorithms faces
two difficulties. First the relations between transfer function parameters and time
domain data are complex, and it is not straightforward to obtain model param-
eters from data, especially when the parameters number increases. For instance,
approaches for the identification of integral second order systems showing inverse
response have been based on the solution of a set of nonlinear equations [6] or the
numerical evaluation of an integral [11]. The second difficulty faced is to evalu-
ate what parameters can be dismissed without compromising the model accuracy.
Normally, it is not correct to dismiss large or small parameters because, due to the
Buckingham pi theorem, what influences the model behavior is not their absolute
value but the relative value among parameters. In the following section we show
how dimensional analysis helps to circumvent these problems.

5.2 General procedure

The dimensionless representation of dynamical systems reduces the number of
parameters of the dynamical system. In this way we may gain some simplicity on
the system representation. The parameters reduction by the dimensionless transfer
52 Application of dimensional analysis in systems modeling and control design

function representation is limited to two parameters. Despite the limit on the

parameters reduction capabilities, the dimensionless representation may enable the
graphical representation in two or three dimensions of expressions that, otherwise,
would not be possible to represent graphically. This feature may help in develop-
ing new identification procedures as it is shown in the identification of system with
inverse response.
There is, however, another advantage in representing a dynamical system by
means of dimensionless numbers. The dimensionless numbers obtained may provide
further physical insight than the dimensional counterpart. In case that certain dimen-
sionless numbers, by their construction and their physical meaning, are likely to be
large and/or small, simpler relations may be established.
The general procedure to apply the dimensional analysis for system identifica-
tion and model order reduction is the following:
1. The starting point is a dimensionless representation of the system to be identified
or reduced.
2. Analyse the dimensionless parameters obtained, looking for large and small
dimensionless numbers.
3. Discard the too large or small dimensionless parameters.
4. Formulate useful relations for identification or model reduction using the
remaining dimensionless numbers.
The general procedure is applied to two particular problems to show an example
of application. Both problems are not easily solved by other means. The first exam-
ple of application is the identification of second order system with inverse response
plus time delay. The second application example is the model order reduction for
decentralized control.

5.3 Example 1: Second order inverse response model

identification

5.3.1 Problem statement

Inverse response systems are present in several industrial and chemical processes,
such as boilers, chemical reactors, etc. [7]. These processes present serious chal-
lenges for control design because of their nature of competing dynamics and they
impose fundamental limitations to the performance and robustness attainable lev-
els [10]. Accurate and easy-to-perform identification techniques are then necessary
in order to control and establish fundamental limitations of these processes.
The objective is to derive an identification algorithm for inverse response sys-
tems with two poles and a time delay (see (5.1)), based on the process reaction curve.
The assumed structure of the plant to be identified is

k(− τ3 s + 1)
G(s) = e−hs (5.1)
(τ1 s + 1)(τ2 s + 1)
 Dimensionless systems identification and model order reduction 53

Although in the literature it is possible to find identification algorithms for

systems described by (5.1), they are complex because they are based on complex
experimental data in time and frequency domains [3]. Furthermore, the case of
inverse response of integrating processes has also been investigated. The work of
Reference 6 provides a systematic approach to obtain all the three parameters for
integrating processes with dead time and inverse response. It is accomplished by
numerically solving a set of three nonlinear equations. Another approach for inte-
grating processes is proposed in Reference 11 in which the zero time constant is
identified in closed form by means of an analytic expression and the system gain and
the pole time constant are calculated by means of numerical integral evaluation.
Previous identification procedures are complex due to the data required and
their operations. However, simpler identification algorithms based on simple exper-
iments are well accepted in industry because normally simple models are pur-
sued for control purposes and the economical impact due to experimentation is
minimized.
In what follows we make use of the dimensionless representation of model (5.1)
to derive a very simple identification algorithm. The key point for the algorithm is to
realize that certain dimensionless number a is normally small in second order inverse
response systems, due to the competing dynamics of these systems.

5.3.2 Dimensionless representation of second order inverse

response model
The dimensionless representation of transfer function (5.1) is

G(s̄) (− bs̄ + 1)
Ḡ(s̄) := = e−cs̄ (5.2)
k (s̄ + 1)(as̄ + 1)

where
● s̄ = τ1 s, which can be interpreted either as a time scaling or as a frequency scaling.
● a is the ratio between the fast and slow time constants, that is a = τ2 /τ1 . The
values of a considered are a ∈ [0.1, . . . , 0.5].
● b is the ratio between the zero time constant and the slow time constant, that is
b = τ3 /τ1 . The values of b considered are b ∈ [0.1, . . . , 4].
● c is the ratio between the time delay and the slow time constant, that is c = h/τ1 .
The use of the dimensionless representation (5.2) is a key point enabling the
identification algorithm. It allows the normalized system output to be character-
ized (i.e. y/k) at normalized time instants (i.e. ta /τ1 ) as a function of just three
dimensionless parameters, a, b, and c. Note that the transfer function (5.1) has five
parameters, so dimensional analysis permits a reduction of two parameters by the use
of dimensionless numbers.
The objective is, on the basis of the sole information of, at least, three points of
the step response curve, to obtain an identification algorithm that provides the values
54 Application of dimensional analysis in systems modeling and control design

of k, h, τ1 , b, and a of transfer function (5.2), which is equivalent to identifying the

values of k, h, τ1 , τ2 , and τ3 of transfer function (5.1).

5.3.3 Identification procedure

The proposed identification algorithm provides the values of k, h τ1 , b, and a in a
sequential manner by the use of, at least, three points of the system reaction curve.

Identification of the gain k

Given the total output change y due to an input change u, the system gain is
determined as
y
k= (5.3)
u

Identification of the time delay h

The time delay h can be identified from the step reaction curve following graphi-
cal methods, as presented in Reference 1. In this way both the system gain k and
the time delay h are identified from the reaction curve and are the basis for the
identification of system dynamic behavior as discussed in the following.

Identification of the dominant time constant τ 1

The system inverse response can be understood as the interaction of slow and fast
dynamics [5]. In fact the considered transfer function (5.1) can be equivalently
represented as a sum of two first order systems with inverse sign gains:
 
(− bs̄ + 1) k1 k2
Ḡ(s̄) = e−cs̄ = − e−cs̄ (5.4)
(s̄ + 1)(as̄ + 1) s̄ + 1 as̄ + 1

with
1+b
k1 = (5.5)
1−a
a+b
k2 = (5.6)
1−a

We have an inverse response system if the following condition is accompli-

shed, which is the dimensionless counterpart of the dimensional one provided in [5]:

1 1+b
> >1 (5.7)
a a+b

In physical processes showing inverse response behavior it is normally the case

that the competing dynamics are different in the sense that one dynamic is much
slower than the other one. In fact from expression (5.7) if both dynamics are equal
 Dimensionless systems identification and model order reduction 55

(i.e. τ1 = τ2 , or equivalently a = 1) the process cannot have inverse response behavior.

Moreover, in case that both dynamics are similar ( i.e. a ≈ 1), that implies a very
high gain of the competing dynamics k1 and k2 due to the denominator term (1 − a)
in (5.5) and (5.6). Thus, it is expected that in physical processes the cause of inverse
response is mainly due to the difference of two moderate gains competing at distinct
dynamics instead of two dynamically similar high gain processes.
The system step time response for absolute time values ta greater than the
system delay h, expressed by means of dimensionless constants, is given by
   
y b − 1 −(ta −h)/τ1 b − a −(ta −h)/τ1 a
=1+ e + e (5.8)
ku 1−a a−1
which can be rewritten using the normalized output y% = y/ku = y/y and the
relative normalized time t  = (ta − h)/τ1 as follows:
   
b − 1 −t  b − a −t  /a
y% = 1 + e + e (5.9)
1−a a−1
Note that in the rest of the chapter, the time variables t  are not absolute but relative
(i.e. t  = ta − c).
Assuming, as discussed above, that constant a is small (e.g. a < 0.5), that is

both poles have well-differentiated dynamics, it implies that the term e−t /a in (5.9)

tends to zero faster than e−t . In this case it is reasonable to approximate the step
response (5.9) for large enough values of time t  by
 
b − 1 −t̄ 
y% ≈ ȳ% := 1 + e (5.10)
1−a
In order to compare graphically both outputs, in Figure 5.1 we present the
exact and approximate outputs for a = 0.1 and a = 0.5 (in both cases b = −1). As
can be seen the smaller the a parameter, the better the approximation of (5.10) for
smaller time values.
Before proceeding, it is important to stress the distinction between tx and t̄x .
Parameter tx is the normalized time for the normalized output achieving the value
of x% as given by (5.9). Parameter t̄x is the normalized time for normalized out-
put achieving the value of x% as given by approximated equation (5.10). As time
increases, the value of t̄x tends to tx .
Now, taking logarithms from (5.10):
 
b−1
ln (y% − 1) = ln − t̄  (5.11)
1−a
and particularizing (5.11) for two points {yx% , tx } and {yy% , ty } of the reaction curve
with tx and ty large enough, we can identify τ1 as
ty − tx
τ1 =   (5.12)
yx% −1
ln yy% −1
56 Application of dimensional analysis in systems modeling and control design

Step responses a = 0.1

1
Exact
0.5 Approximated
Normalized output

−0.5

−1

−1.5
0 1 2 3 4 5 6
Normalized time

Step responses a = 0.5

1
Exact
Approximated
Normalized output

−1

−2

−3
0 1 2 3 4 5 6
Normalized time

Figure 5.1 Normalized exact and approximated step responses, for a = 0.1 and
a = 0.5

Identification of constant b
The constant b, once the dominant time constant τ1 is known, can be identified by
taking the point of minimum value of the step response ( yp , tp ). It follows that the
derivative of the output at that point is zero, then
   
dy 1 − b −tp a − b −tp /a 1
= e + e =0 (5.13)
dt 1−a a−1 a

From (5.9) and (5.13) we finally obtain

1 − yp%
b̄ = 1 −  (5.14)
e−tp

Identification of constant a
Finally constant a is identified by analyzing the variation of the relative normal-
ized time t  from (5.9) with respect to parameters a and b. In Figure 5.2 it can be
observed how the normalized time varies for output values from 10% to 90% as
 Dimensionless systems identification and model order reduction 57

a = 0.1 a = 0.3
4 5

4
3
90% 3
2
t′

t′
80%
2
70%
60%
1 50%
40% 1
30%
20%
0 10% 0
−4 −3 −2 −1 0 −4 −3 −2 −1 0
b b
Exact
Approximation
a = 0.55 a = 0.9
5 6

4 5

4
3
t′

3
t′

2
2
1 1

0 0
−4 −3 −2 −1 0 −4 −3 −2 −1 0
b b

Figure 5.2 Normalized time as a function of b for a constant

a function of parameter b assuming a constant. The curves are clearly not linear;
however, a quadratic form can be fairly adjusted as can be seen in Figure 5.2. On
the other hand, Figure 5.3 presents the normalized time variation, also for output
values between 10% and 90% as a function of parameter a assuming now a constant
value of b. It can be seen that in this case the curves can be well adjusted by a linear
function.
From the results of Figures 5.2 and 5.3, it is proposed to adjust the normalized
time by means of the following expression:

tx = m1x + m2x a + (m3x + m4x a)b + (m5x + m6x a)b2 (5.15)

with mix , i ∈ [1, 6] parameters adjusted by least squares. The rationale of (5.15) is
that for b constant (5.15) is linear in a, while for a constant (5.15) is quadratic in b,
as required from Figures 5.2 and 5.3. Moreover, it has enough degrees of freedom
to capture the curvature changes as a function of a and b. In Figure 5.4 the values
mix , i ∈ [1, 6] adjusted by least square approximation can be seen. The approximation
58 Application of dimensional analysis in systems modeling and control design

b = −0.1 b = −0.6
4 5
90%

4
3 80%
70% 3
2 60%
t′

t′
50% 2
40%
30%
1 20% 1
10%

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
a a
Exact
Approximation
b = −1.6 b = −3.6
5 6

5
4

4
3
t′

t′

2
2

1 1
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
a a

Figure 5.3 Normalized time as a function of a for b constant

is performed on the basis of the dimensionless transfer function (5.2). Thus, from
function (5.15) and values b and τ1 previously identified, a is given by

tx − (m1x + m3x b + m5x b2 )

ā = (5.16)
m2x + m4x b + m6x b2

General identification algorithm

The general identification algorithm provides the transfer function parameters by
using three or four points of the step response curve, (tp , yp ), (tx , yx ), (ty , yy ), and
possibly (tz , yz ). First, identify from the whole step response the system gain as
k = u
y
and the delay h, if any. Choosing four points from the step response, the
transfer function parameters are

ty − t x
τ1 =   (5.17)
yx% −1
ln yy% −1
 Dimensionless systems identification and model order reduction 59

2.5

m1
2 m2
m3
m4
m5
1.5
m6

1
mij

0.5

−0.5

−1
10 20 30 40 50 60 70 80 90
Percentage of the normalized output

Figure 5.4 Parameters mij

1 − yp%
b = 1−  (5.18)
e−tp
tz − (m1z + m3z b̄ + m5z b̄2 )
a= (5.19)
m2z + m4z b̄ + m6z b̄2

We propose the following identification algorithm for simplicity and general-

ity. The points considered for calculating dominant time constant τ1 are (t47% , y47% )
and (t90% , y90% ). The constant a is calculated using again (t47% , y47% ), thus just three
points are required.

τ1 = 0.60(t90% − t47% ) (5.20)

1 − yp%
b = 1− 
e−tp
t −n
a = 47%x
m

where the constants m and n are given in Table 5.1. Note that m = m2x + m4x b + m6x b2
and n = m1x + m3x b + m5x b2 .
60 Application of dimensional analysis in systems modeling and control design

Table 5.1 Constants m and n for the proposed identification

algorithm

b m n b m n

0.1 0.9094 0.6905 2.1 0.4798 −0.8682

0.2 0.8910 0.6258 2.2 0.4549 −0.9607
0.3 0.8722 0.5597 2.3 0.4297 −1.0546
0.4 0.8532 0.4922 2.4 0.4042 −1.1500
0.5 0.8338 0.4233 2.5 0.3784 −1.2467
0.6 0.8141 0.3530 2.6 0.3522 −1.3448
0.7 0.7940 0.2814 2.7 0.3258 −1.4443
0.8 0.7737 0.2083 2.8 0.2990 −1.5452
0.9 0.7530 0.1338 2.9 0.2718 −1.6475
1.0 0.7320 0.0580 3.0 0.2444 −1.7512
1.1 0.7107 −0.0192 3.1 0.2166 −1.8563
1.2 0.6890 −0.0979 3.2 0.1886 −1.9627
1.3 0.6671 −0.1779 3.3 0.1602 −2.0706
1.4 0.6448 −0.2593 3.4 0.1314 −2.1799
1.5 0.6222 −0.3421 3.5 0.1024 −2.2905
1.6 0.5992 −0.4263 3.6 0.0730 −2.4025
1.7 0.5760 −0.5119 3.7 0.0433 −2.5160
1.8 0.5524 −0.5989 3.8 0.0133 −2.6308
1.9 0.5285 −0.6872 3.9 −0.0170 −2.7470
2.0 0.5043 −0.7770 4.0 −0.0477 −2.8646

Identification accuracy analysis

In order to gain insight on the identification accuracy we perform a prediction
error evaluation of models identified for the admissible parameter range. Thus, we
calculated the prediction error


∞
S2 = (y(tk ) − ym (tk ))2 (5.21)
k=1

where y(tk ) is the output value of the exact step response and ym (tk ) is the output
model value. In all cases the simulation time of step response has been 25 s, and
the time increment t = 0.025 s. In Figure 5.5 the quadratic prediction error is
shown as a function of a and b.
It can be seen that the error S2 increases as a increases as expected, although
for values of b > 1.5 the tendency becomes nearly erratic. However, it is difficult to
find tendencies around the central zone, what implies that the algorithm makes use
of important information because the residues show no tendency apart from the ones
considered and expected above.
In Figure 5.6 we show the best and the worst identified step responses. The
best model is given for parameters a = 0.4 and b = 2.25 yielding a prediction error
value of S2 = 0.001579. On the other hand the worst identified model is given for
parameters a = 0.2 and b = 4 with prediction error S2 = 0.853486.
 Dimensionless systems identification and model order reduction 61

0.9
0.8
0.7
0.6
0.5
S2

0.4
0.3
0.2
0.1
0
1
0.9
0.8 4
0.7 3.5
0.6 3
2.5
a 0.5 2
0.4 1.5
0.3 1 b
0.5
0.2 0

Figure 5.5 Prediction error S2

5.3.4 Application examples

Numerical example
The proposed identification algorithm is applied to the identification of the plant
given by k = 1, b = 1, τ1 = 1, and time delay h = 1 for three different values of
parameter a = {0.2, 0.35, 0.5}. The objective is to compare the algorithm perfor-
mance with respect to distinct parameter a values. In Figure 5.7 the plant output
is plotted against the identified plant output. The step input is of unit magnitude
and applied at time zero. It can be seen that in all cases the responses are fairly
similar.

Continuous stirred tank reactor

In this section a case-study example is provided. We consider the isothermal contin-
uous stirred tank reactor (CSTR), as the one in Figure 5.8, where the isothermal
series/parallel Van de Vusse reaction [4,12] is taking place. The reaction can be
described by the following scheme:

k1 k2
A −→ B −→ C (5.22)
k3
2A −→ D
62 Application of dimensional analysis in systems modeling and control design

Best identified model

100

50
y, ym, u (%)

−50

−100

−150
0 1 2 3 4 5 6 7 8 9 10

Worst identified model

100
y
50 ym
u
0
y, ym, u (%)

−50
−100
−150
−200
−250
0 1 2 3 4 5 6 7 8 9 10
Time (s)

Figure 5.6 Best and worst identified models

Applying a mass balance, the system can be described by the following model:

dCA (t) Fr (t)

= [CAi − CA (t)] − k1 CA (t) − k3 CA2 (t)
dt V
dCB (t) Fr (t)
=− CB (t) + k1 CA (t) − k2 CB (t) (5.23)
dt V

where Fr is the feed flow rate of product A, V is the reactor volume, which
is kept constant during the operation, CA and CB are the reactant concentrations
in the reactor, and ki (i = 1, 2, 3) are the reaction rate constants for the three
reactions.
In this case, the variables of interest are: the concentration of B in the reactor
(CB as the controlled variable), the flow through the reactor (Fr as the manipu-
lated variable), and the concentration CAi of A in the feed flow (whose variation
can be considered as the disturbance). The kinetic parameters are chosen to be
k1 = 5/6 min−1 , k2 = 5/3 min−1 , and k3 = 1/6 l mol−1 min−1 . Also, it is assumed
that the nominal concentration of A in the feed (CAi ) is 10 mol l−1 and the volume
V = 700 l.
 Dimensionless systems identification and model order reduction 63

a = 0.2
1
Amplitude

0.5
0
−0.5
−1
0 1 2 3 4 5 6 7
Time (s)

a = 0.35
1
Amplitude

0.5
0
−0.5
−1
0 1 2 3 4 5 6 7
Time (s)

a = 0.5
1
Exact output
Amplitude

0.5
Identified output
0
−0.5
−1
0 1 2 3 4 5 6 7
Time (s)

Figure 5.7 Example 1 – Step outputs’ comparison

Using (5.23) and the parameters values, the characterization of the steady-state
for the process can be obtained as it is shown in Figure 5.9, for three concentrations
of CAi , where it is easy to see the non-linearity of the system.
Initially, the system is at the steady-state (therefore the operational point) with
CAo = 2.9175 mol l−1 and CBo = 1.10 mol l−1 . It is assumed that changes in the
set-point would be not bigger than 10% and the possible disturbance in CAi can
variate around ±10%.
Applying the proposed methodology according to (5.20), with a 10% change on
the set-point, the identified process model is

0.3199(− 0.3520s + 1)
G(s) = (5.24)
(0.5619s + 1)(0.3086s + 1)

Figure 5.10 shows the process output and the identified model (5.24) for a step
change in the process input (yu (t)). It is possible to see that the identification pro-
cedure has a very good accuracy, specially during the first time instants of the step
response. The step response of the linearized model obtained on the operating point
64 Application of dimensional analysis in systems modeling and control design

A/C
10

I/P
AV AY
10 10
AT
10 Feed

Coolant

Product

Figure 5.8 Example 2 – CSTR system

is also shown in Figure 5.10. The model obtained by the simple step response iden-
tification is better than the more complex linearized model obtained analytically.
The applicability of the proposed algorithm to obtain accurate models for CSTR
processes with second order inverse response is then justified.

5.4 Example 2: Reduced effective transfer function reduction

for PID decentralized control

5.4.1 Problem statement

Processes in industrial settings are normally multivariable. Multivariable control
systems may be designed using multivariable control structures [10]. However,
multivariable controller design and tuning are far more complicated than of the
well-established single loop control. As a result, one practical approach is to control
multivariable processes by using single input/single output (SISO) control structures,
that is decentralized control. Although decentralized control is less complex than
multivariable control, even for industrial multivariable processes showing simple
 Dimensionless systems identification and model order reduction 65

CSTR − Steady state plots

6
5
4
3
CA

2
1
0
0 100 200 300 400 500 600 700 800 900 1000

1.5

1
CAi = 8 mol/l
CAi = 10 mol/l
CB

0.5 CAi = 12 mol/l

0
0 100 200 300 400 500 600 700 800 900 1000
F, l/min

Figure 5.9 Example 2 – Steady-state characterization for the reactor

74

73.5

73

72.5

72
Process
yu(t)%

71.5 Identified model

Linearized model
71

70.5

70

69.5

69
0 1 2 3 4 5 6
Time (min)

Figure 5.10 Example 2 – Reaction curve for process, linearized model, and
identified model
66 Application of dimensional analysis in systems modeling and control design

dynamics, the design of decentralized controllers has to deal with complex model
dynamics due to loop interaction. The complexity of resulting models dramatically
limits the applicability of PID tuning rules existing in the literature [8], because
they are based on simple models. For instance, consider the two inputs/two out-
puts (TITO) multivariable system shown in Figure 5.11. Furthermore, consider that
transfer functions G11 , G12 , G21 , and G22 are modeled by first order plus dead time
(FOPDT) models, that is the matrix transfer function is given by

⎡ ⎤
k11 e−h11 s k12 e−h12 s
⎢ τ11 s + 1 τ12 s + 1 ⎥
G(s) = ⎢
⎣ k e−h21 s
⎥ (5.25)
21 k22 e−h22 s ⎦
τ21 s + 1 τ22 s + 1

In order to design decentralized controller C11 (see Figure 5.11), the trans-
fer function between output Y1 and input U1 , required to design decentralized
controller C11 , is

Y1 C22 G12 G21

e
G11 = = G11 − (5.26)
U1 1 + C22 G22

where for the sake of readability the complex variable s is not shown. Equation (5.26)
is called effective transfer function (ETF) and it depends on the existing controller of
closed loop pair C22 .

R1 – U1 Y1
C11 G11

G21

G12

R2 Y2
C22 G22
U2
–

Figure 5.11 Decentralized control structure of a TITO system

 Dimensionless systems identification and model order reduction 67

For the sake of simplicity, the EFT dependence on existing controller C22 may
be eliminated under the perfect control assumption, because in this case [10],

Y2 C22 G22
= =1 (5.27)
R2 1 + C22 G22

and the ETF can be simplified as follows:

G12 G21
re
G11 = G11 − (5.28)
G22

Equation (5.28) is known as reduced effective transfer function (RETF), and it depends
on system transfer functions only.
Particularizing the RETF for the FOPDT TITO system, we have

k11 e−h11 s k12 k21 (τ22 s + 1)e−(h12 +h21 −h22 )s

re
G11 = − (5.29)
(τ11 s + 1) k22 (τ12 s + 1)(τ21 s + 1)

At the sight of transfer function (5.29) it is not clear if it can be accurately

approximated by an FOPDT or an second order plus dead time (SOPDT) model,
commonly used as baseline for PID control design. In fact transfer function (5.29)
may even show inverse response behavior that cannot be captured by a minimum
phase model. Dimensional analysis is applied for model order reduction of the RETF,
characterizing the feasibility of approximating the RETF by means of FOPDT/SOPDT
systems. Furthermore, in base of dimensionless numbers relations, it is possible to
ascertain when transfer function (5.29) shows inverse response behavior, so care must
be taken with the controller design procedure.

5.4.2 Dimensionless representation of the reduced effective transfer

function
The first step of the approach is to represent the RETF (5.29) in dimensionless form,
that is given by

e−h̄11 s̄ (τ̄22 s̄ + 1)e−h̄ s̄

re
Ḡ11 = − k̄ (5.30)
(s̄ + 1) (τ̄12 s̄ + 1)(τ̄21 s̄ + 1)

with the dimensionless numbers Ḡ11 re

= G11 re
/k11 , s̄ = τ11 s, k̄ = (k12 k21 )/(k11 k22 ),
τ̄12 = τ12 /τ11 , τ̄21 = τ21 /τ11 , τ̄22 = τ22 /τ11 , h̄11 = h11 /τ11 , and h̄ = (h12 + h21 −
h22 )/τ11 .
The benefit of dimensionless representation is two-fold. On the one hand we
reduce the number of dependent parameters from eight dimensional parameters
68 Application of dimensional analysis in systems modeling and control design

to six dimensionless parameters. Second, and more important, the dimensionless

parameters enable the interpretation in terms of limiting values, what facilitates the
analysis of particular cases. In the rest of the chapter we consider the following
assumptions:
● All the previous dimensionless parameters are positive. The implication is
stability of FOPDT transfer functions and that negative gain signs appear in
pairs.
● For causality considerations, h̄ ≥ 0, that is h12 + h21 − h22 ≥ 0.
● We assume that k̄ < 1. By proper selection of input–output pairs, for instance
consider pairing following the relative gain array (RGA) procedure, it follows
that k̄ < 1.

5.4.3 Inverse response analysis

Given the dimensionless reduced effective transfer function (DRETF) (5.30), the first
point of study is to characterize when (5.30) shows an inverse response behavior.
In this case a nonminimum phase SOPDT model might be used. On the contrary,
if the DRETF does not show inverse response, then an FOPDT or a minimum phase
SOPDT model may be more appropriate to capture the fundamental dynamics of
DRETF. In what follows, we provide conditions for inverse response behavior, first
in the general case and, second, in some particular cases.

General case
Assume without loss of generality that h > h11 . On the contrary, if h < h11 , an
analogue analysis may be performed. The case h = h11 is analyzed in the next
section as a particular case.
In case that h > h11 , the DRETF (5.30) may be rewritten as

1 (τ̄22 s̄ + 1)e−h̄x s̄
re
Ḡ11 = − k̄ e−h̄11 s̄ (5.31)
(s̄ + 1) (τ̄12 s̄ + 1)(τ̄21 s̄ + 1)

where h̄x := h̄ − h̄11 > 0.

Transfer function (5.31) shows inverse response if the difference of transfer
functions between parentheses has zeros on the right half plane (RHP), that is pos-
itive zeros [10]. Note that the time delay h̄11 has no influence on the existence of
inverse response.
In order to characterize the inverse response property of transfer function
(5.31), we approximate the time delay h̄x by a first order Padé approximation,
yielding

1 (τ̄22 s̄ + 1)( −2h̄x s̄ + 1)
re
Ḡ11 = − k̄ e−T̄11 s̄ (5.32)
(s̄ + 1) (τ̄12 s̄ + 1)(τ̄21 s̄ + 1)( 2 s̄ + 1)
h̄x
 Dimensionless systems identification and model order reduction 69

which can be rewritten as

as̄3 + bs̄2 + cs̄ + d
re
Ḡ11 = e−h̄11 s̄ (5.33)
(s̄ + 1)(τ̄12 s̄ + 1)(τ̄21 s̄ + 1)( h̄2x s̄ + 1)

where the third order polynomial of the numerator P(s̄) := as̄3 + bs̄2 + cs̄ + d has
parameters defined as

a = (τ̄12 τ̄21 + k̄ τ̄22 )(h̄x /2)

h̄x  
b = τ̄12 τ̄21 + (τ̄12 + τ̄21 ) − k̄ τ̄22 − (1 + τ̄22 )h̄x /2
2

h̄x h̄x
c = τ̄12 + τ̄21 + − k̄ τ̄22 − +1
2 2

d = 1 − k̄

Moreover, we define the discriminant  of P(s̄) as  = 18abcd − 4b3 d + b2 c2 −

4ac3 − 27a2 d 2 [15].

Theorem 5.1. Assume that P(s̄) has only real valued zeros (i.e.  ≥ 0). Transfer
function (5.33) is minimum phase (i.e. all zeros are left half plane (LHP) zeros) if and
only if
 
τ̄12 τ̄21 + h̄2x + k̄ h̄2x τ̄12 + τ̄21 + h̄2x (1 + k̄)
τ̄22 < min , −1 (5.34)
k̄(1 − h̄2x ) k̄

Proof. By Descartes’ rule of signs we know that (i) the maximum number of posi-
tive zeros is equal to the changes of sign of P(s̄), and (ii) the minimum number of
negative zeros is equal to the changes of sign of P(− s̄).
Note that term a is always greater than zero because we are adding positive
values. Moreover, d is also greater than zero because, by proper pairing, k̄ < 1. As
a result, the number of sign changes is determined by parameters b and c. If b > 0
and c > 0 then the number of sign changes of P(s̄) is zero, thus there are no positive
zeros (i.e. no inverse response). In fact, in this case, the number of sign changes of
P(− s̄) is equal to 3, the number of negatives zeros.
If any of the parameters b or c, or both, is negative, then the number of sign
changes of P(s̄) is 2, whereas the number of sign changes of P(− s̄) is 1. Two possi-
bilities are then open: there are two positive zeros and one negative zero; or there are
two complex conjugate zeros and one negative zero. The existence of complex zeros
is discarded by assumption  ≥ 0. Therefore, there are two positive real zeros and
the system is in nonminimum phase.
Finally, b > 0 is equivalent to
τ̄12 τ̄21 + h̄2x + k̄ h̄2x
τ̄22 <   (5.35)
k̄ 1 − h̄2x
70 Application of dimensional analysis in systems modeling and control design

and c > 0 is equivalent to

τ̄12 + τ̄21 + h̄x
(1 + k̄)
τ̄22 < 2
−1 (5.36)
k̄
and the result follows.

Time delay factorization

Assuming that h = h11 , both delays in (5.30) may be factored, then the RETF may
be written as
 
1 (τ̄22 s̄ + 1)
Ḡ11 =
re
− k̄ e−h̄11 s̄ (5.37)
(s̄ + 1) (τ̄12 s̄ + 1)(τ̄21 s̄ + 1)

In this case, the time delay has no influence on the inverse response manifesta-
tion of the DRETF and the inverse response behavior of transfer function (5.37) may
be analyzed equivalently by transfer function

(τ̄12 s̄ + 1)(τ̄21 s̄ + 1) − k̄(τ̄22 s̄ + 1)(s̄ + 1) −h̄11 s̄

re
Ḡ11 = e (5.38)
(s̄ + 1)(τ̄12 s̄ + 1)(τ̄21 s̄ + 1)

The inverse response behavior occurs when transfer function (5.38) has, at
least, one right half plane zero (RHP). The zeros of the numerator polynomial are
given by

as̄2 + bs̄ + c = 0 (5.39)

with a = τ̄12 τ̄21 − k̄ τ̄22 , b = τ̄12 + τ̄21 − k̄(τ̄22 + 1), and c = (1√− k̄).
√ s̄ = (− b ± b − 4ac)/(2a). For
The zeros of polynomial (5.39) are given by 2

real zeros, by imposing the condition −b + b − 4ac < 0, we assure that the
2

zeros are LHP zeros and no inverse response is possible. By arranging and squar-
ing previous condition we have b2 − 4ac < b2 , yielding the condition ac > 0. This
is written in terms of dimensionless numbers as

(τ̄12 τ̄21 − k̄ τ̄22 )(1 − k̄) > 0 (5.40)

which is a necessary and sufficient condition to discard nonminimum phase behav-

ior in (5.37). As by assumption k̄ < 1, condition (5.40) is simplified to

τ̄12 τ̄21
k̄ < (5.41)
τ̄22

In case that poles are complex conjugates, the condition for LHP zeros is that
b > 0, that is

τ̄12 + τ̄21
k̄ < (5.42)
τ̄22 + 1
 Dimensionless systems identification and model order reduction 71

Pole-zero cancelation
The case of pole-zero cancelation occurs when τ̄12 = τ̄22 (or τ̄21 = τ̄22 ) yields to
further simplifications. In this case transfer function (5.31) may be reduced to

⎛   ⎞
−h̄x
1 + 1 2
s̄
re
Ḡ11 =⎝ − k̄   ⎠ e−h̄11 (5.43)
s̄ + 1 (τ̄21 s̄ + 1) h̄2x s̄ + 1

The inverse response behavior occurs when transfer function (5.43) has an
RHP. The zeros of the numerator polynomial are given by

as̄2 + s̄ + c = 0 (5.44)

with a = (τ̄21 + k̄) h̄2x , b = τ̄21 + h̄2x − k̄(1 + h̄2x ), and c = (1 − k̄).
By solving equation (5.44) we arrive at the condition for minimum phase for real
zeros, given by

h̄x
(τ̄21 + k̄) (1 − k̄) > 0 (5.45)
2

which by assumption we have 0 < k̄ < 1, and the non-minimum phase behavior is
discarded for any RETF.
For complex zeros the condition to discard inverse response is that b > 0, which
yields

τ̄21 + h̄x
2
k̄ < (5.46)
h̄x
2
+1

Time delay factorization and pole-zero cancelation

The case of time delay factorization and pole-zero cancelation yields to further
simplifications. In this case transfer function (5.31) may be reduced to
 
1 1
re
Ḡ11 = − k̄ (5.47)
s̄ + 1 τ̄21 s̄ + 1

The difference of two first order systems yields an inverse response if τ̄121 > k̄1 >
1 [5]. By assumption, we have 0 < k̄ < 1, it is always true that k̄1 > 1, thus, there is
inverse response behavior if and only if

k̄ > τ̄21 (5.48)

72 Application of dimensional analysis in systems modeling and control design

Table 5.2 Summary of no inverse response conditions for each

case considered

Assumptions No inverse response if...

 
τ̄ τ̄ + h̄x +k̄ h̄x τ̄12 +τ̄21 + h̄2x (1+k̄)
≥0 τ̄22 < min 12 21 2h̄x 2 , k̄
−1
k̄(1− 2 )

h̄11 = h̄ ,  ≥ 0 k̄ < τ̄12 τ̄21

τ̄22
τ̄12 +τ̄21
h̄11 = h̄ ,  < 0 k̄ < τ̄22 +1
τ̄12 = τ̄22 ,  ≥ 0 No inverse response
τ̄21 + h̄2x
(or τ̄21 = τ̄22 ),  < 0 k̄ < h̄x
2 +1
h̄11 = h̄ and
τ̄12 = τ̄22 k̄ < τ̄21
(or τ̄21 = τ̄22 )

Summary of inverse response analysis

In Table 5.2 we present a summary of inverse response conditions for each
case considered. Recall that the discriminant for the general case is defined as
 = 18abcd − 4b3 d + b2 c2 − 4ac3 − 27a2 d 2 , whereas the discriminant for the rest
of cases is defined as  = b2 − 4ac.

5.4.4 Reduced order model: general case

One general approach to obtain reduced order models of the RETF is to perform
Maclaurin approximations, as presented in Reference 14. The idea is to expand the
RETF in its Maclaurin series as follows:
re
Ḡ11 = a11 + b11 s̄ + c11 s̄2 + O(s̄3 ) (5.49)

where the polynomial coefficients are

a11 = Ḡ11
re
(0)
re
d Ḡ11 (s̄) 
b11 = 
d s̄ s̄=0

1 d 2 Ḡ11
re
(s̄) 
c11 = 
2 d s̄2 s̄=0

Next, we obtain the analytic polynomial coefficients of the Maclaurin series for
the dimensionless FOPDT model

k e−h̄s̄
Ḡ(s) = (5.50)
k11 τ̄ s̄ + 1
 Dimensionless systems identification and model order reduction 73

that is given by
 
k k k 1 2
Ḡ re = − (τ̄ + h̄)s̄ + h̄ + (h̄ + τ̄ )τ̄ s̄2 + O(s̄3 ) (5.51)
k11 k11 k11 2

For SOPDT processes, the same procedure may be repeated. However, the
SOPDT parameters are not easily obtainable as before [14]. Section 5.4.5 shows how
SOPDT models may be approximated without requiring the Maclaurin approximation.

5.4.5 Reduced order model: particular cases

At the sight of dimensionless number values, we may obtain reduced order models in
a straightforward manner. In fact, if some dimensionless numbers are small or there
is certain relationship among them, simpler models may be easily obtained.

Dimensionless gain
In case that k̄ 1, an approximation to the dimensionless reduced effective transfer
function is simply

1 −h̄11 s̄
e (5.52)
s̄ + 1

As k̄ = (k12 k21 )/(k11 k22 ), the condition k̄ 1 implies that the cross gains are
small and we have low interaction. As a result we may consider the problem as not
coupled.

Slow model
Previous model may be improved in the case of two time scales. We have two time
scales if τ̄12 1 and τ̄21 1. In this case we may disregard fast model dynam-
ics. The DRETF (5.29) is given by the slow model with FOPDT transfer function
given by

(1 − k̄) −h11 s̄
e (5.53)
s̄ + 1

However, note that when τ̄12 1 and τ̄21 1, it might be possible to per-
form time scale decoupling [9], and the original loop pairing might not be the
most adequate, despite RGA pairing.

Time delay factorization and pole-zero cancelation

Recall that under the assumption h = h11 , equivalently h11 + h22 = h12 + h21 , the
delay can be factored, so the dimensionless reduced effective transfer function is
74 Application of dimensional analysis in systems modeling and control design

given by (5.37). In this case the exact dimensionless reduced effective transfer func-
tion is given by the following transfer function:

(τ̄12 s̄ + 1)(τ̄21 s̄ + 1) − k̄(τ̄22 s̄ + 1)(s̄ + 1) −h̄11 s̄

re
Ḡ11 = e (5.54)
(s̄ + 1)(τ̄12 s̄ + 1)(τ̄21 s̄ + 1)

that is too complex to be used for control design with commonly used PID tuning
rules.
The case of pole-zero cancelation by τ̄12 = τ̄22 (or τ̄21 = τ̄22 ) yields to further
simplifications. In this case transfer function (5.54) may be reduced to

(τ̄12 − k̄)s̄ + (1 − k̄) −h̄11 s̄

re
Ḡ11 = e (5.55)
(s̄ + 1)(τ̄12 s̄ + 1)

Transfer function (5.55) is exactly modeled by second order transfer function.

We have the following cases:
● (τ̄12 − k̄) < 0. In this case a non-minimum phase SOPDT model is able to
exactly represent the dimensionless reduced effective transfer function. There
is inverse response.
● (τ̄12 − k̄) > 0. In this case a minimum phase with LHP zero SOPDT model is
able to exactly represent the dimensionless reduced effective transfer function.
● (τ̄12 − k̄) = 0. In this case an SOPDT model is able to exactly represent the
dimensionless reduced effective transfer function.

Summary of model order reduction

In Table 5.3 we present a summary of reduced order models.

Table 5.3 Summary of model order reduction for

each case considered

Assumptions Model
1 −h̄11 s̄
k̄ 1 e
s̄ + 1
(1 − k̄) −h̄11 s̄
τ̄12 1 and τ̄21 1 e
s̄ + 1
(τ̄12 − k̄)s̄ + (1 − k̄) −h̄11 s̄
h̄11 = h̄ and τ̄12 = τ̄22 e
(s̄ + 1)(τ̄12 s̄ + 1)
(or τ̄21 = τ̄22 )
(1 − k̄) −h̄11 s̄
h̄11 = h̄ and τ̄12 = τ̄22 e
s̄ + 1
(or τ̄21 = τ̄22 ) and τ̄12 = 1
 Dimensionless systems identification and model order reduction 75

5.4.6 Application examples

Vinante and Luyben (VL) column
Consider the Vinante and Luyben distillation column [13]

⎡ ⎤
−2.2e−s 1.3e−0.3s
⎢ 7s + 1 7s + 1 ⎥
G(s) = ⎢
⎣ −2.8e−1.8s
⎥ (5.56)
4.3e−0.35s ⎦
9.5s + 1 9.2s + 1

The RETF dimensionless parameters are given in Table 5.4.

Table 5.4 Vinante and Luyben DRETF dimensionless numbers

k̄ τ̄ 12 τ̄ 21 τ̄ 22 h̄11 h̄

0.38 1 1.35 1.31 0.14 0.25

0.8

0.6

DREFT
0.4
FOPDT Maclaurin
FOPDT slow
0.2
SOPDT

0
0 1 2 3 4 5

Bode diagram
0
Magnitude (dB)

−20 DREFT
FOPDT Maclaurin
FOPDT slow
−40
SOPDT

−60
10−2 10−1 100 101 102 103
Frequency (rad/s)

Figure 5.12 Step response and Bode magnitude plot of reduced models of Vinante
and Luyben DRETF
76 Application of dimensional analysis in systems modeling and control design

The DRETF has no inverse response behavior because the discriminant is  =

0.07 > 0, and applying Theorem 5.1 (condition (5.34)), we have

1.31 < min{1.46, 5.31} (5.57)

which discards inverse response behavior.

Note that τ̄21 ≈ τ̄22 , thus by the pole-zero cancelation case, we can also discard
non-minimum phase behavior using condition (5.46), that yields

0.38 < 1.33 (5.58)

which is true and discards inverse response behavior of DRETF.

As inverse response behavior is discarded, two FOPDT models and one minimum
phase SOPDT model are proposed: first, the ‘FOPDT Maclaurin’ approximation that
requires no assumptions; second, the ‘FOPDT Slow’ model obtained assuming that
k̄ 1 and given by (5.53); third, the ‘SOPDT’ model derived by assuming (i) τ̄21 =
τ̄22 , and (ii) h̄11 = h̄ . Figure 5.12 shows the step responses and the magnitude Bode
diagram of the models considered.

Polymerization reactor
Consider now the polymerization reactor [2]

⎡ ⎤
22.89e−0.2s −11.64e−0.4s
⎢ 4.572s + 1 1.807s + 1 ⎥
G(s) = ⎢
⎣ 4.689e−0.2s
⎥ (5.59)
5.80e−0.4s ⎦
2.174s + 1 1.801s + 1

The RETF dimensionless parameters are given in Table 5.5. Note that τ̄12 ≈ τ̄22 .
Furthermore, we have h̄11 = h̄ , so we consider the time delay factorization and
pole-zero cancelation case. Note also that in this case k̄ < 0, hence the results must
be particularized to this case by realizing that no inverse response is possible because
two first order transfer functions are being added in (5.47).
The same three models as previous examples are used and the results can be seen
in Figure 5.13. Note that in this case, the SOPDT model is equal to the DRETF.

Table 5.5 Polymerization reactor DRETF dimensionless numbers

k̄ τ̄ 12 τ̄ 21 τ̄ 22 h̄11 h̄

−0.41 0.39 0.47 0.39 0.0437 0.0437

 Dimensionless systems identification and model order reduction 77

1.5

DREFT
1
FOPDT Maclaurin
FOPDT slow
0.5 SOPDT

0
0 1 2 3 4 5

Bode diagram
10

0
Magnitude (dB)

−10 DREFT
FOPDT Maclaurin
−20 FOPDT slow
−30 SOPDT

−40
10−2 10−1 100 101 102
Frequency (rad/s)

Figure 5.13 Step response and Bode magnitude plot of reduced order models of
Polymerization Reactor DRETF

References

[1] K.J. Åström and T. Hägglund. PID Controllers: Theory, Design and Tuning.
Instrument Society of America, 1995
[2] J. Garrido, F. Vázquez, and F. Morilla. ‘Centralized inverted decoupling
for TITO processes.’ In 15th IEEE International Conference on Emerging
Technologies and Factory Automation, Bilbao, Spain, 2010
[3] H.P. Huang, M.W. Lee, and C.L. Chen. ‘A system of procedures for iden-
tification of simple models using transient step response.’ Industrial and
Engineering Chemistry Research, 40(8):1903–1915, 2001
[4] C. Kravaris and P. Daoutidis. ‘Nonlinear state feedback control of sec-
ond order nonminimum-phase nonlinear systems.’ Computers & Chemical
Engineering, 14(4–5):439–449, 1990
[5] K. Linoya and R.J. Altpeter. ‘Inverse response in process control.’ Industrial
and Engineering Chemistry Research, 54(7):39–43, 1962
[6] W.L. Luyben. ‘Identification and tuning of integral processes with dead-
time and inverse response.’ Industrial and Engineering Chemistry Research,
42:3030–3035, 2003
78 Application of dimensional analysis in systems modeling and control design

[7] T.E. Marlin. Process Control. Designing Process and Control Systems for
Dynamic Performance. McGraw-Hill, 2000
[8] A. O’Dwyer. Handbook of PI and PID Controller Tuning Rules. Imperial
College Press, 2006
[9] B.A. Ogunnaike and W. Harmor Ray. Process Dynamics, Modelling and
Control. Oxford University Press, 1994
[10] S. Skogestad and I. Postlethwaite. Multivariable Feedback Control. Analysis
and Design. Wiley, 1996
[11] A. Srivastava and A.K. Verma. ‘Identification of integrating processes with
deadtime and inverse response.’ Industrial and Engineering Chemistry
Research, 46:8270–8272, 2007
[12] J.G. Van de Vusse. Plug-flow type reactor versus tank reactor. Chemical
Engineering Science, 19:964, 1964
[13] C.D. Vinante and W.L. Luyben. Experimental studies of distillation decou-
pling. Kemin Teollisuusmaalaus, 1972
[14] T.N.L. Vu and M. Lee. Independent design of multi-loop pi/pid controllers
for interacting multivariable processes. Journal of Process Control, 20:922–
933, 2010
[15] E.W. Weisstein. CRC Concise Encyclopedia of Mathematics. Chapman &
Hall/CRC, 2003
 Chapter 6
Homogeneity of PID tuning rules
with A. Ibeas, C. Pedret and S. Alcántara

6.1 Introduction
We have seen that any physical quantity, apart from its value, has a unit that
measures its physical dimension. Physical dimensions and units, like length (m),
mass (kg), and time (s), relate mathematics and the physical world by physical equa-
tions. Valid physical equations describing physical system behavior are dimensionally
homogeneous. In the same way, valid controller equations, the controller being a
physical system relating measured outputs with control actions, are also dimension-
ally homogeneous. In contrast, the controller tuning rule, that is the rule providing
the controller parameters as a function of model parameters, is just a mathematical
relationship and, a priori, there is not any physical constraint requiring dimensional
homogeneity.
In fact, homogeneous tuning rules form a subset of the possible tuning rules,
as nonhomogeneous tuning rules are discarded. As a result, is the homogeneity
property of a tuning rule limiting the achievable performance? Could be the case
that nonhomogeneous tuning rules would provide a better closed loop performance?
For instance, for PID controllers, dimensionally homogeneous tuning rules and
nonhomogeneous tuning rules can be found in Reference 2.
In the present chapter we demonstrate that tuning rules derived from mini-
mization of indexes depending on closed loop transfer functions yield homoge-
neous and nonhomogeneous tuning rules, depending on the closed loop functions
considered in the minimization index.
However, when index minimization yields nonhomogeneous tuning rules, the
index can be transformed by dimensional weighting factors, recovering homogeneity
property of the tuning rule. As a result, no optimality is lost by using homoge-
neous tuning rules, and the prevalence of dimensional homogeneous tuning rules is
explained.
However, note that when dimensional weighting factors are used in the min-
imization index, the dimensionless weights depend on the plant parameters value
instead of being a constant factor. As a result, homogeneous tuning rules have a prop-
erty lacking on nonhomogeneous tuning rules, that is dimensionally homogeneous
tuning rules preserve dimensional similarity.
80 Application of dimensional analysis in systems modeling and control design

In this way, closed loops of similar plants tuned with homogeneous tuning rules
are also similar, which implies that their closed loop responses are related by gain
and time scaling factors. In other words, closed loop responses of dimensionally
similar plants are equal except for gain and time scaling when dimensionally similar
tuning rules are used. Consider for instance that experimental results show a 5%
overshoot on the step response. If a dimensionally homogeneous tuning rule is used,
we can expect the same 5% overshoot on the prototyped airplane. In this case equality
follows because overshoot is a dimensionless parameter.
It can be concluded that homogeneous tuning rules are more advantageous
than nonhomogeneous ones because, although both tuning procedures may yield
the same tuning performance, homogeneous tuning rules also preserve dimensional
similarity, whereas nonhomogeneous tuning rules do not.

6.2 Homogeneous PID tuning rules

6.2.1 Dimensionless controller parameters

Consider the following ideal form of PID controller:
 
1
C(s) = kC 1 + + τD s (6.1)
τI s

with the controller parameters kC , τI , and τD .

The dimensionless representation of the controller is given by following steps
(i)–(iii) presented in Chapter 3, and in step (iv), C(s̄) is divided by kC , thus
 
C(s̄) 1
C̄(s̄) := = 1+ + τ̄D s̄ (6.2)
kC τ̄I s̄

with dimensionless numbers defined as τ̄I := ττp1I , τ̄D := ττp1D , and s̄ := τp1 s.
The relation between the dimensional and dimensionless controller is given by

C(s, θC ) = kC C̄(s̄, C ) (6.3)

with θC = {kC , τI , τD } and C = {τ̄I , τ̄D }.
Note that the dimensionless controller numbers C have been obtained by the
use of plant parameter τp1 . This is mandatory in order to have the same dimension-
less variable s̄ as the plant, which allows the frequency scale compatibility between
dimensionless plant and dimensionless controller.

6.2.2 Homogeneous tuning rules characterization

In this section we characterize the function of tuning rules that are dimension-
ally homogeneous, so we can compare homogeneous and nonhomogeneous tuning
 Homogeneity of PID tuning rules 81

rules. We assume that the controller parameters depend on the plant parameters
θG := {k, h, τp1 , . . . , τpn , τc1 , . . . , τcm , l} and, possibly, on closed loop specifications
θλ . Examples of θλ are the closed loop time constant λ and the sensitivity function
peak Ms . The tuning rule fi , i = 1, . . . , 3, sets the controller parameters dependence
on model parameters θG and tuning parameters θλ

kC = f1 (θG , θλ ) (6.4)
τI = f2 (θG , θλ ) (6.5)
τD = f3 (θG , θλ ) (6.6)

We assume that the tuning-related parameters dimensions are either time dimen-
sion [θλ ] = T or dimensionless [θλ ] = 1. For instance [λ] = T and [Ms ] = 1. In
this way, parameters of set θG and θλ in (6.4)–(6.6) have only two distinct dimen-
sions, the gain dimension K and the time dimension T . Moreover, by definition
[G(s)] := [Y ]/[U ], where Y is output dimension and U is input dimension, and
[C(s)] := [U ]/[Y ]. As a result, [G(s)C(s)] = 1, that is dimensionless. Recall that
[G(s)] = [kτp1l
] and [C(s)] = [kC ], thus [kC kτp1
l
] = 1, that is, the controller gain kC
l
is made dimensionless multiplying by kτp1 .

Definition 6.1. A tuning rule is homogeneous if it can be written in dimensionless

form as follows:
l
kC kτp1 = 1 (G , λ ) (6.7)
τI
= 2 (G , λ ) (6.8)
τp1
τD
= 3 (G , λ ) (6.9)
τp1
Recall that the controller parameters dimensionless representation is based
on the homogeneity assumption. Although any valid physically based equation is
homogeneous, a tuning rule may be nonhomogeneous.
The following two examples show homogeneous and nonhomogeneous tuning
rules.

Example 6.1. The Ziegler–Nichols tuning rule [7] is

kC = 1.2τ/kh
τI = 2h
1
τD = h
2
It can be written in dimensionless form as
 −1
h
kC k = 1.2
τ
82 Application of dimensional analysis in systems modeling and control design

τI 2h
=
τ τ
τI 1h
=
τ 2τ

Example 6.2. On the contrary, the Calender et al. tuning rule [2, p. 78] is

1.066
kC =
kh
τI = 1.418h
τD = 0.47h

and kC k is no longer dimensionless but has inverse time dimension. As a result

the tuning rule is not homogeneous and cannot be written in dimensionless form.

Example 6.3. Table 6.1 presents the dimensionless form of generic PID tuning
rules for a first-order plus dead-time (FOPDT) process and for a second-order plus
dead-time (SOPDT) process. Moreover, the tuning rule for the SOPDT process also
considers the tuning parameter λ defined as the closed loop time constant. As a
result dimensionless tuning rules for FOPDT process depend only on one dimen-
sionless parameter, whereas dimensionless tuning rules for SOPDT process depend
on three dimensionless parameters. Note also that in both cases the plant gain only
affects the controller gain. The rest of controller parameters do not depend on the
plant gain.

Table 6.1 Homogeneous tuning rules for two distinct plants. Tuning
rules for the first-order plus dead-time (FOPDT)
processes depend only on one dimensionless parameter
h/τ . On the other hand, homogeneous tuning rules for
second-order plus dead-time (SOPDT) processes, with
closed loop time constant tuning parameter λ, depend on
three dimensionless parameters

Dimensionally homogeneous tuning rules functions

ke−hs ke−hs
C FOPDT = SOPDT =
(τ s + 1) (τ1 s + 1)(τ2 s + 1)
   
l h τ2 h λ
kτp1 kC 1 1 , ,
τ τ1 τ1 τ1
   
τI h τ2 h λ
2 2 , ,
τ τ τ1 τ1 τ1
   
τD h τ2 h λ
3 3 , ,
τ τ τ1 τ1 τ1
 Homogeneity of PID tuning rules 83

6.2.3 Dimensionless controller representation with homogeneous

tuning rules
Dimensional and dimensionless controllers tuned with homogeneous tuning rules
are related by

C(s, θC ) = kC C̄(s̄, G , λ ) (6.10)

The relationship follows by recalling that the general dimensionless controller form

is related to its dimensional counterpart by C(s, θC ) = kC C̄(s̄, C ). Moreover, as a
result of tuning rule homogeneity, we have C = i (G , λ ), i = 1, . . . , 3, and the
result follows.

6.3 Closed loop transfer functions

Once tuning rule homogeneity has been defined, we characterize the closed loop
transfer functions of homogeneous tuning rules. The objective is to obtain the dimen-
sionless form of closed loop transfer functions. Thereafter, it will be possible to
ascertain the homogeneity of tuning rules obtained by minimization of indexes
depending on closed loop transfer functions.

6.3.1 Loop transfer function

The following lemma establishes that the open loop transfer function G(s)C(s)
obtained from homogeneous tuning rules is a function of the plant dimensionless
numbers and of the specifications dimensionless numbers, if any.

Lemma 6.1. Given a transfer function G(s) and a PID controller C(s) with param-
eters kC , τI , and τD , given by an arbitrary homogeneous tuning rule defined in
dimensionless form by

l
kC kτp1 = 1 (G , λ ) (6.11)
τI /τp1 = 2 (G , λ ) (6.12)
τD /τp1 = 3 (G , λ ) (6.13)

the loop transfer function L(s) = G(s)C(s), which depends on the variable s and the
parameters θG and θC , that is L = L(s, θG , θC ), can be represented as a dimension-
less function depending on the normalized variable s̄ = τp1 s and the dimensionless
parameters G and λ , thus

L = L(s̄, G , λ ) (6.14)
84 Application of dimensional analysis in systems modeling and control design

Proof. The relation between dimensional and dimensionless model is G(s) =

l
kτp1 Ḡ(s̄, G ), and the relation between dimensional and dimensionless controller
is C(s) = kC C̄(s̄, G , λ ). It follows that

G(s)C(s) = kC kτp1
l
Ḡ(s̄, G )C̄(s̄, G , λ )
= 1 (G , λ )Ḡ(s̄, G )C̄(s̄, G , λ )
= L(s̄, G , λ ) (6.15)

In case of nonhomogeneous tuning rules the open loop transfer function L(s) is
a function of at least one dimensional parameter θG and/or θλ .

Example 6.4. Consider an FOPDT model. First of all, recall that a general FOPDT
transfer function can be rewritten by means of the dimensionless number h/τ

k h
G(s) = G(s̄) = e− τ s̄ (6.16)
(s̄ + 1)

The ideal PID controller C(s) can be written as

τI τD s 2 + τ I s + 1
C(s) = kC (6.17)
τI s

For any homogeneous tuning rule we have

 
h
kC = k −1 1 (6.18)
τ
 
h
τI = τ 2 (6.19)
τ
 
h
τD = τ 3 (6.20)
τ

Substituting the tuning rules structure of (6.18)–(6.20) into the PID transfer
function (6.17) we obtain
  2 h    
−1 h τ 2 τ 3 τh s2 + τ 2 τh s + 1
C(s) = k 1   (6.21)
τ τ 2 τh s

taking into account that s̄ = τ s, (6.21) yields

       
h 2 τh 3 τh s̄2 + 2 τh s̄ + 1
−1
C(s) = C(s̄) = k 1   (6.22)
τ 2 τh s̄
 Homogeneity of PID tuning rules 85

then multiplying G(s) by C(s) we obtain the open loop transfer function L(s) as
       
1 τh 2 τh 3 τh s̄2 + 2 τh s̄ + 1 −(h/τ )s̄
L(s) = L(s̄) =   e (6.23)
2 τh (s̄ + 1)s̄

which can also be simplified to

    h
h 3 τh s̄2 + s̄ + −1
e−(h/τ )s̄
2
L(s̄) = 1 τ
(6.24)
τ (s̄ + 1)s̄

Thus, the loop transfer function L(s) is a function of the dimensionless

variable h/τ .

6.3.2 Dimensionless closed loop transfer functions

Consider the control loop given in Figure 6.1. The reference r, input disturbance du ,
and output disturbance dy are the exogenous signals, whereas the error e, the control
action u, and the output y are the internal signals [1]. By definition internal signals
and exogenous signals are related by

e = r−y
u = du + Ce
y = dy + Gu

which can be written in matrix form as

⎡ ⎤⎡ ⎤ ⎡ ⎤
1 0 1 e r
⎣−C 1 0⎦ ⎣u⎦ = ⎣du ⎦ (6.25)
0 −G 1 y dy

du dy

r e y
u
C(s) G(s)
–

Figure 6.1 Control structure. Any relationship between exogenous and internal
signals is given by any of the following four closed loop transfer
functions, the sensitivity S(s), the complementary sensitivity T (s), the
control sensitivity SC (s), and the plant sensitivity SG (s). The closed loop
is completely determined by these four transfer functions
86 Application of dimensional analysis in systems modeling and control design

We are interested in describing all the internal signals as a function of exogenous

signals, thus

⎡ ⎤ ⎡ ⎤−1 ⎡ ⎤
e 1 0 1 r
⎣u⎦ = ⎣−C 1 0⎦ ⎣du ⎦ (6.26)
y 0 −G 1 dy

which results in

⎡ ⎤ ⎡ ⎤⎡ ⎤
e 1 −G −1 r
⎣u ⎦ = 1 ⎣C 1 −C ⎦ ⎣du ⎦ (6.27)
y 1 + GC GC G 1 dy

As a result, the relation between internal signals and exogenous signals is com-
pletely determined by the following four transfer functions that determine the closed
loop behavior:

(i) S = (1 + GC)−1 = er = dey = duu = y

dy
(ii) T = GC(1 + GC)−1 = yr
(iii) SC = C(1 + GC)−1 = ur = −udy
(iv) SG = G(1 + GC)−1 = −edu
= dyu

S is the sensitivity transfer function, T is the complementary sensitivity trans-

fer function, SC is the control sensitivity transfer function, and SG is the plant sensi-
tivity transfer function.
Recall that Ḡ = Ḡ(s̄, G ) and C̄ = C̄(s̄, C ) are the dimensionless model and
dimensionless controller, respectively, hence L(s̄) = Ḡ(s̄)C̄(s̄). Dividing by τp1 and
performing the change of variable s̄ → τp1 s, the previous closed loop transfer
functions are S = (1 + kC kτp1 l
L)−1 , T = kC kτp1 l
L(1 + kC kτp1 l
L)−1 , SC = kC C̄(1 +
kC kτp1 l
L)−1 , and SG = kτp1
l
Ḡ(1 + kC kτp1 l
L)−1 .
In case of homogeneous tuning rules it follows that C = i (G , λ ), i =
1, . . . , 3, and in particular kC kτp1 l
= 1 (G , λ ). As a result, the previous
closed loop transfer functions show the following parameters dependence: S(s) =
S(s̄, G , λ ), T (s) = T (s̄, G , λ ), SC (s) = kC S̄C (s̄, G , λ ), and SG (s) = kτp1
l
S̄G (s̄,
G , λ ).
The implication is that S(s) = S(s̄) and T (s) = T (s̄) because they are already
dimensionless. In fact, transfer functions S(s̄) and T (s̄) relate signals with the same
dimensions (reference and output), thus they are dimensionless. On the contrary this
is not the case for transfer functions SC (s̄) and SG (s̄), which have dimensional gain
because SC (s̄) and SG (s̄) relate signals with, in general, distinct dimensions (output
and input).
 Homogeneity of PID tuning rules 87

6.4 Optimality of homogeneous tuning rules

In this section we investigate the conditions of the scalar index I = I (S(s), T (s), SC (s),
SG (s)) which minimization yields homogeneous tuning rules. An example of a
scalar index is the minimization of the mixed sensitivity minC ||wS S; wT T ; wu Sc ||∞
[4, p. 63]. As a result we explain not only the prevalence of homogeneous tuning
rules but also their equivalent optimality properties when compared with nonhomo-
geneous tuning rules.

Lemma 6.2. Any tuning rule that minimizes a scalar index that depends on the loop
transfer function L(s) only, that is I = I (S(s)), I = I (T (s)), or I = I (S(s), T (s)), is
homogeneous.

Proof. The transfer function L(s̄) is a function of L(s̄) = L(s̄, G , C ). As a

result transfer functions S(s̄) and T (s̄), or any combination, can be expressed in
dimensionless form as functions of the dimensionless numbers G and C , that is

I = I (s̄, G , C ) (6.28)

Assume that the optimal solution is given by ∗C . From (6.28), ∗C depends only
on G , hence the optimal tuning rules are homogeneous.

Lemma 6.3. Any tuning rule that minimizes a scalar index that depends on the control
sensitivity I = I (SC (s)) or on the plant sensitivity I = I (SG (s)) is homogeneous.

Proof. First note that SG (s̄) = k S̄G (s̄, G , C ), and the constant term k does not
affect the minimization result. As a result the optimum ∗C is a function of G
only. In the case of SC (s̄) = kC S̄C (s̄, G , C ), multiplying by kτp1
l
does not affect
the optimal solution because kτp1 is also a constant, hence kτp1 SC (s̄) = kτp1
l l l
kC S̄C (s̄,
l
G , C ), and kτp1 kC ∈ C . Thus, the minimization problem is homogeneous.

Lemma 6.4. Tuning rules that minimize any scalar index that combines SC (s)
and SG (s) or any loop transfer function L(s) with SC (s) and/or SG (s) are not
homogeneous.

Proof. Assume that we want to minimize the index I = I (SC , SG ), that is

l
min I (kC S̄C (s̄, G , C ), kτp1 S̄G (s̄, G , C )) (6.29)
C

l
Multiplying (6.29) by constant kτp1 yields an equivalent minimization prob-
lem because the minimum is not modified by a constant, thus
l l
min kτp1 I (kC S̄C (s̄, G , C ), kτp1 S̄G (s̄, G , C )) (6.30)
C
88 Application of dimensional analysis in systems modeling and control design

Moreover, by the property of homogeneous functions [5] we have (6.30) equal to

l w l w l
min I ((kτp1 ) kC S̄C (s̄, G , C ), (kτp1 ) kτp1 S̄G (s̄, G , C )) (6.31)
C

If w = 1 then

min I (S̄C (s̄, G , C ), k 2 τp1

2l
S̄G (s̄, G , C )) (6.32)
C

which is not homogeneous.

If w  = 1 then (kτp1
l w
) kC ∈ l w
/ C , (kτp1 l
) kτp1 ∈ l w
/ G , and ((kτp1 l w
) kC )/((kτp1 l
) kτp1 )
∈
/ C . Thus, the minimization problem is not homogeneous.

6.4.1 Weighting factors

In the preceding section the homogeneity theorems were discussed without the
inclusion of weighting factors when constructing the indices to be minimized. How-
ever, the inclusion of weighting factors in the minimization indexes permits to suit
the optimization problems to be solved. The weighting factors used have a twofold
purpose: on the one hand, the numerical value that adjusts the function to be mini-
mized, and on the other hand, the units of the weighting factor that can set either a
homogeneous or a nonhomogeneous minimization problem.

Lemma 6.5. Any tuning rule that minimizes a homogeneous scalar index given by
I = I (L, αSC , βSG ), with α, β weighting factors with dimensions [α] = KT l and
[β] = (KT l )−1 , is given by

C = (G , ᾱ, β̄) (6.33)

with ᾱ = α/(kτp1
l
) and β̄ = βkτp1
l
.

Proof. After scaling s̄ = τp1 s it follows that [L(s̄)] = 1, that is dimensionless. In the
same way [αSC (s̄)] = 1 and [βSG (s̄)] = 1, hence

I = I (L(s̄, G , C ), αkC S̄C (s̄, G , C ), βkτp1

l
S̄G (s̄, G , C )) (6.34)

l
Now defining kC kτp1 = CK and recalling that CK ∈ C , (6.34) can be writ-
ten as
α
I = I (L(s̄, G , C ), l
l
CK S̄C (s̄, G , C ), βkτp1 S̄G (s̄, G , C )) (6.35)
kτp1

In (6.35) the dimensionless controller parameters C depend on G , ᾱ =

l
α/(kτp1 ), and β̄ = βkτp1
l
, which are dimensionless parameters too.
 Homogeneity of PID tuning rules 89
l
The theorem implication is that for distinct values of kτp1 , the homogeneous
tuning rules are the solution of minimization problems where the dimensional weight-
ing factors α and β are changed in order to keep their dimensionless counterparts
constant.
The result explains the prevalence of homogeneous tuning rules in the bibliog-
raphy as can be seen in Reference 2. First the most commonly used minimization
indexes yield homogeneous tuning rules by themselves. Second, any minimization
index can yield a homogeneous tuning rule by proper selection of weighting factors
units. Moreover, the minimization of homogeneous indexes can be achieved by means
of homogeneous tuning rules, as a result no optimality is lost by using homogeneous
tuning rules instead of the more general nonhomogeneous ones. In the next section
we derive the implications of using homogeneous tuning rules.

6.5 Homogeneous and nonhomogeneous tuning rules

For a single plant, homogeneous tuning can solve any minimization problem. As a
result, nonhomogeneous tuning rules provide no advantage. In fact any minimiza-
tion problem can be cast in dimensionless form by merely adding the appropriate
dimensional weighting factors. So we can conclude that there is no limitation on the
achievable performance if homogeneous tuning rules are used.
However, homogeneous and nonhomogeneous tuning rules are far from being
equivalent. The difference between homogeneity and nonhomogeneity arises when
investigating the tuning rules properties over sets of plants that share the similarity
property. Consider for instance the set of LTI plants G(s) of fixed but nevertheless
arbitrary structure as can be seen in Figure 6.2. The set is partitioned into subsets of
dimensionally similar plants. Each subset, due to similarity, has a single dimensionless
representation that is marked by a black dot.
Now consider the closed loop transfer functions obtained from each plant
by any homogeneous tuning rule. The homogeneous tuning rule maps the sub-
set of dimensionally similar plants to the subset of closed loop transfer function
Gcl (s) = {L, S, T , SC , SG } that is also similar, in such a way that yields a single dimen-
sionless closed loop representation. On the contrary, nonhomogeneous tuning rules
map dimensionally similar subsets G(s) into dimensionally nonsimilar sets on Gcl (s).
As a result the closed loop similarity is lost.
The use of homogeneous tuning rules is useful when, for instance, control sys-
tems are designed on the basis of models because experimental results can be easily
transferred to the prototype. This is not the case if nonhomogeneous tuning rules
are used. We remark that no optimality is lost in the use of homogeneous tuning
rules. The previous discussion is formalized in the following lemmas.

Lemma 6.6. Any homogeneous tuning rule maps the set of dimensionally similar
plants defined by its dimensionless number G into a set of dimensionally sim-
ilar closed loop systems, defined by the closed loop transfer functions Gcl (s) =
{L, S, T , SC , SG }.
90 Application of dimensional analysis in systems modeling and control design

Homogeneous tuning rule

Gcl(s)

Nonhomogeneous tuning rule

G(s)

Dimensionless representation

Set dimensionally similar transfer functions

Gcl(s)
Set nondimensionally similar transfer functions

Figure 6.2 Open loop–closed loop transformation. The set of open loop systems
G(s) (left) is partitioned into dimensionally similar subsets. The black
dot indicates the dimensionless representation. The set of closed loop
systems Gcl = {S(s), T (s), SC (s), SG (s)} (right) is also partitioned into
dimensionally similar subsets. Homogeneous tuning rules preserve
dimensional similarity. An infinite set of dimensionally similar plants is
transformed by homogeneous tuning rules into closed loops transfer
functions, which are also dimensionally similar. If the tuning rule is
nonhomogeneous the dimensionally similar plants are mapped into
closed loop transfer functions, which are no longer dimensionally
similar, hence there is no single dimensionless representation for the
closed loop transfer functions

Proof. The closed loop transfer functions are S = S(G , C ), T = T (G , C ),

SC = kC SC (G , C ), and SG = kSG (G , C ). Now, when a tuning rule is homo-
geneous C = (G , λ ), we can write S = S(G , λ ), T = T (G , λ ), SC =
kC SC (G , λ ), and SG = kSG (G , λ ). As a result the set of dimensionally sim-
ilar plants G has the same set of dimensionally similar closed loop transfer
functions.

Lemma 6.7. A nonhomogeneous tuning rule maps the set of dimensionally simi-
lar plants defined by its dimensionless number G into closed loop sets Gcl (s) =
{L, S, T , SC , SG }, which are not dimensionally similar among them.
 Homogeneity of PID tuning rules 91

Proof. If a tuning rule is nonhomogeneous then C = (θG , θλ ). As a result S =

S(θG , θλ ), T = T (θG , θλ ), SC = kC SC (θG , θλ ), and SG = kSG (θG , θλ ). As a result the
set of dimensionally similar plants G has not the same set of dimensionally similar
closed loop transfer functions.

Example 6.5. Consider the unstable time-delayed second-order transfer function

given by
k
G(s) = e−hs (6.36)
(1 − τ1 s)(1 + τ2 s)
We define two dimensionally similar transfer functions, the model Gm (s) and
the prototype Gp (s), related by the gain scaling factor Sk and the time scaling factor
St as follows:

k p = Sk k m
τ1p = St τ1m
τ2p = St τ2m
h p = St h m

We investigate the closed loop properties of the model and the prototype trans-
fer functions by PID control tuned with the homogeneous tuning rule of Rotstein
and Lewin [3] and the nonhomogeneous tuning rule of Wang and Jin [2,6]. In both
cases the tuning rule requires the desired closed loop time constant λ. The tuning rule
of Rotstein and Lewin [3] is


 
τ1 λ τλ1 + 2 + τ2
kC =
λ2 k
 
λ
τI = λ + 2 + τ2
τ1


λ τλ1 + 2 τ2
τD =
 (6.37)
λ τλ1 + 2 + τ2

By inspection, the tuning rule is homogeneous.

The tuning rule of Wang and Jin [2,6] is
kτ1 (τ1 − h)(τ2 + h)
kC =
(λ + h)3
λ3 + 3hλ2 + 3(τ1 τ2 + hτ1 − hτ2 )λ + hτ2 (τ1 − h)
τI =
(τ1 − h)(τ2 + h)
(τ2 + h − τ1 )λ3 + 3τ1 τ2 λ(λ + h) + τ1 τ2 h2
τD = (6.38)
τI (τ1 − h)(τ2 + h)
92 Application of dimensional analysis in systems modeling and control design

In this case the tuning rule is no longer homogeneous because the controller
gain has the same units as the plant gain. However, the integral and derivative time
are homogeneous relations.
In what follows km = 1, τ1m = 5, τ2m = 3, and hm = 0.4. The particular mea-
surement units are not important. The selected closed loop time constant of the model
is λm = 2.5. Model and prototype are related by the gain scaling factor Sk = 1.2 and
the time scaling factor St = 2. The closed loop step responses for the homogeneous
tuning of Rotstein and Lewin can be seen in Figure 6.3. As expected both model
and prototype step responses are dimensionally similar. This is shown by tempo-
rally stretching the model step response and seeing that it equals the prototype step
response.
Now consider the nonhomogeneous tuning rule of Wang and Jin applied to the
same model and prototype. Figure 6.4 shows the closed loop step response. In this

1.4
Model
Prototype
1.2 Time scaled model

0.8
Output

0.6

0.4

0.2

–0.2
0 5 10 15 20 25
Time

Figure 6.3 Homogeneous tuning rule step responses. When two distinct but
dimensionally similar plants, the prototype Gp (s) and the model Gm (s),
are controlled by proportional integral derivative (PID) control tuned
with the Rotstein and Lewin homogeneous tuning rule, the closed loop
behavior is also dimensionally similar. Consider the closed loop step
responses for the model (-) and the prototype (- -). Visually both figures
share the same shape. In fact, if the model response is time scaled (o)
by scaling factor St it is equal to the prototype time response. Thus both
closed loops are similar
 Homogeneity of PID tuning rules 93

1.6
Model
1.4 Prototype
Time scaled model

1.2

0.8
Output

0.6

0.4

0.2

–0.2
0 5 10 15 20 25
Time

Figure 6.4 Nonhomogeneous tuning rule step responses. When two distinct but
dimensionally similar plants, the prototype Gp (s) and the model Gm (s),
are controlled by proportional integral derivative (PID) control tuned
with the Wang and Jin nonhomogeneous tuning rule, the closed loop
behavior is no longer dimensionally similar. In fact at the sight of the
closed loop step responses for the model (-) and the prototype (- -), the
time scaled model response (o) is no longer equal to the prototype.
Moreover, both overshoots are distinct, hence dimensional similarity is
discarded

case both responses are no longer similar because they cannot be made equal by time
scaling. Note also that both step responses have distinct overshoot.
However, consider now a model and a prototype with the same time scaling
factor St = 2 but with new gain scaling factor Sk = 1, thus both model and pro-
totype have the same gain. Figure 6.5 shows the closed loop step response. Now,
both closed loop responses show again the similarity property, because the non-
homogeneity was caused by the controller gain. However, note that for distinct
gains similarity does no longer hold. As a conclusion nonhomogeneous tuning
rules map dimensionally similar plants into nondimensionally similar closed loops,
thus nonhomogeneous tuning rules do not preserve similarity. In this case prototype
controller design based on controlled model experiments is infeasible. Moreover,
the controller maintenance of families of dimensionally similar products that uses
nonhomogeneous tuning rules is made unnecessarily complex.
94 Application of dimensional analysis in systems modeling and control design

1.6
Model
1.4 Prototype
Time scaled model

1.2

0.8
Output

0.6

0.4

0.2

–0.2
0 5 10 15 20 25
Time

Figure 6.5 Nonhomogeneous tuning rule step responses. The nonhomogeneous

Wang and Jin closed loop step responses for the model (-) and the
prototype (- -) are in this case dimensionally similar as can be checked
by the time scaled model (o). The reason is that Wang and Jin tuning
rule is nonhomogeneous for the controller gain kC but homogeneous for
the integral τI and derivative time τD . In this case both model and
prototype have the same gain value, so the gain scaling factor Sk is
equal to unity. As a result the closed loops are dimensionally similar

References
[1] J.C. Doyle, B.A. Francis, and A.R. Tannenbaum. Feedback Control Theory.
Macmillan Publishing Company, 1992
[2] A. O’Dwyer. Handbook of PI and PID Controller Tuning Rules. Imperial
College Press, 2006
[3] G.E. Rotstein and D.E. Lewin. ‘Simple PI and PID tuning for open-loop
unstable systems.’ Industrial Engineering Chemistry Research, 30:1864–1869,
1991
[4] S. Skogestad and I. Postlethwaite. Multivariable Feedback Control. Analysis
and Design. Wiley, 1996
[5] T. Szirtes and P. Rózsa. Applied Dimensional Analysis and Modeling.
Butterworth-Heinemann, 2007
 Homogeneity of PID tuning rules 95

[6] H. Wang and X. Jin. ‘Direct synthesis approach of PID controller for second-
order delayed unstable processes.’ In Proceedings of the 5th World Congress on
Intelligent Control and Automation, Hangzhou, China, pages 19–23, 2004
[7] J.G. Ziegler and N.B. Nichols. ‘Optimum settings for automatic controllers.’
Transactions ASME, 64, 1942
 Chapter 7
Dimensionless PID tuning rules comparison

7.1 Introduction
The PID tuning is one of the most important issues in control theory and engineer-
ing due to the spread use of PID control in the industry [1]. In fact there are hundreds
of tuning rules, derived in many different ways and with distinct characteristics [6].
In the same way, there are even more industrial processes each one having its
own properties and particularities. Due to this variability it is well known that no
single tuning rule is desirable in all instances [5]. On the contrary, there are tuning
rules that are more suited for certain processes characteristics. As a result, the ques-
tion that arises to the control engineer is to decide which tuning rule is better for the
control engineering problem at hand. This topic of practical importance has been long
recognized although the received attention in the bibliography has been scarce [4].
In this chapter we propose a framework for PID tuning rules comparison by
using dimensional analysis. The proposal establishes the framework in which to
perform the analysis for any user-selected indexes. The key point enabling the com-
parison framework yields in the dimensionless representation of both transfer
function models and transfer function controllers. As a result it follows that the
indices used for comparison are dimensionless with the following benefits:
● The set of dimensionless parameters is the minimum size set in which compari-
son can be performed exactly. Any other parameters set must be of greater size,
thus complicating unnecessarily the comparison procedure.
● Each dimensionless parameter has a physical significance to be exploited in the
comparison process. Moreover, dimensionless parameters avoid the problem of
relative values of dimensional parameters. For instance, given an first order plus
dead time (FOPDT) model with delay h equal to 1 second, is the delay a control
issue? As stated, the question is meaningless because the answer depends on the
value h/τ , being τ the time constant.
● The fact that comparison indices are dimensionless also implies that the indices
no longer refer to a single model but to an infinite set of plants, the set of all
dimensional similar plants. As a result the comparison results increase generality.
● Finally, the extrapolation of comparative indices to the particular problem at hand
is straightforward, requiring a mere index scaling by a constant factor.
98 Application of dimensional analysis in systems modeling and control design

7.2 Elements of the comparative framework

Given the control structure of Figure 7.1, the plant G(s), and a set of PID tuning
rules, we want to assess and compare the performance of each tuning rule in the most
meaningful and economic way. The scope of the comparison procedure considers:
● Plants G(s): The set of linear time invariant (LTI) plants of arbitrary order.
For instance in Table 7.1 we present some of the most common process plants
used in control. This includes delayed second order plants minimum and non-
minimum phases. The complexity of each plant, measured as the number
of dimensionless numbers G , is also presented in Table 7.1. Note that for

du dy

r e y
u
PID (s) P (s)
−

Figure 7.1 Control structure

Table 7.1 Plants considered

Dimension Abbreviation Plant Dimensionless numbers

k
G ∈ R1 FOPDT e−hs h/τ
(τ s + 1)
k
SO τ2 /τ1
(τ1 s + 1)(τ2 s + 1)
k
G ∈ R2 SOPDT e−hs h/τ1 , τ2 /τ1
(τ1 s + 1)(τ2 s + 1)
k(− τ3 s + 1)
SOPIR τ2 /τ1 , τ3 /τ1
(τ1 s + 1)(τ2 s + 1)
k
TO τ2 /τ1 , τ3 /τ1
(τ1 s + 1)(τ2 s + 1)(τ3 s + 1)
k(− τ3 s + 1)
G ∈ R3 SOPIRDT e−hs h/τ1 , τ2 /τ1 , τ3 /τ1
(τ1 s + 1)(τ2 s + 1)
k
TOPDT e−hs h/τ1 , τ2 /τ1 , τ3 /τ1
(τ1 s + 1)(τ2 s + 1)(τ3 s + 1)
k(− τ4 s + 1)
TOPIR τ2 /τ1 , τ3 /τ1 , τ4 /τ1
(τ1 s + 1)(τ2 s + 1)(τ3 s + 1)
k
FO τ2 /τ1 , τ3 /τ1 , τ4 /τ1
(τ1 s + 1)(τ2 s + 1)(τ3 s + 1)(τ4 s + 1)
 Dimensionless PID tuning rules comparison 99

comparative purposes, it is equally complex to compare an FOPDT plant with

an second order (SO) plant.
● Controller tuning rules: Any controller tuning rule that can be expressed in
dimensionless form, that is any homogeneous tuning rule. Although there are
nonhomogeneous tuning rules, from the results of previous chapter and the lit-
erature review [6], the vast majority of existing tuning rules are homogeneous.
● Comparative indices: Any index derived from the open loop transfer function
C(s)G(s). This is very general because it includes the following indexes as
particular instances:
∞
– IAE = 0 |e(t)|dt
∞
– ISE = 0 e(t)2 dt
∞
– TV = 0 |u̇(t)|dt
√
– Sensitivity function bandwidth wB , |S( jwB )| = 1/ 2.
– Sensitivity function peak Ms = maxw |S( jw)|.

7.3 Dimensionless comparative framework

The dimensionless framework in which to perform the PID tuning rules comparison
is presented. The comparison procedure is as follows:
1. For a given model G(s) obtain the dimensionless transfer function Ḡ(s̄). Check
the values of the dimensionless numbers. Look for large or small dimensionless
numbers in order to assess their relative influence and possibly discard them
(e.g. relative fast poles).
2. For a given tuning rule obtain its dimensionless form by calculating the functions
1 , 2 , and 3 , as follows:
l
kC kτp1 = 1 (G , λ ) (7.1)
τI
= 2 (G , λ ) (7.2)
τp1
τD
= 3 (G , λ ) (7.3)
τp1
with τp1 the time constant chosen for obtaining the dimensionless transfer func-
tion Ḡ(s̄).
3. Obtain the dimensionless loop transfer function Ḡ(s̄)C̄(s̄).
4. Choose the indexes derived from the loop transfer function Ḡ(s̄)C̄(s̄). Note that
well-known indices as discussed above are obtained from Ḡ(s̄)C̄(s̄).
5. Calculate and plot the indexes mentioned above as a function of the plant
dimensionless numbers and the tuning parameters, if any. The set of dimen-
sionless number is the minimum possible set.
6. Compare and select the tuning rules by means of the previously calculated
indexes, using additional information on particular values of dimensionless
numbers.
100 Application of dimensional analysis in systems modeling and control design

7. Recall that the information is in dimensionless form. However, dimensional

values for indices comparison are extracted by mere scaling in order to obtain
their dimensional value.

7.4 Dimensionless elements

In this section we review and derive the form of the dimensionless elements requi-
red for the comparison framework.

7.4.1 Loop transfer function

As shown in Chapter 6, the open loop transfer function G(s)C(s) is a function of the
plant dimensionless numbers G and of the specifications dimensionless numbers
λ , if any, thus

G(s)C(s) = L(s̄, G , λ ) (7.4)

7.4.2 Dimensionless closed loop transfer functions

Recall also from Chapter 6 that the relation between dimensional and dimensionless
closed loop transfer function is
● S(s) = S(s̄, G , C )
● T (s) = T (s̄, G , C )
● SC (s) = kC S̄C (s̄, G , C )
● SG (s) = kτp1l
S̄G (s̄, G , C )

7.4.3 Dimensionless integral errors

The relation between integrals of dimensional and dimensionless functions is given
by the following lemma:

Lemma 7.1. Given the dimensional function f (t) with dimensionless form given by
Sk f¯ (t̄) with t = St t̄, and Sk , St constants, then the integral
 tf  t̄f
f (t)dt = Sk St f¯ (t̄)d t̄ (7.5)
t0 t̄0

with integration limits t̄0 = t0 /St and t̄f = tf /St .

Proof. The result follows by direct substitution of f (t) by Sk f¯ (t̄) and t by St t̄.
This general result is particularized to integral errors and to integral control
actions, which are well-known performance measures of controlled loops.
 Dimensionless PID tuning rules comparison 101

Integral errors
Given the closed loop error e(t) = r(t) − y(t), a classical performance measure is
given by the integration of distinct error functions f (e(t)), that is
 tf
f (e(t))dt (7.6)
t0

Depending on the function f (.) we have distinct integral errors such as

● IAE when f (e(t)) = |e(t)|
● ISE when f (e(t)) = e(t)2

Lemma 7.2. Given the dimensional error signal due to output disturbance, i.e.
e(t) = S(t)d(t), with d(t) the disturbance signal at the output, the value of the integral
error is given by
 tf  t̄f
f (e(t))dt = kd Stc+1 f (ē(t̄))d t̄ (7.7)
t0 t̄0

with kd the disturbance gain, c the number of integrators in S(s)d(s), and ē(t̄) the
dimensionless error ē(t̄) = L−1 {S̄(s̄)d̄(s̄)}.

The result provides, for a whole set of similar plants and similar disturbances,
the values of the dimensionless integral error as a function of the transfer function
dimensionless numbers ē(t̄). In order to obtain the dimensional value of the error
e(t) it is simply required to multiply ē(t̄) by kd Stc+1 , which only depends on the
disturbance properties.

Integral control actions

Another
∞ interesting integral refers to the variations of control actions TV =
0
|u̇(t)|dt, as it is a measure of the control activity level.

Lemma 7.3. Given the dimensional control action signal (i.e. u(t) = SC (t)d(t)), the
relation between dimensional and dimensionless integrals is
 tf  t̄f
|u̇(t)|dt = kd kC Stc−1 ˙ t̄)|d t̄
|ū( (7.8)
t0 t̄0

where kC is the controller gain, kd the disturbance gain, and c the number of integra-
tors in SC (s)d(s).

Proof. The proof follows by noting that u(t) = SC (s)d(s) and the dimensionless
form is given by ū(t̄) = kd kC Stc S¯C (s̄)d̄(s̄). Finally consider the change of variable
t = St t̄ in the integral and the result follows.
102 Application of dimensional analysis in systems modeling and control design

The only difference with the preceding case is that now the dimensional form
depends also on the controller gain kC .

7.4.4 Indexes
In this section the dimensionless indexes to perform the comparison are presented.
Despite the variety of possible indexes an important point is that the set of indexes
used should capture the fundamental trade-off of any control design. The dimension-
less indexes considered in the following are:
● Dimensionless crossover frequency w̄c . |Ḡ C̄( j w̄c )| = 1. √
● Dimensionless closed loop bandwidth w̄B . |S̄( j w̄B )| = 1/ 2. It is an indicator
of disturbance rejection performance. √
● Dimensionless complementary sensitivity bandwidth w̄BT . |T̄ ( j w̄BT )| = 1/ 2.
It is an indicator of tracking performance and noise mitigation.
● Peak of dimensionless sensitivity function M̄ . M̄ = maxw̄ |S̄( j w̄)|. It is an
indicator of closed loop robustness toward model uncertainty.
 t̄
● Dimensionless integral absolute error t̄0f |ē(t̄)d t̄|.
 t̄
● Dimensional variation of the control input kC t̄0f |ū( ˙ t̄)|d t̄.
However, it should be remarked that any index derived from the dimension-
less loop transfer function just depends on the plant and controller dimensionless
numbers.

7.5 Application example

In this section we compare tuning rules for FOPDT models that require a perfor-
mance tuning parameter, the desired closed loop time constant λ. The tuning rules
considered are:
● Skogestad. Analytic tuning rules are derived from internal model control (IMC)
considerations in Reference 8. In fact the derived controller for FOPDT models
is a proportional integral (PI). The objective is to obtain simple but still good
performing tuning. The tuning rules are
◦ kC = τ/k(λ + h)
◦ τI = min {h, 4(λ + h)}
◦ τD = 0
● Rivera et al. The PID tuning rule proposed in Reference 7 is based on the IMC
principle assuming an FOPDT model with the delay term approximated by a
first order Padé. The tuning rules are
◦ kC = (τ + 0.5h)/k(λ + 0.5h)
◦ τI = τ + 0.5h
◦ τD = τ h/(2τ + h)
● Dahlin. The tuning rules of Reference 2 are equal to the ones proposed by
Rivera et al. except for the kC term:
◦ kC = (τ + 0.5h)/k(λ + h)
◦ τI = τ + 0.5h
◦ τD = τ h/(2τ + h)
 Dimensionless PID tuning rules comparison 103

● Lee et al. In Reference 3 the tuning rules proposed are also derived from the
IMC principles.
◦ kC = τI /k(λ + h)
◦ τI = τ + h2 /2(λ + h)
◦ τD = h2 (1 − 3τhI )/2(λ + h)

7.5.1 PID tuning rules dimensionless characterization

In this section we characterize the above mentioned tuning rules by means of
dimensionless parameters h̄ = h/τ and λ̄ = λ/τ . Note also that l = 0, then
 
h λ
kC k = 1 ,
τ τ
 
τI h λ
= 2 ,
τ τ τ
 
τD h λ
= 3 ,
τ τ τ

In Table 7.2 we present the values kC k, τI /τ , and τD /τ for each one of the tuning
rules.
In what follows we perform the comparison of the dimensionless indices
considered.

7.5.2 Dimensionless sensitivity bandwidth comparison

In this section we compare the bandwidth provided by each one of the tuning rules
as a function of the controllability index and the control relaxation. The results can
be seen in Figure 7.2 and the following conclusions can be established:
● Skogestad, Rivera, and Dahlin provide very similar bandwidth values, despite
Skogestad proposes a PI and Rivera and Dahlin a PID.
● Lee et al. behave in a similar way as Skogestad, Rivera, and Dahlin but the
bandwidth values provided are lower.
In general there are neither qualitative nor quantitative differences among the
sensitivity function bandwidth provided by the four tuning rules.

Table 7.2 Dimensionless form of the PID tuning rules

Tuning rule 1 = kk C 2 = τ I /τ 3 = τ D /τ

Skogestad 1/(λ̄ + h̄) min [1, 4(λ̄ + h̄)] 0

Rivera (1 + 0.5h̄)/(λ̄ + 0.5h̄) 1 + 0.5h̄ h̄/(2 + h̄)
Dahlin (1 + 0.5h̄)/(λ̄ + h̄) 1 + 0.5h̄ h̄/(2 + h̄)
h̄(3 + 4(λ̄ + h̄))
Lee et al. (λ̄ + (3/2)h̄)/(λ̄ + h̄)2 1 + (h̄)2 /(2(λ̄ + h̄))
((λ̄ + h̄)(6(λ̄ + h̄) + 2h̄))
104 Application of dimensional analysis in systems modeling and control design

Skogestad Rivera

2 2

1 1
wb

wb
0 0
2 2
1.5 1.5
1 1 1 1
h/τ 0 0.5 λ/τ h/τ 0 0.5 λ/τ

Dahlin Lee et al.

2 2

1 1
wb

wb

0 0
2 2
1.5 1.5
1 1 1 1
h/τ 0 0.5 λ/τ h/τ 0 0.5 λ/τ

Figure 7.2 Sensitivity function bandwidth wb

7.5.3 Dimensionless sensitivity peak comparison

Figure 7.3 shows the sensitivity peak function of each tuning rule as a function of
the controllability index and the control relaxation. It can be seen that:
● Skogestad and Dahlin have a very similar behavior of the sensitivity function
peak.
● Rivera has a similar behavior for lower h̄ values than the other tuning rules but
as h̄ increases Rivera rapidly increases the sensitivity peak.
● Lee et al. provide in general the lowest sensitivity peak. However, for small λ̄
and h̄ it provides higher sensitivity peaks than the other tuning rules.

In summary, each tuning rule provides a high sensitivity function peak at distinct
points.

7.5.4 Dimensionless integral absolute error

Figure 7.4 shows the IAE for the tuning rules. The points of interest are:
● Dahlin provides in general the lower IAE for all λ̄ and h̄.
 Dimensionless PID tuning rules comparison 105

Skogestad Rivera

3 3

2 2
M

M
1 1

0 0
2 2
1.5 1.5
1 1 1 1
h/τ 0 0.5 λ/τ h/τ 0 0.5 λ/τ

Dahlin Lee et al.

3 3

2 2
M

M
1 1

0 0
2 2
1.5 1.5
1 1 1 1
h/τ 0 0.5 λ/τ h/τ 0 0.5 λ/τ

Figure 7.3 Sensitivity function peak M

Skogestad Rivera

4 4
IAE

IAE

2 2

0 0
2 2
1.5 1.5
1 1 1 1
h/τ 0 0.5 λ/τ h/τ 0 0.5 λ/τ

Dahlin Lee et al.

4 4
IAE

IAE

2 2

0 0
2 2
1.5 1.5
1 1 1 1
h/τ 0 0.5 λ/τ h/τ 0 0.5 λ/τ

Figure 7.4 Integral absolute error

106 Application of dimensional analysis in systems modeling and control design

Skogestad Rivera

15 15
10 10
TV

TV
5 5
0 0
2 2
1.5 1.5
1 1 1 1
h/τ 0 0.5 λ/τ h/τ 0 0.5 λ/τ

Dahlin Lee et al.

15 15
10 10
TV

TV

5 5
0 0
2 2
1.5 1.5
1 1 1 1
h/τ 0 0.5 λ/τ h/τ 0 0.5 λ/τ

Figure 7.5 Control action variability

● Rivera and Lee et al. provide the highest peaks of IAE. Rivera peaks for high h̄
and low λ̄ while Lee et al. peaks for high h̄ and high λ̄.
● Skogestad IAE increases almost linearly with h̄.

7.5.5 Dimensionless control action variation

Regarding the control action variability presented in Figure 7.5 we can establish that:
● The lowest TV corresponds to Skogestad. The other tuning rules provide higher
values.
● Dahlin tuning rule provides a nearly constant value of TV for all h̄ and λ̄.
● Rivera provides higher values of TV for higher h̄ and lower λ̄ while Lee et al.
provide high TV values for lower h̄ and λ̄.

7.6 PID tuning rules selection

In this section we present by means of an example the practical utility of the pro-
posed comparative framework in order to select the most appropriate tuning rule to
the control problem particularities.
 Dimensionless PID tuning rules comparison 107

Given the following FOPDT model:

1
G(s) = e−2s (7.9)
5s + 1

the controllability index is h/τ = 2/5 = 0.4. In Figure 7.6 the four indices consid-
ered are represented, that is the bandwidth wb , the sensitivity peak M , the absolute
integral error IAE, and the control action variability TV for each one of the tuning
rules considered.
At the sight of Figure 7.6 we can study two scenarios: first tunings with
λ/τ < 1, that is closed loops faster than the open loop dynamics. In this scenario, for
the indices considered, the most suited choice is the Skogestad tuning rule because it
provides a very low control action variability TV while keeping the rest of indices very
similar to Rivera and Dahlin tuning rules. In the second scenario with λ/τ > 1, that
is closed loop slower than open loop dynamics, Skogestad still provides the minimum
control variability value TV ; however, Lee et al. are able to provide a lower IAE.
As a result, depending on the considered cost function the choice would be between
Skogestad and Lee et al.

Sensitivity peak Bandwidth

2.5 1.4

1.2
2
1
M

w¯b

0.8
1.5
0.6

1 0.4
0.5 1 1.5 0.5 1 1.5
λ/τ λ/τ
Integral absolute error Control action variation
2.5 25

20
2
15
1.5
TV
IAE

¯
¯

10
1
5

0.5 0
0.5 1 1.5 0.5 1 1.5
λ/τ λ/τ

Figure 7.6 Dimensionless indices for ‘-’ Skogestad, ‘- -’ Rivera, ‘- ·’ Dahlin, and
‘·’ Lee et al.
108 Application of dimensional analysis in systems modeling and control design

References
[1] K.J. Åström and T. Hägglund. PID Controllers: Theory, Design and Tuning.
Instrument Society of America, 1995
[2] E.G. Dahlin. ‘Designing and tuning digital controllers.’ Instrumentaion
and Control Systems, 41(6):77–81, 1968
[3] Y. Lee, M. Lee, S. Park, and C. Brosilow. ‘PID controller tuning for desired
closed-loop responses for SISO systems.’ AIChE, 44(1):106–115, 1998
[4] A. Leva, C. Cox, and A.E. Ruano, ‘Hands-on PID autotuning: a guide to better
utilisation,’ ed. 1a , Austria: IFAC, 2002
[5] G.K. McMillan. Good Tuning: A Pocket Guide. The Instrumentation, Systems
And Automation Society, 2000
[6] A. O’Dwyer. Handbook of PI and PID Controller Tuning Rules. Imperial
College Press, 2006
[7] D. Rivera, M. Morari, and S. Skogestad. ‘Internal model control: 4. PID
controller design.’ Industrial Engineering Process Design and Developments,
25:252–265, 1986
[8] S. Skogestad. ‘Simple analytic rules for model reduction and PID controller
tuning.’ Journal of Process Control, 13(4):291–309, 2003
 Chapter 8
Control of dimensionally similar systems
with A. Claramonte

8.1 Introduction
The use of scaled models is prevalent in engineering applications because they allow
for experimental tests to be performed without requiring the physical system (i.e.
prototype) to be constructed. The model is built dimensionally similar with respect
to the prototype. Consequently, the experimental results on the model can be extrap-
olated to the prototype for analysis and design purposes [2,11]. Dimensional analysis
applications can be found in fluid mechanics [12], chemistry [13], and robotics. For
instance, dimensionally similar models have been used in robotics for optimum design
of jumping robots [8] and for study of energy efficiency of walking machines [9].
The analysis, evaluation, and design of control systems can also be benefitted
by using dimensionally similar models. The use of models permits for experimental
tests to be performed when the prototype does not exist or it is difficult or dangerous
to perform experiments on it. Moreover, there are families of manufactured prod-
ucts that, by construction, share some similarity relation. From a control perspective
the objective is to find an expedient way for analysis, design, and maintenance of
controllers for the whole family of products.
In the bibliography we may find particular applications of dimensional analysis
to control design on vehicle dynamics [4] and robotics [7]. In Reference 4 a scaled
testbed is constructed for vehicle control application, showing the dynamic simi-
larity with actual vehicles. The design of a robust controller for distinct vehicles is
investigated in Reference 5 from a dimensionless approach. Another contribution is
presented in Reference 7 where scaling laws for continuous time and discrete time con-
trollers are developed for a single flexible link manipulator. The approach taken in all
cases considers dimensionless numbers derived from physical parameters. However,
it is difficult to generalize this approach because on the one hand the dimensionless
numbers depend on physical parameters such as inertias or capacitances, and, on the
other hand, the number of parameters forming the pi groups is problem dependent.
However, the case is different for the transfer function framework. In the trans-
fer function approach physical parameters are transformed into gains, time delays,
poles, and zeroes time constants [1]. In this way, only two scaling factors are
required to relate two dimensionally similar transfer functions, the gain scaling fac-
tor Sk , and the time scaling factor St . Although we always have to consider physical
110 Application of dimensional analysis in systems modeling and control design

parameters when applying dimensional analysis and similarity, once dimensional

similarity is established, the transfer function framework greatly simplifies and
generalizes the scaling of controllers.
The analysis and design of control systems in the transfer function framework
are pervasive in control applications because normally the system state is not fully
measurable and dynamical controllers are required. In this case, for continuous time
dimensionally similar systems controlled by means of digital controllers, a simi-
lar performing control level can be achieved by modifying the controller in two
straightforward ways: (i) by adjusting the controller gain and (ii) by adjusting the
controller sampling time. Moreover, the adjusting factors are directly given by the
system scaling factors.
However, the construction of fully similar models is costly and sometimes it
is not even possible [12]. In this case we no longer have complete similarity but
partial similarity. In this chapter we show how partial similarity can be treated using
robust control theory by modeling the lack of similarity as uncertainty.

8.2 Control of dimensionally similar systems

Consider a model defined mathematically by the transfer function Gm (sm ), which is
dimensionally similar to a prototype Gp (sp ). Assume that the controller Cm (sm ) has
been designed and tested on the model and that it achieves a desired closed loop
performance. The point to be investigated is, based on the knowledge of Cm (sm ), the
determination of the controller Cp (sp ) applied to the prototype (see Figure 8.1) that
achieves a similar closed loop performance.
The second problem to be considered is the assessment of the resulting proto-
type controller robustness properties when the dimensional similarity is partial. In

w(sm) ∆(sm)

rm ym
Cm(sm) Gm(sm)
−

rp yp
Cp(sp) Gp(sp)

Figure 8.1 Model and prototype control loops

 Control of dimensionally similar systems 111

fact there are instances that it is very difficult and even impossible to design a model
that is completely similar to a prototype. The best that can be achieved is partial
similarity. In this case we may consider that the model is uncertain, represented as
multiplicative uncertainty in Figure 8.1.

8.3 Complete similarity

In this section we consider the problem of controller design for systems that follow
a similarity relation, when complete dimensional similarity exists. If we design a
controller in order to fulfill some performance level on a model, then the question
is how the currently existing model controller should be modified in order to achieve
a similar performance level on the prototype system.

8.3.1 Continuous time control

The model-prototype controllers equivalence is presented by the following theorem:

Theorem 8.1. Consider two continuous time transfer functions Gp (sp ) and Gm (sm ).
Assume that Cm (sm ) is the continuous time model controller achieving the desired
transfer function Lm (sm ) = Gm (sm )Cm (sm ). If the prototype controller Cp (sp ) is
obtained as Cp (sp ) = Sk−1 St−l Cm (St sp ) then, if Gp (sp ) and Gm (sm ) are continuous time
similar, then the open loop transfer functions of the model and the prototype are
similar, that is Lm (s̄) = Lp (s̄).

Proof. In case of continuous time similarity we have

Lp (sp ) = Gp (sp )Cp (sp )
= Sk Stl Gm (St sp )Sk−1 St−l Cm (St sp )
= Gm (St sp )Cm (St sp )
= Lm (St sp )
= Lm (sm ) (8.1)
This implies that Lp (s̄) = Lm (s̄), that is model and prototype open loop transfer
functions are equal except for a time scaling factor.
In words Theorem 8.1 states that the only modifications to be performed to an
existing model controller are to modify its gain with Sk−1 St−l and to change the model
controller Laplace variable by St sp .

8.3.2 Discrete time control

The model-prototype controllers equivalence on sampled data control is presented by
the following theorem:
112 Application of dimensional analysis in systems modeling and control design

Theorem 8.2. Consider two sampled data transfer functions Gp (zSD ) and Gm (zSD ),
with sampling times (Ts )p and (Ts )m , respectively. Assume that Cm (zSD ) is the
discrete time model controller with sampling time (Ts )m achieving the desired trans-
fer function Lm (zSD ) = Gm (zSD )Cm (zSD ). Now if the prototype controller Cp (zSD ) is
obtained as Cp (zSD ) = Sk−1 St−l Cm (zSD ) with sampling time (Ts )p = St (Ts )m , then if
Gp (zSD ) and Gm (zSD ) are sampled-data similar then Lp (zSD ) = Lm (zSD ).

Proof. In the case of discrete time similarity we have Gp (z) = Sk Stl Gm (z), so the loop
transfer function Lp (z) can be written as

Lp (z) = Gp (z)Cp (z)

= Sk Stl Gm (z)Sk−1 St−l Cm (z)
= Gm (z)Cm (z)
= Lm (z) (8.2)

Then the discrete transfer functions Lm (z) and Lp (z) are equal in a strict discrete
time sense. Moreover, due to sampled data similarity it holds that the dimensionless
number zSD is also equal in the model and the prototype, that is zSD = esp (Ts )p =
esm (Ts )m .

Theorem 8.2 states that the only modifications to be performed to an existing

controller are to modify its gain by means of the inverse gain scaling factor Sk−1 St−l
and to scale the prototype controller sampling time by the time scaling factor St .

Application example
Consider two continuous transfer functions, the model and the prototype, given by

(15s + 1)(0.005s + 1)
Gm (s) =
(3.5s + 1)(0.5s + 1)(0.1s + 1)
5(24s + 1)(0.008s + 1)
Gp (s) =
(5.6s + 1)(0.8s + 1)(0.16s + 1)

We have five time constants (three poles and two zeros) and the dimensionless
numbers are τp2 /τp1 , τp3 /τp1 , τz1 /τp1 , τz2 /τp1 , and τp1 s.
In Table 8.1 the dimensionless numbers for the model and the prototype are
shown. In can be seen that τp2 /τp1 , τp3 /τp1 , τz1 /τp1 , and τz2 /τp1 are equal, so we have
complete similarity. The time scaling factor is given by St = (τp1 )p /(τp1 )m = 5.6/
3.5 = 1.6. The gain scaling factor is given by Sk = kp /km = 5.
Next, we calculate the zero order hold equivalent discrete transfer function of
the model Gm (z) and of the prototype Gp (z). The model sampling time selected
for the model is (Ts )m = 0.02 s. The prototype sampling time is calculated as
(Ts )p = St (Ts )m = 1.6 · 0.02 = 0.032 s in order to obtain similarity in the discrete
 Control of dimensionally similar systems 113

Table 8.1 Dimensionless numbers of model and prototype

Dimensionless number Model Prototype

τp2 /τp1 0.5/3.5 = 0.14 0.8/5.6 = 0.14

τp3 /τp1 0.1/3.5 = 0.03 0.16/5.6 = 0.03
τz1 /τp1 15/3.5 = 4.28 24/5.6 = 4.28
τz2 /τp1 0.005/3.5 = 0.001 0.008/5.6 = 0.001
τp1 s 3.5sm 5.6sp

time transfer functions. The resulting discrete time transfer function for the model
Gm (z) is given by

0.02341z 2 − 0.01639z − 0.006977

Gm (z) = Sk−1 Gp (z) = (8.3)
z 3 − 2.774z 2 + 2.556z − 0.7821
On the basis of the above discrete time model we design a minimum time
controller for the model Cm (z), given by

2.469 · 104 z 3 − 6.849 · 104 z 2 + 6.311 · 104 z − 1.931 · 104

Cm (z) = , (Ts )m = 0.02
z 3 − 578z 2 + 404.7z + 172.3
(8.4)

The controller for the discrete time prototype is given by Theorem 8.2 resulting in

4938z 3 − 1.37 · 104 z 2 + 1.262 · 104 z − 3863

Cp (z) = , (Ts )p = 0.032 (8.5)
z 3 − 578z 2 + 404.7z + 172.3
In Figure 8.2 the outputs of the model and the prototype for a step input in
the reference can be seen. In both cases it takes three samples to make the output
to follow the reference, as expected. We can conclude that Cp (z) is also minimum
time controller for the prototype Gp (z).

8.4 Partial similarity

Under partial similarity, model and prototype have distinct dimensionless represen-
tation (i.e. Ḡm (s̄)  = Ḡp (s̄)) because some dimensionless numbers have distinct value
in the model and the prototype. For the sake of clarity we assume that model and
prototype gains are known. In case both gains are not perfectly known the same
approach can be used but adding an uncertain gain scaling factor. Moreover, time
delays on the model and prototype are converted to rational expressions by Padé
approximation.
The uncertainty is characterized as multiplicative dynamic uncertainty [10]
because normally the dimensionless numbers related with dominant time constants
114 Application of dimensional analysis in systems modeling and control design

Step response
600
Model step response
Prototype step response

500

400

300

200

100

0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35
Time (s)

Figure 8.2 Model and prototype closed loop step responses

are similar, whereas dimensionless time constants related to fast dynamics are no
longer similar. Moreover, dynamic uncertainty considers high-frequency unmodeled
dynamics and it is computationally more tractable than the parametric counterpart [3].
Henceforth the uncertain model is

Ḡm (s̄) = Ḡm0 (s̄)(1 + W (s̄)Δ(s̄)) (8.6)

where the weight W (s̄) is in general a rational transfer function and Δ(s̄) is any
stable transfer function, which at any frequency is not larger than 1 in magnitude (i.e.
|Δ( j w̄)| ≤ 1, ∀w̄). The weight W (s̄) can be calculated as an upper bound of l(w̄),
that is |W ( j w̄)| ≥ l(w̄), ∀w̄, where l(w̄) is defined as
 
 Ḡp − Ḡm0 

l(w̄) = max   (8.7)
Ḡ 
m0

Consequently, we have the nominal model Gm0 (s̄) together with the uncer-
tainty description |W (j w̄)| ≥ l(w̄), ∀w̄, so robust analysis and design procedures
[10] can be applied. Finally we show that when a controller achieves robust stabil-
ity (RS) and robust performance (RP) on the uncertain dimensionless model, it also
achieves RS and RP on the dimensional prototype.
 Control of dimensionally similar systems 115

Theorem 8.3. If the controller designed on the model Cm0 (s̄) achieves RS or RP on
Gm (s̄) = Gm0 (s̄)(1 + W (s̄)(s̄)), then the controller Cp (s̄) = Sk−1 St−l Cm0 (s̄) achieves
RS or RP on Gp (s̄).

Proof. Assume that Cm0 (s̄) achieves RS, then W (s̄)Tm (s̄)∞ < 1. The comple-
mentary sensitivity function is Tm (s̄) = Gm0 (s̄)Cm0 (s̄)(1 + Gm0 (s̄)Cm0 (s̄))−1 . Now by
multiplying and dividing by Sk Stl and recalling that Gp (s̄) = Sk Stl Gm0 (s̄) and Cp (s̄) =
Sk−1 St−l Cm0 (s̄) we have Tp (s̄) = Gp (s̄)Cp (s̄)(1 + Gp (s̄)Cp (s̄))−1 . Finally the change of
variable s̄ = St sp , which is a time scaling: does not modify the magnitude then
W (s)T p (s)∞ < 1. The RP proof follows a similar reasoning and it is omitted.

8.5 Experimental case study

The ‘ecp Industrial Plant Emulator’ [6] is a three-shaft system where drive and load
inertias can be adjusted (see Figure 8.3). The transfer function between the angular
output velocity and the motor drive voltage input is given by

θ̇ F
= (8.8)
V Js + 1

where F is the dynamic friction and J is the total system inertia.

Figure 8.3 Physical model configu ation

116 Application of dimensional analysis in systems modeling and control design

Figure 8.4 Physical prototype configu ation

We consider two systems, the model shown in Figure 8.3 and the prototype
shown in Figure 8.4. The model has two weights: the driver pulley (the small one)
and the load pulley, which fix the model inertia Jm .
The prototype is physically configured by adding two extra weights to each
pulley, so the drive pulley and the load pulley have now four weights each (see
Figure 8.4). As a result the prototype increases the system inertia with respect to
the model.
First we identify model and prototype inertia by a simple step response with
step magnitude u = 0.6 V, yielding Jm = 10.2 s, and Jp = 16.25 s. On the other
hand, model and prototype gains are obtained by closed loop identification
with a proportional controller kC = 0.1 yielding Fm = 115.93 rad/sV and Fp =
112.64 rad/sV. The reason is that low step voltages rapidly saturate the motor drive
and a closed loop experiment permits to identify the gains more accurately. The result
confirms that model and prototype gains hardly vary because they are related to the
dynamic friction. As a result, the gain scaling factor is Sk = Fp /Fm = 1.029 and the
time scaling factor is St = Jp /Jm = 1.59.
Once the model is identified, a PI controller is designed for the model with
settling time equal to 1 s and damping ratio of 0.7, which yields

1 + 0.26s
Cm (s) = 2.66 (8.9)
s
 Control of dimensionally similar systems 117

25
Model output
20 Prototype output
15
Output

10

0
0 0.5 1 1.5 2 2.5 3 3.5 4
Time (s)

10
8
Control action

6
4
2
0
−2
0 0.5 1 1.5 2 2.5 3 3.5 4

Figure 8.5 Model and prototype outputs and control actions

The controller is discretized by zero order hold, yielding

0.6916z − 0.69
Cm (z) = (8.10)
z−1

The controller is digitally implemented on the computer, with sampling time

of TSm = 0.0006 s. The model controlled output and model control action can be seen
in Figure 8.5, achieving the design specifications.
Next we tackle the problem of prototype controller design. The prototype con-
troller Cp (s) is obtained by transformation of the model controller Cm (s). Recall
that if the discrete prototype controller is equal to Cp (zSD ) = Sk−1 Cm (zSD ) with
(Ts )p = St (Ts )m , then model and prototype closed loops are similar. Because SK ≈ 1
and St = 1.59, we implement the same discrete digital controller (8.10) with the
only modification being that the sampling time is set to (Ts )p = 0.00096 s.
The prototype controlled output and prototype control action are shown in
Figure 8.5. Now the prototype settling time is of 1.6 s, as expected. Moreover,
the damping ratio is kept equal to 0.7 because the damping ratio is a dimensionless
parameter, and thus equal under similarity.
We empirically show that model and prototype control outputs and control
actions are similar. Recall that two outputs are similar if the only transformations
required to make them equal are a gain scaling and a time scaling. Due to the fact
118 Application of dimensional analysis in systems modeling and control design

25
Model output
20 Prototype output

15
Output

10

0
0 0.5 1 1.5 2 2.5 3 3.5 4
Time (s)

10
8
Control action

6
4
2
0
−2
0 0.5 1 1.5 2 2.5 3 3.5 4

Figure 8.6 Prototype time scaled output

Sk ≈ 1 just a time scaling by a factor of St = 1.59 is required. Figure 8.6 shows

the model input and output together with the prototype input and output time scaled
by tp = St−1 tm . The results show that in fact model and prototype closed loops are
similar.
Finally we assume that there is partial similarity and consider that the proto-
type inertia is not perfectly known, that is Jm = (1 ± 10%)Jp . The dimensionless
nominal model and the dimensionless prototype are

1 1
Ḡm (s̄) = , Ḡp (s̄) = (8.11)
s̄ + 1 (1 ± 10%)s̄ + 1

The uncertainty is described by l(w̄).

 
 (s̄ + 1) 
l(w̄) = max  − 1 (8.12)
∈±10% s̄ + 1

Figure 8.7(a) shows the magnitude of 20 combinations of the uncertain l(w̄),

together with the upper bound error weight given by the first order transfer function

0.1s
W (s) = (8.13)
s+1
 Control of dimensionally similar systems 119

(a) Bode diagram

−20
−30
Magnitude (dB)

−40
−50
−60
−70
−80
−90
−100
10−2 10−1 100 101 102 103 104 105
Frequency (rad/s)
(b) Bode diagram
−20
−30
Magnitude (dB)

−40
−50
−60
−70
−80
−90
−100
10−2 10−1 100 101 102 103 104 105
Frequency (rad/s)

Figure 8.7 Relative error for 20 combinations of uncertain zero. ‘+’ line is the
magnitude of the fi st order weight

The condition of robust stability W (s)T (s)∞ < 1, with T (s) = Gm (s)Cm (s)
(1 + Gm (s)Cm (s))−1 , is accomplished as can be seen in Figure 8.7(b). As a result
the controller Cp (s) does guarantee robust stability when applied on the prototype
under partial similarity.

References
[1] P. Balaguer, C. Pedret A. Ibeas, and S. Alcántara. ‘Controller parameter depen-
dence on model information through dimensional analysis.’ In 48th Conference
on Decision and Control, Shanghai, 2009
[2] G.I. Barenblatt. Dimensional Analysis. Gordon and Breach Science Publish-
ers, 1987
[3] S.P. Bhattacharyya, H. Chapellat, and L.H. Keel. Robust Control. The
Parametric Approach. Prentice Hall, 1995
[4] S. Brennan and A. Alleyne. ‘The Illinois roadway simulator: a mecha-
tronic testbed for vehicle dynamics and control.’ IEEE/ASME Transactions
on Mechatronics, 5:349–359, 2000
120 Application of dimensional analysis in systems modeling and control design

[5] S. Brennan and A. Alleyne. ‘Dimensionless robust control with applications

to vehicles.’ IEEE Transactions on Control Systems Technology, 13:624–630,
2005
[6] ecp Educational Control Products. Industrial Plant Emulator Model 220,
http://www.ecpsystems.com/controls_emulator.htm, 2008
[7] M. Ghanekar, D.W.L. Wang, and G.R. Heppler. ‘Scaling laws for linear
controllers of flexible link manipulators characterized by nondimensional
groups.’ IEEE Transactions on Robotics and Automation, 13:117–127, 1997
[8] M. Higashimori, M. Harada, M. Yuya, I. Ishii, and M. Kaneko. ‘Dimensional
analysis based design on tracing type legged robots.’ In Proceedings of the
2005 IEEE International Conference on Robotics and Automation, Barcelona,
Spain, 2005
[9] M. Kaneko, S. Tachi, K. Tanie, and M. Abe. ‘Basic study on similarity in
walking machine from a point of energetic efficiency.’ IEEE Journal of
Robostics and Automation, RA-3:19–30, 1987
[10] S. Skogestad and I. Postlethwaite. Multivariable Feedback Control. Analysis
and Design. Wiley, 1996
[11] T. Szirtes and P. Rózsa. Applied Dimensional Analysis and Modeling.
Butterworth-Heinemann, 2007
[12] F.M. White. Fluid Mechanics. Mcgraw-Hill College, 4 Har/Dsk edition, 1998
[13] M. Zlokarnik. Scale-up in Chemical Engineering. Wiley-VCH, 2002
 Chapter 9
Adaptive systems

9.1 Introduction
Adaptive control systems are used to overcome the problem of plant uncertainty
due to, for instance, slowly varying time parameters [4]. One approach to adaptive
control is gain scheduling, where an appropriate controller is chosen on the basis of
scheduling variables that are continuously observed [2]. The complexity of a gain
scheduling controller increases with the number of scheduling parameters. Dimen-
sional analysis may reduce the problem complexity formulation by using a dimen-
sionless representation that reduces the number of scheduling parameters, as shown
in Reference 3.
Another common technique for adaptive systems is the model reference adap-
tive control (MRAC). In this case the controller parameters are continuously updated
in order to keep a fixed closed loop performance, specified by a reference model [2].
Despite the fact that a constant reference model is common practice, better MRAC
schemes can be achieved by considering time varying reference models during oper-
ation. The motivation for a time varying reference model can be better understood
by means of an example. Consider the adaptive attitude control of a spacecraft. Due
to fuel consumption the axis inertia moments diminish with time. On the other hand,
when the fuel is refilled in the spacecraft (assume it can be done during normal oper-
ation) the spacecraft inertia increases. In this case it could not be sensible to have a
constant reference model because:
● Due to inertia increase, it can be the case that too much fuel is expended or even
actuators can be saturated if closed loop dynamics are too demanding.
● In case of inertia decrease it could be possible to achieve a faster closed loop
performance level, more suited for landing maneuvers.
As a result it can be difficult to choose a valid reference model with suited
dynamics for the whole spacecraft inertia operation range and it could lead either
to conservative results when low inertia is present or to actuator saturation and fuel
wasting under the high inertia condition.
A more convenient approach is to have a variable reference model depending on
the spacecraft inertia. In this case it is possible to set faster closed loop dynamics
when the inertia is small with no controller saturation. On the other hand, when the
spacecraft has high inertia, the reference model dynamics are set to slower values in
order to prevent actuator saturation.
122 Application of dimensional analysis in systems modeling and control design

In this chapter we show how dimensional analysis concepts are used to derive a
new model reference adaptive control with the following features: (i) the reference
model is no longer constant but time varying: (ii) there is a bounded control effort,
measured as the shift between the current system poles and the reference model poles.
Finally (iii) the reference model variation is no longer arbitrary but follows a simi-
lar pattern. In this way although the reference model is no longer constant, several
properties of the reference model are still kept constant. In fact the reference model
describes similar dynamics.
The similar model reference adaptive control (SMRAC) is one approach to
MRAC schemes in which actuator wearing and saturation are concerns. The SMARC
guarantees a bounded control effort during operation at the cost of modifying the
reference model but maintaining output dimensionless properties constants.

9.2 Actuator limitations and dimensionally similar

model reference

9.2.1 Control effort

Normally, the actuator limitations are set as bounds on the control signal magni-
tude or its rate of change. However, we introduce a new way of setting the actuator
capacity by means of the control effort concept. The control effort c, as defined in
Reference 1, is the shift of the poles from the open to the closed loop system


n
c=− (bi − ai ) (9.1)
i=1

where n is the system order, and bi and ai are the closed loop poles and the open
loop poles, respectively. Equation (9.1) can be rewritten as

c = Bn − An (9.2)
 
with Bn = − ni bi and An = − ni ai .
The relation between the closed loop poles and control effort can be classified
accordingly to its sign as
● c > 0. In this case the closed loop poles are faster than the open loop poles.
● c = 0. The closed loop poles are equal to the open loop poles.
● c < 0. The closed loop poles are slower than the open loop poles.
In the MRAC the number of open loop poles (i.e. plant poles) is not necessarily
equal to the number of closed loop poles (i.e. the model reference poles) because
the denominators of the plant and the reference model are not necessarily equal. In
this case a new definition of the control effort is required and stated as follows:

c = Bnm − Anp (9.3)

 Adaptive systems 123
 n
with Bnm = − ni m bi and Anp = − i p ai , where bi and ai are poles of the closed
loop and the poles of the open loop, respectively. In this case the above classification
is no longer valid as the control effort can be equal to zero with closed loop poles
distinct to the open loop poles. Anyway, it is true that any relative increase in the
closed loop poles stability comes by an increase in the control effort, for constant Anp .

9.2.2 Similar model reference adaptive control

Dimensionally similar models
Recall that if two transfer functions, the model Gm (sm ) and the prototype Gp (sp ), are
dimensionally similar, then they are related as

Gp (sp ) = Sk Stl Gm (St sp ) (9.4)

where Sk and St are the gain and time scaling factors, respectively, and l is the
number of plant integrators. In the same way, given a nominal transfer function
G0 (s0 ), any transfer function is derived from G0 (s0 ) by
1. Multiplying G0 (s0 ) by the scaling factors Sk Stl
2. Substituting the variable s0 by St s, that is s0 → St s
is a transfer function dimensionally similar to G0 (s0 ), for any Sk > 0 and St > 0.
In general, given an nth-order nominal transfer function G0 (s0 ) with relative degree
equal to n∗

k0 (s0 + a0n+1 ) . . . (s0 + a0n+m )

G0 (s0 ) = (9.5)
(s0 + a01 ) . . . (s0 + a0n )

G(s) is similar to G0 (s) if

k(s + an+1 ) . . . (s + an+m )

G(s) = (9.6)
(s + a1 ) . . . (s + an )

with
∗
k = Sk Stl−n k0
ai = St−1 a0i , i ∈ [1, . . . , n + m]
Sk > 0
St > 0 (9.7)

In words, the G(s) gain is scaled by the gain scale factor Sk and by n∗ times the
inverse time scaling factor. The poles and zeroes of G(s) are scaled by the inverse of
the time scaling factor. Note that St > 1 means that the system G(s) is slower than
G0 (s). On the contrary, St < 1 means that the system G(s) is faster than G0 (s).
124 Application of dimensional analysis in systems modeling and control design

Now two important lemmas are established for similar transfer functions.

Lemma 9.1. If transfer function G0 (s) is coprime, then any transfer function G(s)
obtained by similar transformation from G0 (s) is also coprime.

Proof. The result follows because similar transformation scales all poles and zeroes
by the same positive scaling factor St−1 .

Lemma 9.2. If G0 (s0 ) is a strictly positive real (SPR) transfer function, then any
transfer function G(s) obtained by similar transformation from G0 (s0 ) is also SPR.

Proof. By definition (see for instance Reference 5), G0 (s0 ) is SPR if and only if
(i) G0 (s0 ) is a strict stable transfer function and (ii) the real part of G0 (s0 ) is strict
positive along the jw0 axis, that is, ∀w0 ≥ 0 Re[G0 (jw0 )] > 0. If G(s) is derived from
G0 (s0 ) by a similarity transformation it follows from the preceding discussion that
∗
G(s) = Sk Stl−n G0 (s0 → St s).
∗
The multiplication by the constant Sk Stl−n keeps the poles unaltered. The vari-
able substitution s0 → St s means that the poles (and also the zeroes) of G(s) are
scaled by the inverse of the time scaling factor. As St > 0 the stability properties are
unchanged. As a result if G0 (s0 ) is strictly stable, so is G(s).
By assumption ∀w0 ≥ 0 Re[G0 (jw0 )] > 0. It then follows that ∀St w0 ≥ 0
∗
Re[G0 ( jSt w0 )] > 0, for St > 0. Moreover, as Sk > 0 ∀St w0 ≥ 0 Sk Stl−n Re[G0 ( jSt w0 )]
> 0, which by similarity is analogous to ∀w ≥ 0 Re[G( jw)] > 0.

Adaptive scheme
Let an nth-order linear time invariant (LTI) plant with input–output pair {up (t), yp (t)}
be described by the transfer function
Zp (s)
Wp (s) = kp (9.8)
Rp (s)

Zp (s) is an mp degree monic polynomial and Rp (s) is an np degree monic polynomial,

with np ≥ mp . It is assumed that Zp (s) and Rp (s) are coprime.
Let a linear time varying (LTV) model reference transfer function with input–
output pair {r(t), ym (t)} be defined by
Zm (s, t)
Wm (s, t) = km (t) (9.9)
Rm (s, t)
with

Zm (s, t) = (s + am(nm +1) (t))(s + am(nm +2) (t)) . . . (s + am(nm +mm ) (t))
Rm (s, t) = (s + am1 (t))(s + am1 (t)) . . . (s + amnm (t)) (9.10)

Zm (s, t) is an mm degree monic polynomial and Rm (s, t) is an nm degree monic

polynomial, with nm ≥ mm . Let Wm (s, t0 ) be the initial reference model (i.e. t ≥ t0 )
 Adaptive systems 125

with Zm (s, t0 ) and Rm (s, t0 ) coprime. Wm (s, t) varies in a similar manner with respect
to Wm (s, t0 ), to be precisely defined in the following section.
Our objective is to design an adaptive control system with an adaptive con-
troller with bounded control effort by varying the model reference in a dimensionally
similar manner, in order to keep certain input–output properties constant. It is
required that the adaptive control guarantees the boundedness of the tracking error
and its convergence to zero, where r(t) is assumed to be bounded, piecewise contin-
uous and persistently exciting of appropriate order to guarantee the converge of the
plant-identified parameters to their real values.
Figure 9.1 shows the proposed adaptive scheme. It is an indirect MRAC with
time varying model reference Wm adapted by the control effort of the closed loop in
such a way that the control effort is bounded by cmax .
The assumptions regarding the plant and the reference model are summarized as
follows:

● Plant assumptions
(P1) Zp is a Hurwitz polynomial.
(P2) The relative degree n∗p = np − mp is known.
(P3) A minimum value of the high frequency gain kpmin is known.
(P4) An upper bound n of np is known.

ym
Wm
−
e

up yp
r Wp
C

Identi-
fication

θc θp θm
Control
Design
effort

cmax

Figure 9.1 Similar model reference adaptive control (SMRAC)

126 Application of dimensional analysis in systems modeling and control design

● Reference model assumptions

(M1) The relative degree n∗m = nm − mm is greater than or equal to n∗p , that is,
n∗m ≥ n∗p .
(M2) At any time instant, Wm (s, t) is dimensionally similar to Wm (s, t0 ) for
t ≥ t0 (i.e. the variations of Wm (s, t) are not arbitrary but follow a similar
pattern).
Regarding plant assumptions, P2 and P4 are required to establish the existence
of a controller that solves the closed loop matching problem. Assumptions P1 and
P3 guarantee that the generated adaptive law is stable. On the one hand P1 prevents
unstable pole-zero cancelations. On the other hand P3 guarantees a bounded control
action. In fact, perfect control requires the feedforward term km /kp , which can become
unbounded if kp is not limited below. Finally reference model assumption M1 is
necessary to achieve perfect tracking, while assumption M2 is a design property
enabling the SMRAC approach.

9.3 SMRAC for first order plants

In order to make the main ideas of the SMRAC approach clear, we first develop
in detail the SMRAC for first order linear time-invariant plant. In Section 9.4 we
generalize the results for arbitrary order plants.
Consider the first order LTI plant

ẏp (t) = ap yp (t) + kp u(t) (9.11)

where u(t) and yp (t) are the input and output of the plant, respectively. The plant
parameters ap and kp are constant or slowly varying, real, and unknown. The plant
is controllable (i.e. kp  = 0) and it is not necessarily stable (i.e. values of ap > 0 are
allowed).
The reference model is a similar time varying system described by

ẏm (t) = am (t)ym (t) + km (t)r(t) (9.12)

where r(t) and ym (t) are the reference input and output, respectively, of the refer-
ence model. The parameters am (t) and km (t) are known time varying scalar constants
with am (t) < 0 (i.e. the reference model is stable). The reference model parame-
ters variation is not arbitrary but similar, in such a way that, at any time instant the
relation between the model reference parameters is given by

am (t) = St (t)−1 a0m

km (t) = Sk (t)St (t)−1 k0m

with Sk > 0, St > 0 the gain and time scaling factors, respectively, and a0m and
k0m initial or nominal values of the reference model.
 Adaptive systems 127

The time scaling factor St is modified regarding the currently estimated control
effort c(t) (calculated using the parameter estimation âp (t))

c(t) = −(a0m − âp (t)) (9.13)

as follows
⎧
⎨1 if c(t) ≤ cmax
St (t) = a0m (9.14)
⎩ if c(t) > cmax
âp (t) − cmax

If c(t) ≤ cmax then the model reference Wm is kept constant. On the contrary,
if c(t) > cmax , the time scale factor is modified. As a result, it is guaranteed that
the control effort is bounded by cmax , independently of the value of âp (t), varying due
to either estimation or slow plant parameters changes.
It is assumed that if ap > 0 (i.e. the plant is unstable) then cmax > ap (t), because
for stabilization purposes it is required that am = c − ap < 0. If the system is stable
ap < 0 then it is just required that cmax ≥ 0. In any case St > 0.
The gain scaling factor Sk can be kept constant and equal to unity because its
value does not affect the control effort. It can also be modified if a variable model
reference steady state gain is desired.
Our objective is to make the control error ec = yp − ym tend to zero with time
with bounded control effort, using an indirect control scheme in which the model
reference, when required, is no longer constant but varies in a similar way. The
control input is generated as

u(t) = θ (t)yp (t) + k(t)r(t) (9.15)

where θ (t) and k(t) are the controller parameters to be adjusted with an appro-
priate adaptive law that accomplishes the above mentioned objectives. Following
classical indirect MRAC, the controller parameters that provide the desired model
reference behavior of the closed loop are given by

am − ap (t)
θ (t) =
kp
km (t)
k(t) = (9.16)
kp

Because the plant parameters are unknown, the controller parameters are calcu-
lated using the plant parameters estimates

am (t) − âp (t)

θ(t) =
k̂p (t)
km (t)
k(t) = (9.17)
k̂p (t)
128 Application of dimensional analysis in systems modeling and control design

The parameter adaptation law

â˙ p = γ1 yp (t)e(t) (9.18)

˙
k̂p = γ2 up (t)e(t) (9.19)

with γ1 and γ2 positive constants, together with

am (t) = St−1 a0m

km (t) = Sk St−1 k0m

and
⎧
⎨0 if c(t) ≤ cmax
Ṡt (t) = −a0m
⎩ â˙ p if c(t) > cmax
(âp (t) − cmax )2

yields a stable adaptive parameter estimation with bounded control effort.

Note that the controller parameter θ (t) is bounded by cmax /kpmin .
The stability of the SMRAC is established as follows. The identification model
is described by Reference 5

ŷ˙ p (t) = am (t)ŷp + (âp (t) − am (t))yp (t) + k̂p (t)u(t) (9.20)

with am (t) < 0 for stability of the reference model. The above model yields the fol-
lowing identification error (output error): ei = yp (t) − ŷp (t). The error variation with
time is

ėi = ŷ˙ p (t) − ẏp (t) = am (t)ei (t) + (− âp + ap )yp (t) − (− k̂p − kp )u(t) (9.21)

which can be written as

ėi = am (t)ei (t) − ãp yp (t) − k̃p u(t) (9.22)

where ãp = âp − ap and k̃p = k̂p − kp are the plant parameters error.
Consider the following Lyapunov function:
1 2
V = (e + γ1 ã2p + γ2 k̃p2 ) (9.23)
2
with time derivative given by
˙
V̇ = am (t)e2 + (ã˙ p − yp e)ãp + (k̃p − ue)k̃p (9.24)

If the plant parameters are adjusted using the adaptive laws ((9.18) and (9.19))
the stability of the adaptive scheme follows because

V̇ = am (t)e2 < 0 (9.25)

 Adaptive systems 129

To show this consider the parameter am (t):


a0m if c ≤ cmax
am (t) = (9.26)
St−1 a0m if c > cmax with St > 0

In both cases am (t) < 0 yielding system stability because the time scaling fac-
tor St as calculated in (9.14) is always positive.

9.4 SMRAC for arbitrary order plants

9.4.1 SMRAC control scheme

Consider the following np order LTI plant:

Zp (s)
Wp (s) = kp (9.27)
Rp (s)

and the nm order LTV reference model transfer function

Zm (s, t)
Wm (s, t) = km (t) (9.28)
Rm (s, t)

together with the assumptions established in Section 9.2.2.

The reference model changes with time following a similar pattern. As a result
the parameters of the time varying polynomials Zm (s, t) and Rm (s, t), defined as

Zm (s, t) = (s + am(nm +1) (t))(s + am(nm +2) (t)) . . . (s + am(nm +mm ) (t))
Rm (s, t) = (s + am1 (t))(s + am1 (t)) . . . (s + amnm (t)) (9.29)

are given by

ami (t) = St−1 a0mi i ∈ [1, . . . , nm + mm ]

km (t) = Sk St−1 k0m (9.30)

with Sk , St the gain and time scaling factors, respectively, and a0mi and k0m initial
values of the reference model (i.e. nominal reference model). We keep, without
loss of generality, the gain scaling factor constant and equal to 1, that is, Sk = 1, in
order to keep a constant reference model high frequency gain.
The estimated parameters of the plant p̂ = [ĉ0 , c̄ˆ , d̂0 , d̄ˆ T ]T , as parameterized
T

by the observer shown in Figure 9.2, are continuously updated by the following
adaptive law:
˙
k̂ = −sgn(kp )1 ζ̄ū (9.31)

p̄˙ˆ = −1 ζ̄o (9.32)

130 Application of dimensional analysis in systems modeling and control design

u yp
Wp
1
–
–
F, g W ĉT W
– ŷp

W ĉ0 d̂ T W F, g
–

– –
1ζoTζo d̂0 W

Figure 9.2 Observer parameterization on the SMRAC

where k̂ = ĉ0−1 , p̄ˆ = [c̄ˆ , d̂0 , d̄ˆ T ]T , 1 = yp − ŷp , and ζo , ζū measurable signals,
T

which for the sake of clarity are defined in Section 9.5.

The current control effort estimation is given by
n
c(t) = Anm0m − Âpp (t) (9.33)
m n np
with Anm0m = − ni=1 a0i (i.e. a0i poles of W0m (s)) and Âpp = − i=1 âpi (i.e. âpi poles
np
of Ŵp (s)). The poles Âp are calculated by means of the estimated parameters p̂ and the
representation shown in Figure 9.2.
The time scaling factor St is modified regarding the currently estimated control
effort c(t) as follows:
⎧
⎨1 if c(t) ≤ c
⎪ max

St (t) = nm
Am0
⎪
⎩ max n if c(t) > cmax
c + Âpp (t)

If c(t) ≤ cmax then the reference model Wm is kept constant and equal to the
nominal model Wm (s, t0 ). On the contrary, if c(t) > cmax , the time scale factor is
modified. As a result, it is guaranteed that the control effort is bounded by cmax ,
n
independently of the value variations of Âpp , due to either estimation or slow plant
parameters changes.

Assumption 9.1. In case of Wp stable (i.e. all api < 0 i ∈ [1, . . . , np ]), then cmax ≥ 0.
This assures that the closed loop poles Anmm are, at least, equal to the open loop ones.

Assumption 9.2. In case of Wp unstable (i.e.

n n n there is at least one api > 0 i ∈
[1, . . . , np ]), then cmax > −Apupu with Apupu = − 1pu apu and apu > 0, that is the unsta-
ble poles of Wp . This assumption is necessary in order to be able to shift the open
loop unstable poles to closed loop stable poles. Analytically it can be shown by rewrit-
n n
ing (9.33) as c = Anmm − Apsps − Apupu separating the stable and the unstable poles. It
 Adaptive systems 131
n n
then follows that, disregarding the plant stable poles, Anmm = c + Apupu . If cmax > Apupu
then Anmm > 0, thus Wp is stable.

The consequence of both assumptions is that the scaling factor St is always

positive. Finally the control action that maintains a bounded control effort is cal-
culated as

u = k̂ r + p̄ˆ T (9.34)

9.4.2 SMRAC stability analysis

Consider the observer

ẇ1 = Fw1 + gup , w1 (0) = 0

ẇ2 = Fw2 + gyp , w2 (0) = 0 (9.35)

] and p̄ˆ = [c̄ˆ , d̂0 , d̄ˆ T ]T .

T
with w̄o = [w1o
T T T
, yp , w2o
We modify a standard observer yielding the augmented observer presented in
Figure 9.2, with output

ŷp = W (ĉ0 Wu − p̄ˆ T WIw0 + 1 ζ̄oT ζ̄o ) (9.36)

and Wm = W W 1 with W SPR and ζ̄o = W w̄o the filtered observer state. The con-
trol action is given by u = k̂(r + p̄ˆ T ), resulting

ŷp = W (ĉ0 W k̂(r + p̄ˆ T ) − p̄ˆ T WIw0 + 1 ζ̄oT ζ̄o ) (9.37)

which can be further simplified by linearity, yielding

ŷp = ym + W 1 ζ̄oT ζ̄o (9.38)

The augmented observer error 1 = yp − ŷp is then written as

1 = yp − ŷp = yp − ym − W 1 ζ̄oT ζ̄o (9.39)

Now consider the tracking error e1 = yp − ym (see Figure 9.3). The control
action u = k̂(r + p̄ˆ T ) can be rewritten as u = (k + ψ1 )(r + p̄T + 
¯ T w̄o ) where ψ1 =
¯ ˆ
k̂ − k and  = p̄ − p̄ = [ĉ − c , d̂o − do , d̂ − d ] are the parameter errors. The
T T T T

tracking error is

kp
e1 = ¯ T w̄o + ψ1 (r + p̄w̄o + 
Wm (k  ¯ w̄o )) (9.40)
km

1
The dependence of Wm , W and W on time is dropped for readability purposes. It is written explicitly
when needed.
132 Application of dimensional analysis in systems modeling and control design

ym
Wm

–
r e1

u–
k Wp
yp

– T–
pˆ w0

Figure 9.3 Reference model and plant error

¯ w̄o and for perfect tracking k = km /kp , resulting

from Figure 9.3, ū = r + p̄w̄o + 

kp
e1 = ¯ T w̄o + ψ1 ū)
Wm (k 
km
 
kp
= W W ψ1 ū + W  ¯ T w̄o
km
 
kp
= W ¯
ψ1 ζ̄ū +  ζ̄o
T
(9.41)
km

Finally, recalling that ζo = W w̄o is the filtered observer state and ζū = W ū is the
filtered input ū, 1 can be written as
 
kp
1 = W ¯ T ζ̄o − 1 ζ̄oT ζ̄o
ψ1 ζ̄ū +  (9.42)
km

The following adaptation laws are proposed due to the SPR of W :

˙
ψ̇1 = k̂ = −sgn(kp )1 ζ̄ū (9.43)
˙¯ = p̄˙ˆ = − ζ̄
 (9.44)
1 o

Now it is shown that the adaptive system is stable. Recall that (9.42) can be
rewritten in state space form as
 
kp
˙ = A1  + b1 ¯ T ζ̄o − 1 ζ̄oT ζ̄o
ψ1 ζ̄ū +  (9.45)
km
1 = hT1  (9.46)

with hT1 (sI − A1 )−1 b1 = W SPR.

 Adaptive systems 133

We assume a Lyapunov candidate equation of the form

|kp | 2
V =  T P + ¯ T
ψ + ¯ (9.47)
km 1
which has a time derivative equal to
 
kp
V̇ =  T
(AT1 P + PA1 ) + 2 Pb1
T ¯ T ζ̄o − 1 ζ̄oT ζ̄o
ψ1 ζ̄ū + 
km
kp
−2 ¯ T ζ̄o
ζ̄ū − 21  (9.48)
km ψ 1  1

As before, since W is SPR ∀t due to the similarity transformation, it then fol-

lows by the Kalman–Yakubovich lemma that AT1 (t)P(t) + P(t)A1 (t) = −Q(t) with
Q(t) positive definite ∀t and that P(t)b1 (t) = h1 (t) also ∀t. Hence,

V̇ = − T Q(t) − 212 ζ̄oT ζ̄o ≤ 0 (9.49)

so that , ψ1 , and  are bounded and , ψ̄˙ 1 , and  ˙¯ are also square inte-
grable functions. It then follows by signal growth rates arguments (see for instance
Reference 5) that all signals are uniformly bounded, then limt→∞ 1 (t) = 0 which
implies that limt→∞ e1 (t) = 0. However, the convergence of the parameter estimates
to the real parameter values cannot be guaranteed because it depends on the persis-
tently exciting nature of signal r.

9.4.3 SMRAC operation modes

So far we have derived the SMRAC with bounded control effort, meanwhile main-
taining a nominal reference model W0m as constant as possible. In fact, the variations
of the nominal reference model are only performed when the control effort is about
to exceed.
On the other hand we can easily devise a scheme with constant control effort
during the whole operation range, equal to the maximum allowable control effort
cmax (t). In this case we are constantly modifying the reference model Wm (t) and the
time scaling factor St (t) is modified as follows:
nm
Am0
St (t) = n (9.50)
cmax + Âpp
with no control effort estimation required. Note that in this case an indirect approach
is no longer mandatory.
Note that by definition, St (t) is always positive and now, it is possible that St (t) > 1
because we are demanding a reference model ‘faster’ than W0m . In any case, the
stability is guaranteed and their selection follows the user desires. In the first case the
reference model is kept as constant as possible and only modified when the estimated
control effort is violated. In the second case the reference model is continuously
modified in order to have a constant control effort operation.
134 Application of dimensional analysis in systems modeling and control design

9.5 Application example

We consider the system describing angular rigid body dynamics relating the angular
velocity θ̇ to the applied torque T . The transfer function is

θ̇ (s) J −1
G(s) = = (9.51)
T (s) s + C/J

where J is the inertia and C the viscous damping. We assume that the damping is
constant and equal to unity (i.e. C = 1) but the inertia is time varying J ∈ [0.5, 2] as
follows:

⎧
⎪
⎪ for t ∈ [0–150]
0.5
⎨
1.5
J (t) = 0.5 + (t − 150) for t ∈ [150–250]
⎪
⎪ 100
⎩2 for t ∈ [250–350]

(a) 0.5
Output error

−0.5
0 50 100 150 200 250 300 350

(b) 2
1
Output

0
−1
−2
0 50 100 150 200 250 300 350

(c) 2
Control action

−2

−4
0 50 100 150 200 250 300 350
Time (s)

Figure 9.4 Signals

 Adaptive systems 135

(a) 2.5
2
1.5
St

1
0.5
0
0 50 100 150 200 250 300 350

(b) 2
a
Plant parameters

1 b
0
−1
−2
0 50 100 150 200 250 300 350

(c) 4
Theta
Controller gains

2 k

−2
0 50 100 150 200 250 300 350
Time (s)

Figure 9.5 Parameters

The nominal values considered for the model reference model are a0 = −2 and
k0 = 2, so

2
Wm (s, t0 ) = (9.52)
s+2

and the reference model is time varying as a function of St as follows:

Sk (2/St )
Wm (s, t) = (9.53)
s + 2/St

In the example we keep Sk = 1, ∀t in order to keep the steady-state gain

unchanged. The maximum control effort bound selected is cmax = 0.2. Note in this
case that, due to plant stability, no lower bound on cmax is required for stabilization
purposes, so the only requirement is cmax ≥ 0. The reference signal r is a square wave
of zero mean value, amplitude equal to 1 and period 40 s.
Figure 9.4(a) presents the output error evolution. It tends to zero with time
as expected. During the time interval t ∈ [0–150] the output error is due to initial
error parameters estimation, while during the time interval t ∈ [150–250] the error is
136 Application of dimensional analysis in systems modeling and control design

mainly due to slow parameter variation tracking. The identified parameters are shown
in Figure 9.5(b) together with the real parameters evolution.
Next Figure 9.5(a) presents the model reference time scaling factor St evolu-
tion. The scaling factor St is equal to unity for t ∈ [60–180]. During that period
the reference model is equal to the nominal model Wm (s, t0 ) previously described.
However, during other operation periods the scaling factor reaches values greater
than unity, so the reference model becomes slower, first due to estimation error and
second due to inertia variation. This can be verified by Figure 9.4(b). It can also be
seen in Figure 9.4(c) that the control action increases because inertia is very mild.
Figure 9.5(c) shows also the time evolution of controller parameters θ (t) and k(t).
The θ(t) value is the feedback and the k(t) is the feedforward.
Now we compare the evolution of two SMRAC schemes with distinct control
effort bounds, cmax = 0.2 and cmax = 2. Figure 9.6(a) shows the model reference
scaling factor St for both control effort bounds. It can be seen that although for
cmax = 0.2 the St takes values greater than unity as seen before, for cmax = 2 the St
value is equal to unity for all t. It thus follows that the reference model is kept constant
and the results are equal to a standard MRAC.

(a) 2.5
2
1.5
St

1
0.5
0
0 50 100 150 200 250 300 350

(b) 2
1
Outupt

0
−1
−2
0 50 100 150 200 250 300 350

(c) 4
Control action

2
0
−2
−4
0 50 100 150 200 250 300 350
Time (s)

Figure 9.6 Comparison of responses with distinct control effort. Dashed line
cmax = 0.2 and solid line cmax = 2
 Adaptive systems 137

2.5
2
1.5
St

1
0.5
0
340 340.5 341 341.5 342 342.5 343 343.5 344 344.5 345

2
1
Outupt

0
−1

−2
340 340.5 341 341.5 342 342.5 343 343.5 344 344.5 345

5
Control action

−5
340 340.5 341 341.5 342 342.5 343 343.5 344 344.5 345
Time (s)

Figure 9.7 Comparison details

The comparison of outputs and control actions for both control effort bounds
is presented in Figure 9.6(b) and (c). As expected, the SMRAC scheme with
cmax = 0.2 shows a time response slower than the scheme with cmax = 2 when the
inertia increases. However, the slight increase in the time response comes at the
cost of higher control actions as seen in Figure 9.6(c). In fact the control action for
cmax = 2 shows values around 4 units (spikes) while the control action for cmax = 0.2
is around 1 unit. Details of the presented plots are presented in Figure 9.7. The
comparison shows how the control effort bound is useful to keep the control action
bounded.

References
[1] P.Albertos and M. Olivares. ‘Time delay limitations in control implementations.’
In European Control Conference, Karlsrhue, 1999
[2] K.J. Astrom and B. Wittenmark. Adaptive Control. Addison-Wesley Publishing
Company, 1989
138 Application of dimensional analysis in systems modeling and control design

[3] H. Hailu. Dimensional tranformation: a novel method for gain-scheduling and

robust control. PhD thesis, The Pennsylvania State University, 2006
[4] I.D. Landau. ‘From robust control to adaptive control.’ Control Engineering
Practice, 7:1113–1124, 1999
[5] K.S. Narendra and A.M. Annaswamy. Stable Adaptive Systems. Dover Publica-
tions, Inc., 1989
 Index

actuator limitations 122–3 loop transfer function GC 102

adaptive control systems 121–37 overview 97
see also similar model reference selecting tuning rules 106–7
adaptive control (SMRAC) sensitivity function bandwidth 103,
arbitrary order plants, SMRAC for 104
129–33 sensitivity peak function 104, 105
control scheme 129–31 continuous time control 111
operation modes 133 control effort 122–3
stability analysis 131–3 control engineering 8–9
arithmetic of dimensions 12 see also comparative framework,
integral 12 PID tuning rules
power 12 control of dimensionally similar
product 12 systems 109–11
quotient 12 complete similarity 111–13
experimental case study 115–19
bandwidth: see sensitivity function overview 109–10
bandwidth partial similarity 113–15
Buckingham pi theorem 3, 14–16, 18 control theory 8
see also dimensionless representation see also homogeneity, of PID tuning
rules
closed loop transfer functions 83–6
comparative framework, PID tuning decentralized controller: see reduced
rules 102 effective transfer function
dimensionless 85–6 dimensional analysis 1–2
loop transfer function GC 83–5 application to control problems 6–9
comparative framework, PID tuning scientific application 3–6
rules dimensional matrix 16–17
application example 102–3 dimensional set 17–18
characterization 103 dimensional similarity 2–3, 5–6
closed loop transfer functions 102 control: see control of dimensionally
control action variability 106 similar systems
elements 98–9 defining 19–20
framework 99–100 discrete time 46–8
indexes 104 model law 21
integral absolute error 104, 105 prototype and model systems 19–20
integral errors 100–2 scaling factors 20
140 Application of dimensional analysis in systems modeling and control design

state space 44–6 dimensionless controller

transfer function 41–4 representation 83
dimensionless representation 3, 5 nonhomogeneous rules 89–94
discrete time models 35–7 optimality 87–9
overview 23
properties 23 independent dimensions 14
reduced effective transfer function indexes, comparative framework,
(RETF) 67–8 PID tuning rules 104
state space 27–32 integral 12
transfer function 23–7 integral absolute error, comparative
transfer function vs. state space framework 104, 105
32–5 integral errors, comparative framework
dimensions 100–2
arithmetic: see arithmetic of international system (SI): see SI system
dimensions inverse response analysis, RETF 68–72
classification 11 pole-zero cancelation 71
defined 11 time delay factorization 70, 71
dependent and independent 14 inverse response analysis, second order
derived 12 52–64
fundamental 11 see also second order inverse
power law monomial 14 response model
system of units 13–14
discrete time control 111–13 mathematical equation 1
discrete time dimensional similarity homogeneity and 2
46–8 matrix
sampled-data transfer function 46–7 dimensional 16–17
state space 47–8 dimensional set 17–18
transfer function 46 model reference adaptive control
discrete time models dimensionless (MRAC) 121
representation 35–7 monomial power law 14
state space 36–7 MRAC: see model reference adaptive
transfer function 35–6 control (MRAC)
dynamical systems: see dimensional
similarity; dimensionless nonhomogeneous and homogeneous
representation tuning rules 89–94
see also homogeneity, of PID tuning
first order linear time-invariant plant, rules
SMRAC for 126–9
Froude number 3, 23 optimality, of homogeneous tuning
rules 87–9
homogeneity, of PID tuning rules 8–9 weighting factors 88–9
characterization 80–2
closed loop transfer functions 83–6 partial similarity 113–15
dimensionless controller parameters physical equation 1
80 homogeneity and 2
 Index 141

physical quantity 11 similar model reference adaptive

physical relations 3–4 control (SMRAC) 122, 123–6
PID tuning rules adaptive scheme 124–6
dimensionless comparison 97–107 for arbitrary order plants 129–33
see also comparative framework, dimensionally similar models 123–4
PID tuning rules for first order linear time-invariant
homogeneity of 79–94 plant 126–9
see also homogeneity, of PID single input/single output (SISO) 64
tuning rules SISO: see single input/single output
pole-zero cancelation (SISO)
inverse response analysis, SI system 13
RETF 71 derived units 13
reduced order model 73–4 fundamental dimensions 13
polymerization reactor 76–7 SMRAC: see similar model reference
power 12 adaptive control (SMRAC)
power law monomial: see monomial state space dimensional similarity
power law 44–6
Prandtl number 3 discrete time 47–8
product 12 state space dimensionless representation
proportional integral derivative (PID) 27–32
tuning rules: see PID tuning discrete time models 36–7
rules interpretation 30–2
inverted pendulum 29–30
quotient 12 linear time invariant (LTI) 27
system transformation 28
reduced effective transfer function transfer function vs. 32–5
(RETF) transformations 30–2
application examples 75–7 system identification and model
dimensionless representation 67–8 reduction 8
inverse response analysis 67–72 application examples 61–4
problem statement 64–7 difficulties 51
reduced order model 72–4 general procedure 51–2
pole-zero cancelation 73–4 overview 51
time delay factorization 73–4 reduced effective transfer function
Reynolds number 3 reduction 64–77
second order inverse response model
second order inverse response model identification 52–64
52–64 system of units 13–14
application examples 61–4
dimensionless representation 53–4 time delay factorization
identification procedure 54–61 inverse response analysis, RETF 70,
problem statement 52–3 71
sensitivity function bandwidth reduced order model 73–4
103, 104 TITO: see two inputs/two outputs
sensitivity peak function 104, 105 (TITO)
142 Application of dimensional analysis in systems modeling and control design

transfer function dimensional similarity parameters 24

41–4 vs. state space 32–5
discrete time 46–7 two inputs/two outputs (TITO) 66
sampled-data transfer function 46–7
transfer function dimensionless units 11
representation 23–7 see also SI system
dimensionless form 26–7
dimensionless numbers 25–6 Vinante and Luyben distillation column
discrete time models 35–6 75–6
 Control Engineering Series 90

Application of Dimensional Analysis in

Systems Modeling and Control Design

Application of Dimensional Analysis in
Systems Modeling and Control Design Application of
Dimensional
Dimensional analysis is an engineering tool that is widely applied to numerous engineering
problems, but has only recently been applied to control theory and problems such as
Pedro Balaguer is a lecturer
at the Department of Industrial
Analysis in Systems
Modeling and
identification and model reduction, robust control, adaptive control, and PID control. Systems Engineering and
Design, Universitat Jaume I de
Application of Dimensional Analysis in Systems Modeling and Control Design provides an Castelló, Spain. His research
introduction to the fundamentals of dimensional analysis for control engineers, and shows interests include the application
how they can exploit the benefits of the technique to theoretical and practical control of dimensional analysis to
problems. Topics covered include dimensional analysis and dimensional similarity, dynamical control problems, PID control,

Control Design
systems dimensionless representation, dimensionless systems identification and model order supervision of iterative and
reduction, homogeneity of PID tuning rules, dimensionless PID tuning rules comparison, adaptive control systems,
and energy cost and energy
dimensional analysis control fundamentals, control of dimensionally similar systems, and
consumption optimization.
adaptive control in the presence of input saturation.
Balaguer

Pedro Balaguer
The Institution of Engineering and Technology
www.theiet.org
978-1-84919-621-5

Application of Dimensional Analysis.indd 1 17/06/2013 14:04:01