137
University of Calgary PRISM: University of Calgary's Digital Repository Graduate Studies Legacy Theses 1997 Optimal design of multirate systems Shu, Huang Shu, H. (1997). Optimal design of multirate systems (Unpublished doctoral thesis). University of Calgary, Calgary, AB. doi:10.11575/PRISM/22106 http://hdl.handle.net/1880/26848 doctoral thesis University of Calgary graduate students retain copyright ownership and moral rights for their thesis. You may use this material in any way that is permitted by the Copyright Act or through licensing that has been assigned to the document. For uses that are not allowable under copyright legislation or licensing, you are required to seek permission. Downloaded from PRISM: https://prism.ucalgary.ca

Optimal design of multirate systems

  • Upload
    others

  • View
    9

  • Download
    0

Embed Size (px)

Citation preview

Page 1: Optimal design of multirate systems

University of Calgary

PRISM: University of Calgary's Digital Repository

Graduate Studies Legacy Theses

1997

Optimal design of multirate systems

Shu, Huang

Shu, H. (1997). Optimal design of multirate systems (Unpublished doctoral thesis). University of

Calgary, Calgary, AB. doi:10.11575/PRISM/22106

http://hdl.handle.net/1880/26848

doctoral thesis

University of Calgary graduate students retain copyright ownership and moral rights for their

thesis. You may use this material in any way that is permitted by the Copyright Act or through

licensing that has been assigned to the document. For uses that are not allowable under

copyright legislation or licensing, you are required to seek permission.

Downloaded from PRISM: https://prism.ucalgary.ca

Page 2: Optimal design of multirate systems

THE U ~ R S I T Y OF CALGARY

Optimal Design of Multirate Systems

by

Huang Shu

A THESIS

SUBMITTED TO THE FACULTY OF GRADUATE STUDLES

M PARTIAL FULFILL-MEXT OF THE REQb?REMENTS FOR THE

DEGREE OF DOCTOR OF PEILOSOPHY

CALGARY, ALBERTA

APRIL, 1997

@ Huang Shu 1997

Page 3: Optimal design of multirate systems

Acquisitions and Acquisitions et BiMi~graptrÈ Sewke~ - services bibliographiques

The author has granted a non- exclusive licence dowing the National Library of Canada to reproduce9 loan, distrrbute or seil copies ofhismer thesis by any meam and in any fom or fomat, making ttUs thesis available to interested persons.

L'auteur a accordé me licence non exclusive permettant à la Bibliothèque nationaie du Canada de ~ ~ , ~ 9 distIi'buerou vendre des copies de sa thèse de cpelqye manière a sous qgeIwe forme que ce soit pour mettre des exemplaires de cette thèse à la disposition des personnes intéressées.

The author reaias ownership of the L'auteur conserve la propriété du copyright in hismer thesis. Neither droit d'auteur qui protège sa thèse. Ni the thesis nor substantid extracts la ttiéie ni des extraits substantieis de fiom it may be printed or otherwise celle-ci ne doivent être imprimés ou reproduced with the author's autrement reproduits sans son permission. autorisation.

Canada

Page 4: Optimal design of multirate systems

ABSTRACT

Multirate -stems are o h used to achieve a cost advantage in implementation.

This thesis is devoted to both theory and application of optimal design of multirate

systerns-

In the theory part, we develop new algorithms for mdtirate optimal design mith

X2 and 31, criteria. In particular, we present state-space solu~ions to multirate X2-

optimal and suboptimal control and a solution to multirate ?dm-suboptimal control in

tems of a bilinear tansformation. Explicit formulas are given in terms of solutions of

Riccati equations and nest matrices are used to handle the causality constraint arising

h m multirate system structures. Compared to existing results in the frequency

domoin, the coctrollers obtained have the advantage of comput ational efficiency and

ease of implementation. Based on the new dgorithms, a software package is developed

in MATLAB for a general multirate 3d. design-

In the application part, we study via 31, optimization two different systems in

control and signal processing: a power system and a hybrid filter bank. For the

power system, we present a thorough study of power system stabilizers with four

types of 31, stabilizers designed and investigated, which include analog design based

asd discrete design based, single input and multi-input, and single-rate and multi-

rate. The multirate stabilizers are designed based on the new algorithm developed

and are superior to the single-rate ones in the sense that they require substantially

lower sampling rates. For d designs, we propose a systematic method of choosing

weight ing functions to meet certain operating requirements. The st abilizers designed

provide an implementation advantage - they have low complexity and require only

slow samphg rates - and outperform the conventional stabilizers on a series nonlinear

dqmamic tests. For the multirate filter bank. we present a direct design method for

the hybrid structure which consists of continuous- and discrete-time systems. Specifi-

c a synt hesis filters are designed to min-e the worst-case energy gain of the error 111

Page 5: Optimal design of multirate systems

system, suitably weighted, between the hybrÎd filter b d and an ideal system. This

mdtirate hybrid design problem is converted into a single-rate discrete-the one of

optimization, which is then solved by the standard 36, design technique. An

example is discussed in detail to illustrate the design process. The filter bank design

also represents a new application of ?&, contrd theory

Page 6: Optimal design of multirate systems

1 a m deeply indebted to my supervisor Dr. T. Chen for his tremendous support, en-

couragement, and motivation throughout the course of this research. I wodd dso lilce

to t h a d him for the extremely carefiil reading and correcting of Mnous manuscripts

related to this work and for ail the valuable suggestions.

1 wodd like to thanli Dr. O.P. Malik for his suggestions. .&O a special note of

thanks goes to Jian He and Shen Chen for several interesting discussions during the

course of this work. The test resdts for the power system in this thesis are based on

the power system simulation program developed by Jian He-

I am gratefid to my research coueagues, N. Rafee and A. Saadat-Mehr, for proof-

reading the manuscript of this t hesis.

1 would like to adcnowledge the financial support for this research, provided in part,

by the Natural Sciences and Research Council of Canada, made a d a b l e by Dr. T.

Chen. and by the Depaztment of Electrical and Cornputer Engineering through the

Graduate Research Scholarship.

Finally, I owe special thanks to m y d e Lin and my son Jimmyfor being a constant

source of encouragement, support and optimism throughout rny doctoral work.

Page 7: Optimal design of multirate systems

To My Parents.

Page 8: Optimal design of multirate systems

CONTENTS

- 0 APPROVAL PAGE. .... .... .... .. ..... .... . . - . - .-... .....,. ii

A B C . . . . - . . . . . . . . . - - . . . . . . . . . . . . . . . . . . - . . . . . . iü

DEDICATION.. . .. . . . . .. . - . ,.. - . . . . . . .. - . . . . . .. . . . . . . . .. . . M

TABLE OF CONTENTS.. . . . . . . . . . . . . . ., . . . . . . . . . . . .. . . . . - . . vii

LIST OF TABLES. . . . .. . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ix LIST OF FIGURES. . . . . . . . . . . , . - - . . . . . . . . . . . . - .. . . - . . . . . . . . x

LIST OF SYMBOLS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . - . .. . - - . . xiii

CHAPTERS

1. INTRODUCTION . . . . . . . . . . . . . . . . . . . - . . . . . - . . . . . . . . . . . . . 1 1. t Single-Rate Sampled-Data Systems . . . . . . . . . . . . . . . . . . . 1 1.2 Yultirate Sampled-Data Systems . . . . . . . . . - . . . . . . . . . . 3 1.3 Properties of Multirate Controllers . . . . . . . . . . . . . . . . . . . 6 1.4 O u t h e of the Thesis . . - . . . . . . . . . . . . . - . . . . . . . . . . 9 1.5 Xotation . . . . . . . . . . . . . . . . . . . . . . . - . - . . . . . . . . 11

2. CONSTRAINED 3C2 CONTROL: A STATESPACE APPROACH 13 2.1 Causality and Nest Operator . . . . . . . . . . . . - . . . . . . . . . . I I 2.2 The Unconstrained Case . . . . . . . . . - . . . . . - . . . . . . . . - 16 2.3 Main Results: The Constrained Case . . . . . . - . . . . . . . . . . . 19 2.4 Proof of the Main Results . . . . - . . . . . . . . . . . - . . . . . - . 13 2.5 Conclusions . . . . . . . . . . , - - - . . . . . . . - . - . . . . . . . - 16

3. CONSTRAINED 31, CONTROL: AN ALTERNATIVE APPROACH 28

3.1 An EExting Solution . . . . . . . . . . . . . . . . . . . . . . . . . . . 49 3.2 A Solution Process via Bilinear Transformation . . . . . . . . . . . . 31 3.3 P r o o f o f I n v e r t i b i l i ~ ~ f D ~ ~ a n d D ~ ~ . . . . . . . . . . . . . . . . . . 33 3.4 A Design Procedure for klultirate 3d. Controllers . . . . . . . . . . . 36 3.5 Conc~usions . . . . . . . . - . . . . . . . . . . . . . . . . . . . . . . . 41

Page 9: Optimal design of multirate systems

4 . 31, DESIGN OF DIGITAL POIKER SYSTEM STABILIZERS ... 42 . . . . . . . . . . . . . . . . . . . . . . . 4.1 Motivation and Introduction 4'2

. . . . . . . . . . . . . . . 4.2 Linearized Plant and Weighting Functions 46 . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.21 Linearization 46

. . . . . . . . . . . . . . . . . . . . . . . 1.22 Weighting Functions 48 . . . . . . . . . . . . . . . . . . . 4.3 Design of Robust Digital Stabilizers 51

. . . . . . . . . . . . . . . . 4.3.1 Analog Design and Discretization 54 . . . . . . . . . . . . 1.3.2 DisueteTime Design of SIS0 Stabilizers 60 . . . . . . . . . . . 4.3.3 DisueteTime Design of MIS0 Stabilizers 66

. . . . . . . . . . . . . . 4.3.4 Multirate Design of FNSO Stabilizers 67 -. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.4 Conclusions ta

..... 5 . 'fl. DESIGN OF HYBRLD MULTlRATE FILTER BANKS 77 . . . . . . . . . . . . . . . . . . . . . . . 5.1 MotivationandIntroduction 77

. . . . . . . . . . . . . 5.2 Conversion to a Pioblem of 'flm Optimization 81

. . . . . . . . . . . . . 5.2.1 Conversion to a Discrete-The Problem Si . . . . . . . . . . . . . . . . . . 5-22 Conversion to an ?&,, Problem 83

. . . . . . . . . . . . . . . . . . . . . . . 5.2.3 A Design Procedure 90 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.3 Example 91

. . . . . . . . . . . . . . . . . . . . . . . . . 5.4 Performance Limitation 96 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.5 Conclusions 99

6 . ON CAUSALITY AND ANTICAUSALITY OF CASCADED LIN- ........................... EAR DISCRETETIME SYSTEMS -101

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.1 Motivation 101 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.2 Special Cases 103 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.3 General Case 106

............................... 7 . CONCLUDING REMARKS 110

............ A . P O m R SYSTEM MODEL AND PARAMETERS 121

Page 10: Optimal design of multirate systems

LIST OF TABLES

. . . . . . . . . . . . . 4.1 Parameters for the lineaxized power system mode1 45

. . . . . . . . . . . . . . . . . . . . 4.2 Parameters for the tmed analog CPSS 52

. . . . . . . . . . . 4.3 Nonlineu simulation tests of the CPSS and the figures 53

. . . . . . . . . . . . . . . . . . . A.1 Parameters for the power system mode1 122

Page 11: Optimal design of multirate systems

LIST OF FIGURES

1 2 LineanZed power system mode1 . . . . . . . . . . . . . . . . . . . . . . . . 5

. . . . . . . . . . . . . . . . . . . . . . . 1.3 The hybrid multirate filter bank 6

. . . . . . . . . . . . . . . . . . . . . . . 1.4 Singlerate discrete-time system 8

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.1 The lifted system 13

. . . . . . . . . . . . . . . . . . . 3.1 Standard multirate sampled-data setup 37

. . . . . . . . . . . . . . . . . . . . . . . . 3.2 Mdtirate discrete-the systern 33

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.3 The lifted LTI setup 38

. . . . . . . . . . . . . . 3.4 -4 design procedure for multirate & controuer 40

. . . . . . . . . . . . . . . . . . 4.1 A schematic diagram of the power system 43

. . . . . . . . . . . . . . . . . . . . . . . . 4.2 Linearized power system model 17

. . . . . . . 4.3 Recodgured power system with PSS and weighting functions 50

. . . . . . . . . . . . . . . . . . . 2.4 The IEEE standard CPSS configuration 52

4.5 Torque disturbance with normal load: CPSS (dot). analog stabilizer (solid) . . . . . . . . . . . . . . . reduced-order analog stabilizer (dash-dot) 54

4.6 Torque disturbance with light load: CPSS (dot). analog stabilizer (solid). . . . . . . . . . . . . . . . reduced-order analog stabilizer (dash-dot) 55

4.7 Torque disturbance with lead load: CPSS (dot). analog stabilizer (solid). . . . . . . . . . . . . . . . reduced-order andog stabilizer (dash-dot ) 55

4.8 Voltage disturbance with normal load: CPSS (dot). analog stabilizer (solid). . . . . . . . . . . . . . . . reduced-order analog stabilizer (dash-dot ) 56

4.9 Voltage disturbance with light load: CPSS (dot). analog stebilizer (solid). . . . . . . . . . . . . . . . reduced-order andog stabilizer (dash-dot ) 56

4.10 Voltage dist urbance with lead load: CPSS (dot). analog stabilizer (solid). . . . . . . . . . . . . . . . reduced-order analog st abilizer (dash-dot ) 57

Page 12: Optimal design of multirate systems

4.11 Ground fadt test with normal load: CPSS (dot), analog stabilizer (solid), reduced-order analog stabilizer (dash-dot), discretized analog stabilizer

. . . . . . . . . . . . . . . . . . . with sampling period 20 ms (dash)

. . . . . . . . . . . . . . . . . . . . . 4.12 The singlerate digital control setup

. . . . . . . . . . . . . . . . . . . . . . . . . . . 4.13 The discrete-the setup

4.14 Torque disturbance with normal load: CPSS (dot), digital SISO stabilizer . . . . . . . . . (solid), digital MISO singlerate stabilizer (dash-dot )

4.15 Torque disturbance with Eght load: CPSS (dot), digital SISO stabilizer . . . . . . . . . (solid), digital MISO single-rate stabiher (dash-dot)

4.16 Torque disturbance with lead load: CPSS (dot), digital SISO stabilizer . . . . . . . . . (solid): digital MISO single-rate stabilizer (dash-dot)

4.17 Voltage disturbance tvith n o m d load: CPSS (dot), digital SISO stabilizer . . . . . . . . . (solid), digital MISO singlerate stabilizer (dash-dot )

4.18 Voltage disturbance with light load: CPSS (dot): digital SISO stabilizer . . . . . . . . . (solid), digital MIS0 singlerate stabilizer (dash-dot)

4.19 Voltage disturbance with lead load: CPSS (dot), digital SISO stabilizer . . . . . . . . . (solid), digital MISO single-rate stabilizer (dash-dot)

1.20 Ground fadt test with normal load: CPSS (dot), digital SISO stabilizer . . . . . . . . . (solid) , digital MIS0 singlerate stabilizer (dash-dot )

4.21 Magnitude fiequency responses from Kef to p (solid) and w (dot); p and . . . . . . . . . . . . . . . . . . . . . . . . . . . . . w a x e b o t h i n p . ~ .

. . . . . . . . . . . . . . . . . . . . . 4.22 The multirate digital control setup

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.23 The lifted LTI setup

4.24 Torque disturbance with normal load: CPSS (dot), digital MISO multirate stabilizer with sampling period 120 ms for w (solid), and digital MISO multirate stabilizer with sarnphg period 180 ms for w (dash-dot) . .

4.25 Torque disturbance with light load: CPSS (dot), digital MISO multirate stabilizer with sampling period 120 ms for w (solid), and digital MISO multirate stabilizer 6 t h sampling period 180 ms for w (dash-dot) . .

4.26 Torque disturbance with lead load: CPSS (dot), digital MISO multirate stabilizer with sampling period 120 ms for w (solid), and digital MISO multirate stabilizer with sampling period 180 rns for w (dash-dot) . .

xi

Page 13: Optimal design of multirate systems

4.27 Voltage disturbance with normal load: CPSS (dot), digital MISO multi- rate stabilizer with sarnphg period 120 ms for w (solid), and digital MISO multirate stabiüzer with samphg perÏod 180 ms for w (dash-dot) 73

4.28 Voltage disturbance with Iight load: CPSS (dot), digital MISO d t i r a t e stabilizer with sampling period 120 ms for w (solid), and digital MIS0 multirate stabilizer 6 t h sampling period 180 ms for w (dash-dot) . .

4.29 Voltage disturbance with lead load: CPSS (dot), digital MIS0 multirate stabilizer with samphg period 120 ms for w (solid), and digital MISO mdtirate stabitizer with sampling period 180 ms for u (dash-dot) . .

4.30 Ground fault test with normal load: CPSS (dot), digital MISO multirate stabilizer with sampling period 120 ms for w (solid), and digital MIS0 multirate stabilizer with sampling period 180 ms for w (dash-dot ) . . 71

-- 5.1 The hybrid multirate filter bank. . . . . . . . . . . . . . . . . . . . . . . i i

5.2 The pre-filtered error system. . . . . . . . . . . . . . . . . . . . . . . . . 78

3.3 The equident discrete-time error system. . . . . . . . . . . . . . . . . . 85

5.4 Final equivolent system. . . . . . . . . . . . . . . . . . . . . . . . . . . . 87

5.5 1 ~ 1 (solid), IG~:,I (dot), and /G1l (dash) in dB versus w. . . . . . . . . . . 92

5.7 1 Fol (solid) and I& 1 (dash) in dB versus w / r . . . . . . . . . . . . . . . . 94

3.8 Steady-state error (solid) and desired output (dash) versus k. . . . . . . 95

5.9 Unit step response $ versus k. . . . . . . . . . . . . . . . . . . . . . . . . 96

5-10 Impulse responses verms k for Fo (upper) and FI (lower). . . . . . . . . 97

6.1 The cascade of two systems . . . . . . . . . . . . . . . . . . . . . . . . . 10 1

xii

Page 14: Optimal design of multirate systems

LIST OF SYMBOLS

superscript L

G(z) -

G ( s ) -

II-II

Hardy space

ditto

Hilbert space over the set of real numbers

Hilbert space over the set of integers

nest space

ditto

class of nest operators fiom y,- to Ur

open unit disk

unit circle

red-rational

orthogonal complement

G( l/z)'

G(-s)'

Euclidean nom

2-norm on &, 12: or ?&

-nom on 3C,

transpose of matrix -4

cornplex-conjugate transpose of matrix A

orthogonal complement of matrix A

trader matrix of system G

sampling operator

hold operator

sampler with period h

zereorder hold with perïod h

maximum singular value . .. Xlll

Page 15: Optimal design of multirate systems

inf idmilm

SUI? supremum

S direct s u m

J direct subt raction

Lm image

+ Redheffer star product

( ) linear fiactional transformation

xiv

Page 16: Optimal design of multirate systems

INTRODUCTION

In this chapter, we shall motivate our study of single- and multirate sampled-data

control systerns, present an o u t h e of the thesis and introduce some notation.

1.1 Single-Rate Sampled-Data S ystems

Digit al technology has brought dramatic change in process control instrumentation

and control system strategy. A digital control system, sometimes c d e d a sampled-

data control system, is a feedback system which consists of four basic components: a

plant, a analog-tedigital converter, a data-processing unit or digital controllerl and a

digit al-to-analog converter. The control loop samples and quant izes cont inuous-t ime

signds from the plant (process), processes the digital signals via a control algorithm.

and converts the renilting digital signals into continuous-time ones to regdate the

plant.

The schematic sampled-data control diagram is given in Figure 1.1, where the

four basic components are illustrateci by G (plant), Kd (digital controller), S (A/D

convener) and 31 (DIA converter). This system is hy6Rd in that it involves both

continuous- time and discrete-time signais. G is a continuous-tirne linear t i r n e - ~ ~ ~ a n t

(LTI) plant, which has two analog inputs, the exogenous input w and the control in-

put u, and two analog outputs, the regulated output t and the measured output

y. £Cd is a digital controller, which receives the measured sequence yd and generates

a control sequence ud. S is a sampler: It converts the continuous-the y into the

discrete-time yd by samphg y. 31 is a hold: It converts the discretetime ud into

the continuous-the u by holding ud between each two samples. In this setup? all

the signals involved are possibly vector-valued. To iimprove the clarity: we have used

Page 17: Optimal design of multirate systems

-7 L

continuous lines to represent continuous signais and dotted Iines to represent disnete

signals. This convention will be kept throughout the thesis.

Figure 1.1. Standard sampled-data setup

A/D and D/A conversions involve sophisticated technologies. The above sampler

and hold are idealized models which ignore quantization errors. The simplest A/D

and D/A scheme is to do alI the conversions at different channels symhronously with

the same period. Such systems are cded singlerate systems. .More precisely. a

single-rate scheme is the hllowing:

0 S = Sh: an ided sampler with a k e d period h, defined via

O 3L = HA, a zereorder hold with the period h: defined via

& is updated every h seconds, synchronized with Sh and Hh.

Due to the hybrid nature, the sampled-data -tem is harder to deal with than ei-

ther a purely continuous-t ime systern or a pure- discrete-time system. The sampled-

Page 18: Optimal design of multirate systems

:3

data system may be designed based on specifkations given in continuous tirne, discrete

time, or even both.

Thzre are in gened three methods to design digital controllers for sampled-data

systems. The h t is to do an andog design and then a discretization. This is the

simplest method because the analog specifications are natural and familiar. However,

in order to recover the analog pedormance specifications, usually one must use fast

samplers and computw. This introduces a trade-off between performance and hard-

m e cost. The second is to discretize the plant mode1 and then do a discrete-time

design. This is dso a simple method: If the original continuous plant is LTI, the cor-

responding discretized plant is also LTI. Also, since the sampling rate is incorporated

into the design process, one can attempt design with a slow sampling rate, thereby

reducing implementation cost. In Chapter 1, we will make a concrete cornparison

between these two methods for a power system stabilizer design. The limitation of

the second method is that poor intersample ripple may occur because the method

deds only with information at sampling instants. The third approach to design dig-

ital controllers is by a direct design in continuous tirne. This would be a preferred

approach because it solves the problem with no approximation. However, since a

sampled-data system is hybrid and tirne-varyhg, the method is mathematicdy less

tract able.

Single-rate sampled-data systems have b e n widely studied, as reflected in many

books, see, e.g., Kuo [Ml, Astr6m and Wittenmark (61: Franklin et al- [29] y and Chen

and Francis [17].

1.2 Multirate Sampled-Data Systems

In multi-input: multi-output ( M M O ) systems, it is sometimes advantageous to

use different rates at different sampling and hold chonnels. The resultant systems are

referred to as m u k a t e systems.

Page 19: Optimal design of multirate systems

Muitirate sptems are abundant in indust . Examples indude aerospace control

systems [a], power systems [10], robotic systems [48], chernical processes [77], and

digital Nter b& [72]. One reason to use mnltirate controllen is to achieve a cost

adçantage in implementation. In this thesis, we s h d study two different multirate

systems in control and signal processing.

The first system is a multirate stabüizer for a singlemachine infinite-bus power

system. The power system is in general a nonlineax system and consists of a s p -

chronous generator, a govemor, an exciter, an automatic voltage regulator (-liR)o

and transmission lines. The role of the power system stabilizer (PSS) is to provide ad-

ditional damping to low fiequency disturbances bj- measuring rotor speed and povver

and generating a voltage control signal which is fed back to the AVR and exciter.

By lineoripng the nonlinear model at a certain operating condition, one can obtain

a iomh order model of Figure 1.2, where Ap, Aw, 46, and Au, are the deviations

in power, rotor speed, power angle, and terminal voltage, respectively, AI& is a

reference voltage distur bance: and AT, is a mechanical torque disturbance. Gsing

the machine pasameters (for details, see Chapter 4): it may be found that the tram-

fer function from A L f to 4 p has a wider band of significant frequency components

than the transfer function fiom to Au. Hence, instead of using a single fast

sampling rate, one can sample the power signal fast and the speed signal slowly, hence

reducing the implementation cost.

To convert the powet system in Figure 1.2 to the standard setup in Figure 1.1, we

define (for convenience, we &op d the prefix A)

and r to be a vector of any group of LY., p and 6. Then Figure 1.2 is reconfigured

into Figure 1.1. In mdtirate sampling schemes, the sampler S and the hold 7f have

different rates. In this case, if we take the sampling periods for the power and for

Page 20: Optimal design of multirate systems

Figure 1.2. Linearized power system model.

the speed to be h and 2ht respectively, and the period for the hold to be h, then the

multirate sampler corresponds to two different rate samplers

and the single-input hold reduces to the single-rate one: H = HA.

The second multirate system to be studied in the thesis is a hybrid filter bask

which find applications in subband coding [72] and fast A/D conversion [53]. Shown

in Figure 1.3 is the 2-channel filter bank, which includes LTI analog analysis filters Go

and Gl, LTI digital synthesis fdters Fo and Fi, two slow samplers (Sah) with period

9h, and two expanders (E) with the factor of 2. The system is typically used to

process bandlimited analog signals, z(t ) , with Nyquist fkequency l/ h. Usually~ Go is

lowpass and G1 highpass. Fo and FI axe designed to achieve certain reconstmction

performance characteristics. Because slow samplers are used, such filter b d sys-

tems can achieve overall bit-rate reduction in tansmitting analog signals [Tl] : or- if

designed properly, can approximate the fast A/D converter with period h 1531.

Page 21: Optimal design of multirate systems

Figure 1.3- The hybrïd multirate filter bank.

1.3 Properties of Multirate Controllers

In general, multirate controllers are more complex t han single-rate controllers. The

desired multirate controllers should satisfy three proper5es: periodicity, causalit- and

finite dimensionality [19].

Periodicity is a direct generalization of time invariance from single-rate syst erns

to muhirate systems. As an example, consider the multirate power system stabilizer

with the sampling and hold strategy taken in the preceding section. It is easily

seen that the least common period for the two sampiing charnels and the one hold

channe1 îs T = 2h- To keep the overd system to be T-periodic, the two-input and

one-output stabilizer must be T-periodic. This idea can be extended to the general

MIMO controller case: Let T be the le& common period for all sampling and hold

channels; The overd system is T-periodic 8 the controller fi is T-periodic in real

tirne. It can be checked that to keep the periodicity for all sampling and hold channeisl

the skmpling and hold rates must be rationdy related (i.e., the ratio of any two rates

is rational) [19].

Causality and finite dimensionality are requïred for Kd to be implementable in

real time on microprocessors with finite memory. With finite dimensionaüty, Kd can

be described by state-space difierence equations. For the multirate power system

Page 22: Optimal design of multirate systems

stabilizer, we have the following equations for fi:

(Xote that we have used continuous s isals evaluated at respective sampling instants

to represent the corresponding discrete signais, because two sampling rates CO-exist.)

For causa&- of the mdtirate model, we must have a3 = 0.

Equations in (1.1-1.3) can be used for implementing the mdtirate stabilizer on

computers, the state vector z being updated every T seconds. This controuer &

is two-input, one-output, but it is ad~ontageous to view the model in (1.1-1.3) as a

three-input? t m u t p u t LTI system:

This corresponds to grouping the inputs and outputs properly over every period T.

Then i t follows from (1 .LU) that & has a statespace model given by

This three-input: two-output controller is almost like a standard MIMO. LTI discrete-

time system except for the causality constraint = O, arising from the multirate

structure. The method of obtaining such equident LTI models is formally referred

to as the lifting technique [38,44, 191.

In general, the multirate problem of Figure 1.1 can be recast using the lifting tech-

nique as a discrete-t ime singlerate problem involving only LTI systems in Figure 1.4,

Page 23: Optimal design of multirate systems

where & is an LTI plant and & is an LTI controller to be designed; mr a and

y are respectively the lifted exogenous input, control input, regdated output. and d

measured output. The lifted sigoals a and usually have higher dimensions than ud

and yd? respectively, and yd is nrbject to a causdity constraint: Its direct feedthrough

Figure 1 -4. Singlerate discret e-t ime system

term is a block Iower triangular matrix [44, 59, 74, 191. Multirate controller design

thus is reduced to constrained LTI controller design.

'ilultirate control problems have received much attention in the last four decades-

Recent studies inchde the parametrization of all stabilizing controllers [a' 591, LQG/

LQR designs [2: 45, 31: X2/31, designs [?5? 54, 74, 19, 551. Meyer [44] and Ravi et

al. [SI presented a controller parametrization method for multirate systems which is

the direct extension of the Youla parametrization (se, e.g.? [27]). This method was

used by [Gy 54: 74, 191 to reduce the multirate R2 and 31, control problems into

the constrained R2 and 3C, model rnatching problems in the fiequencq- domain: In

[4 Meyer noted that the causality constraint is a convex one and hence a numericd

technique based on convex optimization [11] was proposed for the multirate LQG

problem; Voulgôris et al. [74] proposed a projection method for the constrained R2

and 3C, p r o b l v Chen and Qiu [19] developed an effective framework based on

nest operators to handle the causality constraint and gave an explicit solution to the

Page 24: Optimal design of multirate systems

9

constrained R, model-matching problems using factorizations associated Nith nest

operators; this framework was also used in 1551 to obtain a characterization of aJ1

suboptimal 36. controllers.

1.4 Outline of the Thesis

ki the next two chapters, we study more computationally efficient methods for

both constrained 7i2 and 31, control problems.

In Chapter 2, a statespace approach to the constrained optimal and suboptimal

3t2 control problems is developed and explicit formulas for controllers are given in

terms of solutions of two Riccati equations. The solution process takes advantage of

nest operators to handle the causality constraint [19] and involves a series of modi-

fications of the standard Youla parametrization. The resdts include the solution to

the standard, unconstrained 3CÎ problem as a special case.

In Chapter 3, a different approach to obtaining constained 31, controllers is pre-

sented. More specifically, it is shown that the bilinear transformation can be used to

convert the constrained discrete-time Ra problem into a continuous-t ime R, pîob-

lem. As a resdt , a constrained discretetime 360 solution can be obtained in tenns

of the standard, unconstrained continuous-time & solution and the methodology

proposed by Chen and Qiu [19] to handle the causality constraint. Based on this

solution, we can take advantage of existing sohaxe, e.g., MATLAB, and modify

amilable ut ility functions to solve multirate 'tl, problems.

Chapter 4 includes a case study for Chapter 3 - a multkate 'fl, power system

stabilizer for a single-machine, infinite-bus power system. Since 31, design of power

system stabilizers has received an increased attention in recent years [JI, 5: 14, f 1: 11,

we present a comprehensive study with four ?Lw stabilizers designed and investigated:

(1) an analog SISO stabilizer measuring the speed signal and then irnplemented digi-

tally: (2) a SISO digital stabilizer designed in discrete time: (3) a rnultivariable digital

Page 25: Optimal design of multirate systems

10

stabilizer measuring both speed and power signds with a single samphg rate: (4)

a multirate digital stabilizer involving two samphg rates and designed using the

method in Chapter 3. Our designs give rise to controllers which have a relativel- low

implementation cost; in partidar, aIl stabilizers designed have low ordeq require

considerably slow samphg rates, and outperform the conventional (analog) power

system stabilizer on a series of nonlineax dynamic performance tests. The impor-

tance of weighting finctions is stressed in the design process.

Chapter -5 studies the multirate filter bank shown in Figure 1.1 using 'H, opti-

mization. Since the system does not f a in the standard sampled-data setup, there

is no method a d a b l e to carry out the design directly. We take the viewpoint that

the analysis filters have already been designed and -thesis filters are then to be de-

signed to minimize the worst-case energy gain of the error system, suitab- weighted.

between the filter bank and an ideal system - a fast A/D converter with some time

d e l - By using the lifting techniques. this hybrid and multirate problem is reduced

to one of 'fl, optimization involving only LTI, discrete-time systems, which is then

solvable. An example is diswsed and studied in detail to illustrate aspects in the

design process. In contrast to digital filter banks, performance limitations exist in

this hybrid filter bank.

Chapter 6 treats a fundamental problem motivated by both the filter bank design

md X2 and N, control: Uihen two finite-dimensional iinear time-invariant (FDLTI)

discrete-the systerns, causal or noncausal, are cascaded, under what conditions is the

cascaded system causal or anticausal? Wë answer this question in general. For the

special cases when one system is causal and the other is anticausal: we give necessary

and sufficient conditions. The results can be used to ver* whether a filter bank is

causal or not if the synthesis filters are chosen to be noncausal.

Finally, Chapter 7 offers some concluding remarks dong Nith some suggestions for

Page 26: Optimal design of multirate systems

1.5 Notation

The notation used throughout the thesis is quite standard.

In discrete time, the frequency-domain spaces 3L2 and 31, are the Hardy spaces

of mat&-vdued functions analytic outside the unit disk, Le., Ir1 > 1, with noms

defined as, respectively,

11611, = SUP W ~ m a z [ & e i ~ ) l ,

where * means the cornplex-conjugate transpose and cm, the maximum singular

value. Prefkv 12 means real-rational; hence RU2 and RR, are the real-rational

subspaces of Xz and respectively. 3Cz can be regarded as a subspace in the

Lebesgue space L2 dehed on the unit circle Ir1 = i. The orthogonal complement of

3L2 in L2 is denoted R$ with RN& its rd-rational subspace.

Similar notation is used in continuous time for the fkequency-domain spaces. For

example, iH, is the Hardy space defhed on open right-half plane Re(s) > O with the

ZR, is the real-rationd subspace of X,. The context will prevent confusion.

To represent signals, L2 denotes the Hilbert space of squareintegrable continuous-

time signals, perhaps vector-valuedt defined over the time set of a l l real numbers; the

inner product and corresponding nom on L2 are

where ' denotes transpose and the nom on t ( t ) is the Euclidean one. Similady. &

denotes the Hilbert space of square-summable discrete-time signals, perhaps vector-

Page 27: Optimal design of multirate systems

valued, defined over the time set of aU integers with the b e r product and nom:

For an LTI system G, either in continuous or discrete t h e , its trânsfer matrix is

denoted G. For a state-space realization (A, B, D) of an LTI system, we use the

packed notation

to represent its transfer matrix D + C ( s I - .4)-lB in continuous tirne or D + C(r I -

A)-'B in discrete tirne. For a discrete transfer rnatrix &), ~ ( z ) " is defined to

be G ( I / P ) ' , which corresponds to the adjoint system of G; for a continuous t r a d e r

rnatrix ~ ( s ) , the correspondhg operation G(s)- is G(-s)'.

FinaUy, given an operator K and *O operator matrices

the linear fractionai transformation associated with P and is denoted

and the Redheffer star product [60] of P and Q is

Here: we assume that the domains and co-domains of the operators are compatible

and the inverse exists. With these defmitions, we have

Page 28: Optimal design of multirate systems

CHAPTER 2

CONSTRAINED 3C2 CONTROL: A STATESPACE APPROACH

As noted in the preceding chapter, a multirate control probiem can be converted

into a singlerate control problem involvïng only LTI discrete-time systems. However,

the LTI controller to be designed is subject to a causality constraint on its direct

feedthrough term. In this chapter, we study design of the constrained controller with

an X2 optimali- criterion'.

With respect to the dismete-tÏme setup in Figure 2.1, the constrôined 3t2 problem

can be stated as follows:

Given G. find a contrder K which stabilizes G, satisfies a causality constraint.

and further minimizes the ?& norm f!rom w to z.

In Section 2.1, we will associate the causa&@ condition to nest operaton.

Figure 2.1. The lifted system

There are several ways for such a constrained 3L2 problem to arise:

'The r d t s in th5 chapter have been published in [6j] -

Page 29: Optimal design of multirate systems

14

In discrete-time periodic systems, a common technique to convert a periodic

design problem into an LTZ one is lifting [38]. However, causaIity in periodic

control systems puts a condition on the direct feedthrough terms of the lifted

systems: They must be (block) lower-triangular. R2 or LQG periodic control

leads naturally to the constained problem [74].

O In discretetime multirate systems, again one uses lifting to get LTI systems

[@, 591; but the LiRed controllers s a t i e a similm causality constraint: Certain

blocks in the feedthrough terms must be zero. An LQG multirate control design

gives rise to the constrained 3Lz problem [45].

In X2-optimal control of sampled-data multirate systems, by a certain reduction

process [75,64] based on a continuous lifting technique: see, e-g., (81, one arrives

at the same discrete-time constrained X2 problem.

Recent studies on the constrained X2 problem are ail in the fiequency domain,

based on the parametrization of d stabilizing controllers [44: 591 to reduce the prob-

lem into a constrained 7i2 model-matching problem [45? 54: 14. 191. In this chapter,

we present the explkit statespace realizations for the optimal and suboptimal 3C2

controllers satisfying the causrtlity constraint.

2.1 Causality and Nest Operator

In this section we s h d characterize the causality condition on by nest operators

as in [NI. Let u = Ky. If the direct feedthrough term in EL is DA-: we have u(0) =

Dr; ~(0). Arising from lifting, u(0) and y (O) are vectors of the form

The elements in the two vectors may occur at different time instants in a real-time

system, sas yi(0) at tirne ti and uj(0) at hj . Then the causality condition means that

Page 30: Optimal design of multirate systems

15

the output ~ ~ ( 0 ) cm only depend OR inputs occurring up to hj-

Let C be the set of t h e instants ti and hi; Iet n + 1 be the number of elements in

this set (not counting repetitîons); order E increasingly (cr < or+&

c ={O, : r=0,1, ..., n).

For r = 0,1, ...' n, define

y, and Ur correspond to, respectively, the inputs and outputs occurred after and

inciuding time or during the Iarger period associated with the lifting. It follows that

{y,) and {Ur) are nests of subspaces s a t i e n g

and that the causaliq condition on K means that DK is a nest operator.

The set of d such operators is wrïtten N({Y,), (U,)) and abbreiiated N((yr)) if

{Yr) = {U,). In the foflowing lemmas, the spaces involved are ail finite-dimemional.

Lernma 2.2 Let D be an opemtor on Y .

Page 31: Optimal design of multirate systems

16

(a) There ezist a unitary operator Ul on Y and an operator & in N({Y,)) svch

thut D = UIRl-

(6) There ezrst a unitary opemtor U2 on y and an operator R2 in N({y,)) such

that D = R2&.

Lernma 2.3 Let D be a nonnegatiue Hennitian opemtor on Y. Then there ezist

operators Ri, R2 E N({yr) ) sach that D = R; Ri = R&.

The results in the lemmas f011ow easily fiom matrix theory by noting the relation-

ship between nest operators and (block) lower-triangular matrices: For an operator

D mapping Y to 24, if we decompose the spaces Y and U in the following way

where 3 and @ denote the direct subtraction and direct sum, respectively, then the

associated matrix representation of D is

D E JV({y,)z {U,)) means that this mat* representation is (block) lower triangular,

i-e., DG = O if i > j-

With these, the constrained 7i1 problem is as follows: Given G: design a stabilizing

K sati-g &(oc) E N({Yr), {U,)) to minimize ~~T'llz.

2.2 The Unconstrained Case

The solution to the constrained ?12 problem will relate to that of the standard?

unconstrained X2 problem.

Page 32: Optimal design of multirate systems

Uk start with B date-space model for G:

The standing assumptions for the unconstrained N2 problem me given as follows:

(i) (-4, Bz) is stabilizable and (Cz, A) is detectable;

(ii) M := Df,Di2 and :Y := are nonsinylar;

(iii) For any r on the unit circle 3Zi: the matrkes

have fdl column and row d, respectivek

These assumptions are standard in 'Hz and 3L, literature [3'7]. Assumption (i) is

n e c e s s q and sufficient for the existence of stabilizing controllers. Based on this

wumption, there evist a state feedback matrLz F and an output injedion matri2 L

such that A + B2 F and A + CCz are stable. The optimal ?L2 solution is obtained

b - suitab- specifyïng F and L (see below). Assumption (ü) means that the control

and sensor weightings must be nonsingular matrices, which implicitly assumes that

the system must have at least as many outputs to be controlled as control inputs and

at l e s t as many =ogenous inputs as measured outputs. This assumption insures

that the 3L2 problem is nonsingular. Findy, 'flroptimal controller c m be obtained

by solving two Riccati equations, whose stabilizing solutions ore guaranteed to be

unique by assump tions (i)- (iii) . Let -4, QI R be real n x n matrices with Q and R symmetric. The solutions of the

algebraic Riccati equation

Page 33: Optimal design of multirate systems

is known [32] to relate to the generahed eigenspaces of the sympiectic pair

If H has no generalized eigenvalues on ûV and the two subspaces

are complementary, where x(H) denotes the eigenspace of H corresponding to gen-

eralized eigenvalues inside D, then &(H) defmes uniquely a stabilizing solution of

(2.2): denoted Ric(X); see, e.g.? [3Tl 171.

Now the ?&-optimal controuer Kwt can be given as follows. Introduce the sym-

plectic pairs & and Sy and the matrices F, Fo, Lt Lo:

Then h,t is given by the following state-space mode1 [16]

Page 34: Optimal design of multirate systems

It cm be verified that A+ B2F and A+CC2 are stable and the closed-loop tmnsfer

The optimal solution can be obtained as follows: Solve the following two Riccati

equations for X and Y, respectively:

and then compte Fo7 L, Lo to get K , . Note that &&o) = Lo7 which does not

belong to N({yr) , {Ur)) in general. Hence the optimal controller for the constrained

problem is tqpically different from this Kqt.

2.3 Main Results: The Constrained Case

In this section, we s h d develop complete state-space solutions to the constrained

R2 control problems based on the three assumptions in the preceding section.

Introduce the same matrices: XI Y: F, Fo, L, Lo, as in Section 2.2. Since A + B2F and A + LC2 are stable: the set of intemallq- stabilizing controllers for G can be

characterized by the following Youla parametrization foxmula [ ~ f ] :

Page 35: Optimal design of multirate systems

To sirnpliS; rnatters to follow2 we modify this parametnzation as follows: With the

mat& Lo defined in Section 2.2, replace the parameter Q by Lo + Q and then absorb

Lo into &; this leads to the following characterization of stabilizing controllers

where

Such controllers May not sati* the causality constraint? which will be considered at

a later stage.

Uiith this controller parametrization, the closed-loop system T' can be given in

terms of Q:

where f is the Redheffer star product G * j [60], whose state-spâce mode1 can be

computed h m the state-space models of G and j:

It follows that

with

Page 36: Optimal design of multirate systems

and

T' =,+T~QT~~ Q €ER2-

Further examining fi, we obtain &er some computation

Tl = TF + P2fL,

where

(The matrix Fo is defined in Section 2.2.) Hence, our constrained 3Cz control problem

becomes: Find a Q E RN2 wïth

to minimize the 3L2 nom of

Xow we consider some causality issues raised in Section 2.1. Introduce factoriza-

tions for the two positive definite matrices R and S by Lemma 2.3,

s = SIS;,

where Ri and Si are invertible nest matrices satisfying Ri E rV({Ur)) and SI E

,Ve({X}). Xote that the matrix RIL& maps y to U.

Recall that hp({yr}, {Ur)) is a subspace in the space of matrices mapping Y to Li:

the latter can be regarded as a Hilbert space with the 2-nom (compatible with the

3L2 nom for transfer matrices). Let NL({J&), {Ur)) be the orthogonal complement

of n i ( {X) . (Ur)). Decompose

RILoSl = W + WL

Page 37: Optimal design of multirate systems

FVe are now set up to state the main result of this chapter.

Theorem 2.1 The optimal controller solaing the constrained 3L2 problem is giuen by

Moreover, the minimum cost is giuen by

The proof will be given in the next section. From the theorem, we note that

Km in (2.7) satides the causaity condition: By the facts that RI E N({U.)) and

S1 E M((yr)) and Lemma 2-1, V E M({X) , {U,)) iff RIVSI E N ( { X ) , {U,));

but RIVSl = W E N({Y,), {Ur)). -41~0, in the expression of & in (2.8): the

hs t two terms correspond to the unconstrahed problem; and hence the last tenn,

Il WLll?. is due to the causality constraint. Based on Theorem 2.1, rhe computation of

the constrained R2 solution just reg& a simple modification to the unconstrained

solution in (2.3).

In control system design nith multiple objectives [39], ofken one is interested in

suboptimal controllers. The proof of Theorem 2.1 yields easily a characterization of

all suboptimal 3C2 controllers. For completeness, this result is stated below.

Theorem 2.2 Let 7,t be given as in Theotern 2.1. Assume y > yWt. Then the set of

al1 stabilicing controllers which satisfy the causality constraint and achieve 11f"112 < Y

is giuen by

Page 38: Optimal design of multirate systems

2.4 Proof of the Main Results

The proof of Theorem 2.1 requires the fouowing lemma.

Lemma 2.4 With the matrices R and S used in Sections 2.2 and 2.3, we have

(a) f ; f F E R3ck and f;P2 = R;

(6) E and = S-

Proof We s h d prove part (a); part (b) follows similady. Define

AF = A+B2F,

Gr = Ci+Dl2F,

BIF = BI + &Foy

D l l ~ = Dll + Dl2F0-

The matrix X satisfies the Kccati equation in (2.5) associated with the s!mplectic

pair Sx , and hence satisfies the equivalent L yapunou equation (after some algebra)

This equation admits a solution in the

00

.Usq it can be vedîed that the following identity holds:

Page 39: Optimal design of multirate systems

24

To see E R;H:kl write in te- of povser series of r:

Each Ek is an infinite sum; however, using the infinite sum solution (2.10): we can

simpw Ec for k 2 O as follows:

Then it follows h m the definitions of Dllr, BIF and Fo that

and h m (2.11) that Ek = O for k > 1. Hence, pcfF E RRi. Similar ilrgument dl

give TTT2 = R. If

Sow we proceed to prove Theorem 2.1.

Proof of Theorem 2.1 We staxt with T, in (-2.6): namely,

Fm = FF + T;(TL + QT;).

Since FcZF E R3c: [Lemma 2.4 (a)], the two tems on the right are orthogonal in

X2 and hence

the last equality following from the fact that PTT; = Ri Ri. Similady, by Lemma 2.4

(b): we have

Page 40: Optimal design of multirate systems

II~JI: = [IF& + I I R ~ ~ ~ I I ; + IIR~QS~II;.

To simplî@ further, redefine R ~ Q s ~ as Q. (Note that R ~ Q S ~ E 7E31* iff Q E RR2

because RI and Si are invertible.) This corresponds to modifying the controller

parametrization into

w here

and then

I I~~UI I ; = I I ~ F I I ~ + ~IR~TLI I~ + 1 1 ~ 1 1 2 - (?.la)

Without the causality constra.int, the optimal Q is obviously zero. With the causal-

ity constraint, we must have

This is the sarne as

by Lemma 2.1. For (2.14) to be tme, ~ ( o o ) must completely cancel WL, the corn-

ponent of RILoSt in NL({X), {&)); Le.: Q must be of the folloming orthogonal

decomposit ion

Substitute (2.13) into (2.12) and absorb -WL into .fo to get a parametrization of

controllers sati&ing the causaliw constraint :

Page 41: Optimal design of multirate systems

Some calculation and the definition of V show

Also, substituting (2.15) into (2.13), we have

since WL and are orthogonal in 7f2.

'Yow the results follow easily: The optimality is achieved when = O: put this

into (2.16) to get kopt as in (2.7)- c

2.5 Conclusions

In this chapter, we have obtained the complete state-space solutions to the optimal

asd suboptimal R2 control problems mbject to a causality constra.int . The solutions

require no additional assumptions: The constrained % problem is solvable iff the

standard: unconstrained 'flz problem is solvable. Compared to the existing solutions

in the fiequency domain [74? 19: 451 the results have the advantage of cornputational

efficiencg and ease of implementation on cornputers. Li fact , the algorit hms developed

in this chapter have been implemented in a MATLAB multirate control software which

is cmently under development [56].

For ease of derivation, we have assumed D22 = O to obtain the constrained X2-

optimal controuer. This results in no loss of generality. In fact, if Dn # O: typically

due to the lifting procedure [44, 191, thus it is an easy matter to veri- that the

optimal contrder l?lnnu,wz for Dn # O is [76]

Page 42: Optimal design of multirate systems

S i

where h',, is the optimal controk for the case taking Dn = O and is given in (2.7).

'iote that hrnwrwt satisfis the causality constraint by Lemma 2.1:

The results in this chapter are usefid in studying multirate control systems: A case

study was given in [56] for design of multirate 3C2 power system stabilizers; another

application to optimal multirate discretization of analog controllers may be found in

[ s i ] - Fina&, we mention that the results in this chapter can be extended directly to

the case involving a multiple objective 7tz measure in the sense of Pareto optimality

dong the lines of [39] and the techniques in this chapter to hande the causality.

Page 43: Optimal design of multirate systems

CHAPTER 3

CONSTRAINED 3t, CONTROL: AN ALTERNATIVE APPROACH

Since a bilinear transformation of a system p r e s e m its 36, nomst one ofien con-

verts discrete R, controuer design into a continuous one via the bilinear transforma-

tion. The reason is that continuous %, design requires considerably less computation

(compare [33] for the continuous-time 3C, solution with [37] for the discrete-time 31,

solution). For the Rw design with a causality constraint on the controller feedthrough

terms, solutions have been given numerically via convex optimization by Voulgaris et

al. [71] and explicitly by nest operators by Chen and Qiu [19, 5S]. Hence the follow-

ing question arises: Can we do the constrained disaete-time 31= design also by the

bilineôr transformation? This chapter attacks this probIem if o d y one suboptimd so-

lution is sought. More spe~ifically~ we discuss how to use the bilinear transformation

toget her with the methodology proposed by Chen and Qiu [19] to obtain a constrained

36. solution. The resdts will be used for designing multirate power system stabilizers

- a topic studied in the next chapter.

In viem of Fieme 2.1, the constrained ?fw suboptimal control problem c m be

stated as follows:

Giwn G and 7 > 0, h d ô stabilizing K with ~ ( m ) E ,V({Y,). (24,)) to achieve

llFmll, < 7: if such K exists at d.

In what follows, we take y = 1.

Similar to the constrained 3C2 probIem, the constrained 7& problem may arise

in N, designs of discrete-time periodic systems, discrete-time multirate systems and

sampled-data multirate systems. For example, to design a sampled-data multirate

system via the 3L, approach, one can reduce the design probkm into a single rate LTI

Page 44: Optimal design of multirate systems

problem of 31, optimizatioc via continuous and discrete lifting techniques [X, 191.

The liftings put the causality constraint on the direct feedthrough terms of the Iïfted

LTI controllers-

3.1 An Existing Solution

Uë first summarize the solution process in [19], which is given in the fiequency

domain. -4 key step there is the parametrization of alI 31, suboptimd controllers in

the unconstrained case. Since such a result is adab le using the state-space methods

[37]. we s h d proceed further from here. The state-space mode1 of the plant is assumed

to be

G ( Z ) = [y-+] - (72 D21

W e have assumed that D22 = O. The extension to the case with Dm # O follows dong

the lines of discussion in Section 2.5.

Let us first drop the causality constraint and find a characterization of a l l stabi-

lizing controllers satisfiing 11 f, 11, < 1, which is given in [3?]:

where

with D~~ and ~~1 being square and invertible. We refei to [37] for details of diecking

the expression of It follows that &oc) depends on ~ ( m ) in an affine way:

In order to get k(m) E Aœ((y,),{Ur)), &oc) must be specified. Since l l ~ l l ~ ~ 2

11 &(m) 11 the equident problem is to find a constant matrix oc) xith 11 Q(W) 11 c i such that k (m) in (3.3) belongs to N({yr), {Ur)).

Page 45: Optimal design of multirate systems

B y Lemma 2.2? introduce rnatrix factorizations (QR factorizat ions)

where Ri, R2: Ui : & are aU invertible, U;, U2 orthogonal, and R1 E N({U,}), R2 E

,V({Yr}). Substitute the factorizations into (3.3) and pre- and post-multiply by RF'

and RF' respectively to get

we rewrite (3.1) as

P = T - W .

It follows that &cc) E ,V({Yr), {Ur)) iff W E ,V({Y,), {Ur}) (Lemma 2.1) and

1 1 ~ ( c o ) 11 < 1 iff 1 1 Pl[ < 1. Therefore, we arrive at the following problem: Given T'

find W E N ( { X ) ? {Ur)) such that

This problem can be related to a mat& distance problem [21, 191: Given T I find a

matrix W E N ( { X } , {&)) to minimize IIT - WII. Let

The problem in (3.5) is solvable, and hence an 'MW-suboptimd controller satis@ing

the causality constraint exists, if p < 1.

To summarïze? assume p 5 1. Define WW such that

Page 46: Optimal design of multirate systems

and

3.2 A Solution Process via Bilinear Tkansformation

The bilinear transformation is a bijective function between lzl > 1 in the r-plâne

and Re(s) > O in the s-plane:

Since the bilinear transformation preserves 31, noms for trader matrices (se, e-g..

[76j), one can convert a discrete-time 36, problem into a continuous-the one.

The standard, unconstrained 76, design via bilinear transformation involves the

following steps:

Step 1 Convert the discrete plant G into a continuous one G, via bilinear

transformation.

Step 2 Design a continuous X, controuer Kc for G,.

Step 3 Do inverse bilinear transformation to convert Kc into K: which is a

discrete RO3 controuer.

To carry out Step 2. one can use the standard r e d t s in [33,22] which is implemented

as the function hinfsyn in MATLAB p-halysis and Spthesis SooIbox [ i l -

Page 47: Optimal design of multirate systems

3 '5

For the constrained R, problern, we need a chamcterization of all 'W, suboptimal

controllers, which is, in te- of the results in 133, 221,

where

~ 5 t h DCl2 and being square and invertible matrices. Hence, the conesponding

characterization of all discrete-time '& controllers is

where

In the above, the bilinear transformation has done the foilowing operations: It con-

and to L~ via (see, e-g., [17])

In the next section, we will prove that D~~ and hl are invertible matrices. These

properties are important in treating the causality constraht.

3 o w we consider the causality issue based on the chasacterization of (3.7) that is,

find ô ~ ( o o ) with 11$(0o)ll < 1 such that %(m) E N({y,) , {Ur)): where

Sote that now &(oc) relates to Q(CG) by a linear fiactional transformation; we can

follow a similar step in [19. 551 to simpli& this relation. In the following, we assume

Page 48: Optimal design of multirate systems

that is normalized with

and perfonn a transformation

Q = ~ ( ~ 0 1 ) -

Ey [60], this transformation translates 1 1 ~ 1 1 ~ < 1 into 118, 11- < 1- It foUoWs that

with

Now it is an easy matter to verify that iz2(oo) = O? and iiz(m) and L2&) are

nonsingdar since we have assumed that D12 and hi are invertible. In this way, we

arrive at an affine rnapping ~ ~ ( o o ) H ~ ( o o ) :

The problem then becomes: fmd a Q ~ ( C Q ) with ~l~~(rn)ll < 1 such that h;(o=) E

,V({K), {Ur)). This is similar to the problem studied in the preceding section asso-

ciated with (3.3) and solvable with the obvious modifications.

3.3 Proof of Invertibility of and D~~

Ln the preceding sectionl the invertibility of Dl* and Dzl played an important role

in deriving the constrained solution: It guarantees that î12(00) and tzl(m) are

invert ible, and hence the associated QR factorizations have invertible factors. In t his

Page 49: Optimal design of multirate systems

34

section, we prove these properties in terms of continuou-time ?& solutions. For ease

of reference, we nunmarize the properties as a claim.

Clairn d12 and Dzi ore invertible.

The proof to follow will be based on a simplified version of 3C, theory proposed in

[2]. This treatment captures the essentid features of the general problem [33], but

simplifies the derivation a great ded. The general proof is much more algebraically

involved and is omitted-

The simpEed plant mode1 considered in [22] is

which implicitly assumes that Ddl = O ônd Dc22 = O. Moreover, the following

assump tions are made:

(i) (-4, Bci) is stabilizable and (Cci, A,) is detectable;

(ii) (-Act Bc2) is stabilizable and (Cd, &) is detectable;

Wit h t hese; the cont inuous-the %--subopt imal controllers are given as follows. As-

sume the conditions (i)-(iv) are satisfied. The set of ail stabilizing controllers such

that IIF(G~;,. &) 11, < 1 are chasacterized by

Page 50: Optimal design of multirate systems

with Xc and Y , sati-g the Riccati equations

respectively? and

zc = I - y,&.

Examing Mc: we see that BCl2 (= 1) and bal (= I ) are invertible. Mso? if we

it can be verified that Ac + BcFc and Ac+ LcC, axe stable (their eigendues are in the

open left-half plane). The claim needs two expressions for A,; whkh are first given

as follows.

Lemma 3.1 A, c m be -tten in two versions:

Proof The first equali- follows easily from (3.12). To see the second one, we need

to apply two Riccati equations (3.13) and (3.14): In view of (3.12)-(3.15),

Page 51: Optimal design of multirate systems

From here, (3.1 7) then follows.

Uë are now ready to prove the claùn.

Proof of the Claim In view of (3.8) and (XI?),

It foIlows that Dlz is invertible iff 1 - (Ac + LeCC) is. The latter is true because the

eigendues of Ac + LcCc can only be negative. Similady, using (3.16): we can show

t bat

821 = [I - (Ac + BcFc)][I - (4 + BcFc + C&Bcl ) l - ' ,

and hence the çtability of rl, + B,Fc implies that & is invert ible. O

3.4 A Design Procedure for Multirate X, Controllers

The constrained ?& solution can be used to design multirate R- controllers - in

the next chapter we will study multirate R- stabilizers for a power system. A t this

point, it is appropriate to summârize a design procedure based on the results of this

chapter.

Let us start with the standard multirate sampled-data setup of Figure 3.1. Uk

assume that G is LTI with a state-space mode1 given. S and X axe mdtirate sampling

Page 52: Optimal design of multirate systems

Figure 3.1. Standard multirate sampled-data setup

and hold operaton defined via

where mi and n j are integers and h is a real number referred to as the base pen'od.

The above operations correspond to sampling the i-th channel of y with period n i h

and holding the j-th channel of u with period njh. Based on the analysis in [19]: S

and U can be fixther factored as

where Sh and Hh are fictitious single-rate sampler and hold, respectively? and

are discrete multirate sampler and multirate hold

defined via

respectively, with Srni and H,,

+ = Smi4 +(k) = +(kmi),

t. = Hn, v(knj + r ) = 4(k), r = 0?1, - O - :nj .

The design procedure goes as follows:

Page 53: Optimal design of multirate systems

Step 1 Disaetize G into Gd. There are in g e n e d two ways to do this. The

first is by the sarnple and hold conversion; this corresponds to the traditional

discretization with sampling period h. The second is by a nom-presening

conversion; see [19]. Either Say, Figure 3.1 is converted into a multirate discrete

Figure 3.2. Multirate discrete-the system

design setup of Figure 3.4.

Step 2 Apply lifting, Le., convert Figure 3.4 into a single-rate discrete LTI

setup of Fiope 33.. The general formulas for the multirate lifting may be found

in, e.g.? [19]. Then we arrÏve at the constrained 3t, controi problem.

F i p e 3.3. The lüted LTI setup

Page 54: Optimal design of multirate systems

39

Step 3 Design & to stabilize G7 s a t i e the causality constraint, and achieve

This is sohable by the discussion in Section 3.2.

MATLAB provides a nice platform to implement the above procedure. In fact.

we have developed a software package to design rnultirate 31, controllers using the

bilinear transformation method proposed in Section 3.2. To summarize, let us brie&

discuss the routines we have written, as illustrated in Figure 3.4.

a plant: Store a statespace mode1 of the continuous-the plant.

a dmet-p: Discretize the plant by traditional sample and hold conversion. If one

would Like to do a direct sampled-data design, a nom-preserving discret k a t ion

method is atailable ini e.g., [1'7].

0 m l i ' Multirate lifting of the discretized plant. According to [19], we need to

lift each of sub-trader finctions of the discretized plant, which is implemented

in the function Zfprt, and then combine them. The function rnlift is to get a

state-space mode1 of the whole 1if'ted plant, çd, by calling Zfprt.

hisynaldc: Discrete-time 'H, design for the lifted plant. The output is a charac-

terkation of aIl unconstrained R, suboptimal controllers. This function carries

out the t h steps given in Section 3.2: It converts the lifted plant into a contin-

uous one via bilinear tansformation (BL); it c d s a continuous design function

hisyn-al which generates a characterization of all continuous 7.1, suboptimal

controllers; it then converts the controllers into discrete ones via the inverse bi-

linear transformation (BL). The new huiction hisyn-al is a modification of the

MATLAB built-in function hinfsyn, which cornputes only one continuous 31,

solut ion.

Page 55: Optimal design of multirate systems

Figure 3.4. A design procedure for multirate 310, controller

e lff2l-d: Simplify the linear fractional relation in (3.9) into an ofiine function in

(3.11).

perm: Permute the columns of kli(OO) and &1(00) aad the rom of L~~(oG) and &+û) in (3.11). The reason to do this is that the elements of the input

and output of the iifted controller may not be ordered in the time order due

to the lifting procedure: hence the causiility constraint may not necessarily

Page 56: Optimal design of multirate systems

41

imply that the direct feedthrough texm of the Mted controller is a block lower

tnangula matrix. The fùnction p e m is buiit to reorder the columns of Lll(oo)

and i21(~) and the rom of Lll(oo) and &cc) according to the t h e order

of the controuer input and output, thus the methodology in Section 3.1 can be

applied to handle the causality constraint.

go-dist: Reduce the &e function in (3.11) into a mat& problem by doing

QR factorizations for itz(ffi) and &O) (reordered) . The M N L AB built-in

function qr can be used for the QR factorization for tI2 (00) - Another function

prl is built for the dual factorkation required for &(m).

disprb: Solve the distance problem in tenns of the solution process given in

pi1 191.

kd-causa[: Finab, calculate the controller which satisfis the causality con-

straiat .

Conclusions

In this chapter, we have shown that the bilinear transformation method: usually

used in the standard, unconstrained discrete 31, design, is also appropriate for the

U- problem Nith a causality constraint on the c o n t r o b feedthrough terms. The

methodology used to handle the causality constraint followed closely that in [19], but

the solution is more computationdy efficient. Based on this, we have developed a

MATLAB software package with severd utility functions and subprograms to design

general discrete-time multirate controllers using 31, optimization.

One limitation of the bilinear transformation method is that it yields only one

constrained 3I, solution. In contrast, the direct discrete-time design approach can

give a,II Xfl, suboptimal solutions [S].

Page 57: Optimal design of multirate systems

3C, DESIGN OF DIGITAL POWER SYSTEM STABILIZERS

This chapter looks at an application of ?& optimization: Robust digital stabilizers

(controllers) are designed for a singlemachine infinite-bus power s yst em' . These

digital stabilizers. including a multirate stabilizer ushg the technique given in the

preceding chap ter, outperform the conventional anaIog stabilizer and require only

low cost in implementation.

4.1 Motkation and Introduction

The system to be studied is the singlemachine, infinitebus power system whose

schematic diagram is shown in Figure 4.1. In the power system? the spchronous

generator generates power which is transmitted through the transmission lines with an

infinite bus the exciter and automatic voltage regulator ( .N IL ) are used to maintain

the terminal voltage profde; the associated governor monitors the shaft frequency

and controls the mechanical power and speed. Often in operation this nonlinear

system is subject to various disturbances due to changes in, e-g., generation schedulest

transmission-line structures, load condit ions, and network interconnections; in order

to provide damping to oscillations caused by these disturbances, the power system

stabilizer (PSS) is instded via modulation of the generator excitation. Design of

robust digital power system stabilizers against different disturbances is the focus of

this chapter.

Conventional power system stabilizers (CPSS) are widely used in industry for their

simplicity in structure and hence ease in implementation via analog circuitry. These

analog stabilizers have low complexity; for example, a tq-pical cIass of stabilizers

'The tesults in this chapter have ben reported in [68].

Page 58: Optimal design of multirate systems

1 - 1 Transmission

Figure 4.1. A schematic diagram of the power system.

axe of fourth order, with a ht-order prefdter, a second-order lead-lag compensator.

and a first-order washout factor in cascade connection. The analog CPSS design

is usually based on an LTI model linearïzed about the normal operating conditions

and employs classical lead-lag compensation techniques, typically graphicd in the

frequency domain. Stabilizers designed using this graphical technique offer a certain

degree of robustness against parameter variations and &ovm disturbances; some

improvement can be made using the s o - d e d enhanced technique [47] which provides

extra compensation within a certain range of operating conditions.

The ?&,-based optimization can considerably Ïmprove performance and robust-

ness of power systems, as is evidenced by recent work on analog stabilizer designs

[SI, 5: 14, 7.1, 11. 'H, optimal control design minimizes the worst-case energ'- gain

(R, nom) of a certain dosed-loop trander mat&, suitably weighted. For properly

selected weight ing functions, the optimal controllers designed have good performance

in the face of uncertainties in plant modeling and/or disturbances; moreover, tradeoffs

between performance and robustness can be studied in this framework. One property

of the 31= optimal controller is that its order equds, roughly speaking, the order of

Page 59: Optimal design of multirate systems

44

the plant plus orders of the weighting functions. This can be a drawback if the plant

and weighting functions are of hi& order; in this c- mode1 reduction techniques

are usually applied after the optimal design: e-g., in [Tl], the power system controUer

designed is initidy of order 20, which is then reduced to 10; in [14], the order of the

stabilizer designed using p synthesis is reduced fiom 35 to 10. In this study. we target

stabilizers with orders comparable to the conventional power system stabilizen.

Our main purpose is to design digital power system stabilizers which can be im-

plemented on microprocessors with low cost. To meet this objective, the digital

stabiIizers designed have low complexity and require only slow A/D and DIA con-

versions. We s h d attempt two approaches to design 36, stabilizers. The fmt is to

design analog îi, controllers and then discretize them for digital implementation.

This approach involves approximation in discretization and normally requires high

sampling rates in order to emulate the analog system - see one of the PSS designs

studied later. The second approach is to fix the sampling rate and discretize the plant.

and then perfonn 'H, control design directly in discrete tirne. This approach has the

advantage that because the samphg rate is incorporated in the design process. one

can attempt design with slow sarnpling rates. However, weighting functions should

be selected properly to avoid overdesigning the system at sampling instants. Sev-

eral of our designs studied later follow this second approach, yielduig power system

stabilizers with high performance and low sampling rates.

Several control schemes and robust stabilizer designs are investigated for digital

implementation. Stabilizers can have access to a single input, the speed signal, or

two inputs, the speed and power signals, and generate a voltage control signal for the

automatic voltage regulator (.4VR.) and exciter; in the two-input case, the sampling

schemes can be single-rate or multirate. Thus four robust digit d designs are studied

in t his chap ter:

Page 60: Optimal design of multirate systems

43

1. The stabilizer is SISO, meanving the speed signal; a robust analog stabilizer

is designed first using 'H, optimization and is then discretized via bilinear

transformation (Tustin's approximation). To maintain good performancet fast

sarnpling is required; the samphg period in this case is 10 ms.

2. This has a sim3a.r control setup to Case 1 - SISO and single-rate, but the plant

and weighting functions are discretized first and the digital stabilizer is designed

based on discrete-time 31, optimization. To keep comparable performance, we

can reduce the sampling rate by a factor of 4, the sampling period being 10 ms.

3. ùi this case, the control setup has two inputs and one output - MIS0 (multi-

input, singleoutput) setup, and a single sampling rate is used. The design

is based on the discretized model as in Case 2. Because an additional power

signal is meanired, the sampling rate can be further reduced without sacrificing

performance; the sampling period in this case is 80 ms.

4. Similar to Case 3, the stabilizer measures the speed and power signas but

we choose a multirate sampling scheme. Since multirate system is the main

subject in this thesis, we present two sampling schedules to compaze: In the

iirst one, the speed and power signals are sampled with periods 120 ms and

60 ms, respectively, and the control signal (DIA conversion) is updated every

60 ms. In the second schedule, the samphg period at the speed is further

increased to 180 ms. One reason for using multirate controllers is to achieve

a cost advantage in impIementation in cases when the signds involved have

different bandwidths.

AU the digit al power system stabilizers designed compare favorably 6 t h the analog

conventional one, as will be shown by a series of pedormance tests via nonlinear

simulations. Moreover? the digital stabilizers have relatively low orders (46) ; this is

Page 61: Optimal design of multirate systems

46

important for rd-t ime implementation. This property is obtained by using a low-

order linearized model in design (this is partially justified by the robustness of the

design approach taken) and carefidy selecting simple weighting functions.

The multirate design problem imlved in Case 4 can be solved using the bilinear

transformation method given in Chapter 3 - a 3IATL.A.B softwaze package has been

dedoped t O implement the multirate design approach. The software takes advantage

of the existing MAT LAB fiinctions: in particularo the cont inuous-time 'H, design

function hinfsyn is used as one of the key functions for design - see Section 3.1 for

det ails. The result h g mult irate robust stabilizers have good performance comparable

with single-rate stabilizers, but use lower skmpling rates.

4.2 Linearized Plant and Weighting Functions

A spchronous generator connected to an infinite bus is in generd described by

noniinear difierential equations with combined order seven. As we mentioned eariier.

for low-order stabilizers: it is advantageous to use a low-order hearized model in

?lm optimization. Hence we use the simplified Park's two-axis model [4]: obtained

by ignoring transients in the stator circuit and the effect of the rotor amortisseur -

see the Appendix at the end of the thesis for the equations and the parameters which

are derived from an experimental singlemachine power system at the University of

Calgary.

Linearizing about the normal operating conditions [power P = 0.9 p.u. and power

factor Pj = 0.95 (hg)]: we obtain a linear model [4] in Figure 4.2 for the power

system? where Ap, Au, 46: and 4~ are, respectivelq; the deviations in power, speed,

power angle, and terminal voltage. The reference voltage disturbance is 4 Kej and

the mechanical torque disturbance AT,. For ease of reference, we will drop the prefix

Page 62: Optimal design of multirate systems

av I

Figure 4.2. Linearized power system model.

Table 4.1. Parameters for the lùieaxized power system model.

Parameter Value Parameter Value '

4 in a.lI variables from now on. Xote also in Figure 4.2 that a simplified model is

given for the AVX and exciter block which is in the form of a first-order system:

Kz Tas + 1'

The parameters in Figure 4.2 can be determined from linearizing the equations in the

Appendix with the given operating conditions; these are shown in Table 4.1.

with this linea,rïzed model in Figure 4.2, the overd order is four. Note that the

power system stabilizer is not included in Figure 4.2; but it measures w or both s. and

p and generates a control signal Au, which is then fed back to the AVR and exciter.

8 LW h', 1.5495

Page 63: Optimal design of multirate systems

4s

This mode1 has ben used for conventional stabilizer design and robust analog sta-

bilizer design [14, 131 based on 36. optimization. However, the orders of the robust

st abilizers designed are st ill considembly higher t han t hat of the convent ional s t abi-

lizers due to the inclusion of weighting functions in design ( recd that the controLler

order equais rougkdy the order of the plant plus those for the weighting functions). To

have reasonable orders for stabilizers via Ra optimization, one must select weighting

functions for good performance while keeping their orders to a minimum - a topic to

be discussed next.

4.2.2 Weighting Funetions

For good PSS design, let us note two desirable properties for the power system

under control:

1. By standard practice, the stabilizers should provide damping to oscillations

due to only low-frequency disturbaaces. Typicdy these occur in the frequency

range of approximately 0.2 to 2.5 Hz, see, e-g., [42]. Thus stabilizen are to be

designed to attenuate only low-fiequency disturbances.

2. In the case that the stabilizer has input w and output u, i t is desirable that

both w and u be zero at steady state when the disturbances are step signals.

This reflects the requirements that at steady state, the power angle 6 (integral

of w) be constant and the stabilizer be disconnected in effect. The latter prop

erty means that weU-designed stabilizers are brought to action o d y during the

transient process.

From the first property, good designs should be focused on at tenuating low-fiequency

components in the disturbance inputs Tm and K.,. For this purpose, two input

weighting functions (fictitious premters) Wt for Tm and W, for Kef are introduced

Page 64: Optimal design of multirate systems

49

for design. Wt and W, are chosen to be lowpass to cover the fiequency band of inter-

est: to keep the order increase to a minimum, they are taken to be first-order lowpass

filters with the same time constants:

(This way the order of the genealized plant including the two weighting functions is

increased only by one due to the fact that both Wt and Wv have the same denomina-

tor.) Note that the target frequency range can be determined by ody one parameter

T. The gains kt and kv reflect relative weighting between attenuated low-fiequenc-

disturbances in Tm and Ker: e-g., if kt is mu& larger than +: more emphasis is

placed on disturbance fiom Tm.

So provide damping to the oscillations, one can use as regulated outputs the power

( p ) : speed (w), power angle (6): or any combination of the t h e . However. the

standard R, technique does not parantee the second propem; that w and u at

steady state converge to zero due to step disturbances: If #(oc) # 0, 8 wilI r m p to

infmity; if the stabilizer is overdesigned so that u(m) # O (it provides overdamping),

&(cc) WU be lower than the expected level which is determined by the parameters of

the power system without the stabilizer. The first case could happen if one used ody

;2 a,nd/or p as the regulated outputs; and the second case could happen if u is not

correctly weighted as a regulated output.

For all desipso we will use three signalç, p, 6, and ul as regulated outputs subject

to appropriate weighting. Then the power system in Figure 4.2 can be recodgured

into an equivalent plant Po with three inputs Tm, K e f : and ul and four outputso the

reeplated outputs p, 6, and u: and the measured output y (y can be w or a two-

dimensional vector consisting of (*. and p); the stabilizer is in feedbadc connection

with Po using y as input and u as output, as is shown in Figure 4.3, where h' is the

stabilizer. Io weight the regulated outputso we introduce a constant weighting c; for

Page 65: Optimal design of multirate systems

Figure 4.3. Reconfigured power system with PSS and weighting functions.

d: For a nonzero c ~ , the ?&, designed closed-Ioop system guarantees that due to step

disturbances &(oc) is weu-defined and finite so that w ( m ) must be zero, because 6

is an integral of u;. Similady, to get u(oo) = O due to step disairbances, we use an

integrator as the weighting function for u:

where E is a s m d positive number whose role is to slightly perturb the weighting

function so that Wu becomes stable (this makes 'fl, problem nonsingular). Findy,

we also use constant weighting for p and normalize it to 1: in this case, experience

shows that c~ shodd be much smaller than 1 to avoid overdesign.

To summarize, we have introduced input weighting functions for disturbances Tm

and Kef as shown in Figure 4.3,

wl and to2 being prefiltered disturbance inputs with no spectral limitations. We have

dso introduced output weighting functions for the original regulated outputs:

Page 66: Optimal design of multirate systems

a l

Absorb the weighting functions into Po to get the generalized plant P. the shaded

blodc in Figure 4.3: defmed as follows:

This way we arrive at the standard setup with the following equations

The associated analog design problem can be stated as follows:

Design an LTI, causal stabilizer K to stabilize the generalized plant P and

minimize the 31, n o m of the 3 x 2 closed-loop transfer matrix fiom w to r .

Of course, to achieve good performance one needs to adjust the weighting functions

Ui,? WVf Wu, and CJ; more ~pecificdy~ the following five parameters are tuned in

design:

TT kt, 5, CJ, ku-

Xotice that the number of puameters involved is even l e s than that of a standard

convent ional power svstem st abilizer.

Find- we mention that a state-space model can be obtained for P which is of order

six [order of Po (four) plus the order increase (two) due to the weighting functions].

This state-space model is our basis for subsequent analog and digital designs.

4.3 Design of Robust Digital Stabilizers

Robust stabilizer designs here are based on the generalized plant P in (4.2) which

includes the linearized power system model Po and the weighting functions. Because

Page 67: Optimal design of multirate systems

.j 2

our choices of the linearized model and the weighting functions, we can easily compute

a sixth-order st atespace model for P :

where DZ2 = O for both cases of y by dadation.

Our digital stabilizers designed Iater will be compared against the analog conven-

tiond stabilizer (CPSS) of the form given by the IEEE standard CPSS type PSSlA

shown in Figure 4.4. This CPSS measures the speed signal (u) and output the volt-

Figure 4.4. The IEEE standard CPSS confgguation.

age signal (u) to the -4VR and exciter. In Figure 4.4, u,, and u,, represent the

control Iimit constraint; with our model, u,, = u,in = 0.1 p.u. The parameters

of the CPSS are tuned carefully for good response in 6 (power angle) when a torque

disturbance is applied to the nonlinear power system model under normal operating

conditions. The parameters are given in Table 4.2. Xote there that Ai and .A2 are

both zero, the CPSS is effectively a fourth-order system.

Table

b l T2 0.02 Ts T3 0.1 Ts 0.005 hfs 0.05

4.2. Parameters for the tuned analog CPSS.

Page 68: Optimal design of multirate systems

Throughout pon-er system stabilizers are tested for three cases:

a Tm J.: a torque disturbance applied with 0.1 p.u. step decrease and removal:

a t: a voltage disturbance with 0.05 pu. step increase and removal:

a Ground fault: three-phase-to-ground short circuit accident and redosure:

and under three load conditions:

a Yormd load conditions: P = 0.9 p-u., P' = 0.85 (lag) (slightiy Lower than

PI = 0.95 for the operating conditions);

a Light load conditions: P = 0.2 p.u., Pf = 0.85 (lag);

a Lead load conditions: P = 0.5 p-u., PI = 0.9 (lead).

Xote that for lag Pf , we have 6 > 0; and for lead Pj: we have b < O. For ground fault

accident, only normal load conditions me tested. In all cases, the power angle b is

monitored. For example, using the malog CPSS, we simulate and show the responses

of b in the seven cases in Figures 4.5411 (dotted cunres) as detailed in Table 4.3.

II Nomal load 1 Light load 1 Lead load 1 I - 1

Tm L II Figure 4.3 1 Figure 4.6 1 Figure 4.7 ] I - , - L - 1 Figure 4.8 1 Figure 4.9 1 Figure 4.10 1

Table 4.3. Nonlinear simulation tests of the CPSS and the figures.

In what foIlows we will design four robust digital stabilizers via 'fl, optimization;

our purpose is to fmd stabilizers whkh outperform the CPSS: have low orders, and

require low sampling rates, thereby providing an advantage in implementation.

Page 69: Optimal design of multirate systems

4.3.1 Andog Design and Discretization

We will b t design a robust analog stabilizer and then obtai. the digital stabilizer

bq. discretizing the analog one. The stabilizer is SISO, meMuing only the speed

signal.

In view of Figure 4.3: the analog 36. design problem can be stated as follows:

Given the generaked P? design an LTI stabilizer K to provide closed-loop stabiliw

and minimize the 3L, nom of the trander môtrix from w to z, denoted f . ( s ) . Recall

that the 'fl, norm of Fm is the supremd maximum singular d u e on the imaginary

Such an optimization problem can be solved by y-iteration; each iteration involves

solving two Riccati equations. The standard solution [33] has been implemented in

MATLAB as function hinfsyn in the p-Analysis and Synthesis Toolbox [f]. In pneral.

tirne (sec)

Figure 4.5. l'orque disturbance with normal load: CPSS (dot), analog stabilizer (solid) reduced-order analog stabilizer (dash-dot )

Page 70: Optimal design of multirate systems

0.14' 1 L 1

0 1 2 3 4 5 6 7 0 9 1 0 time (sec)

Figure 4.6. Torque disturbance with light load: CPSS (dot), analog stabilizer (solid): reduced-order analog stabilizer (dash-dot)

Figure 4.7. Torque disturbance with lead load: CPSS (dot), analog stabilizer (solid), reduced-order analog stabilizer (dash-dot)

Page 71: Optimal design of multirate systems

. - - - 0.48

, Q , 1

0 1 2 3 4 5 6 7 8 9 t O time (sec)

Figure 4.8. Voltage disturbance with normal load: CPSS (dot), analog stabilizer (solid), reduced-order analog stabilizer (dash-dot )

Figure 4.9. Voltage disturbance with light Ioad: CPSS (dot), analog stabilizer (solid), reduced-or der analog s t abilizer (das h-dot )

Page 72: Optimal design of multirate systems

o'*O 2 6 9 1 1 0 time (sec)

Figure 4.10. Voltage disturbance with lead load: CPSS (dot), analog stabilizer (solid), reduced-order analog s tabilizer ( dash-dot )

Figure 4.11. Ground fault test with normal bad: CPSS (dot)? analog stabilizer (solid), reduced-order analog stabilizer (dash-dot ) , discretized analog stabilizer with sampling period 20 ms (dash)

Page 73: Optimal design of multirate systems

one inputs a state-space model for P and gets a state-space model for K fiom the

optimization: the order of K equals that of P. Since the order of our P was kept as

low as six, the optimal stabilizer designed will be also of order six, which is comparable

with the standard CPSS given in Figure 4.4.

Before ninning the program hinfsyn, a little massage is necessary to regularize the

problem. The problem with the original data for P is singular because DI2 and D21

in (1.3) are both zero matrices and do not have MI r d . This can be fked by slight-

perturbing the two zero matrices to have full r d . For example, we can introduce a

fictitious input wj and an output zj weighted by a small number c = 0.1. Thus P is

The problem then becomes nonsingular and the design cm be done based on Prim.

The foms of the weighting functions have been chosen in Section 4.2.2: the param-

eters involved are selected by trial and error (and experience) during the design and

simulation process to achieve good dynamic performance for the closed-loop power

system. In thk analog design, we use

These weighting hctions will help achieve the two desirable properties discussed at

the beginning of Section 4.2.2 for the controIled power system.

The optimal analog stabilizer K can be designed now by 36. optimization using

the M.4TLAB function hinfqn:

Page 74: Optimal design of multirate systems

59

Xote that similar to the washout factor in the CPSS in Figure 4.4, the numerator of

thÏs robust ~ ( s ) contains a derivative factor whose function is to force the output

of K to converge to zero when there axe step disturbances. This justifies our use

of the integrator-Ue weighting function Wu. Nonlinear simulations are conducted to

e d u a t e this analog controller: The seven test results are also contained in Figures 4.5-

4.11 as the solid m e s . W e conchde that this robust analog stabil'zer is considerably

superior to the conventional st abilizer.

For an order cornparison, note that 'fl, based power system stabilizers designed

in the literature have higher orders: in [71, 141 the robust stabilizers both have order

ten, which have alreadq- undergone a model reduction process. Our analog stabilizer

is of order six and we can also consider pedonn model reduction. Using MATLAB

fûnctions hankmr, we compute the fourth-order. optimal Hankel-nom approximation

to the sixth-order stabilizer K , based on a balanced realization of k ( s ) (obtained via

MATLAB function sysba2):

This ha. the sarne order as the CPSS tested. hlthough the derivative factor has

disappeored in &(s), it can be vedied that it still satisfies the second propertt-. Le.:

once connected to the power system, the control input u is still zero at steadÿ state for

step disturbances. The simulation tests are shown in Figures 4.5-4.11 in dot-dashed

cuves: In general, the reduced-order Kr suffers slight peâonnance degradation corn-

pared with the original K; but the test results are still much better than those of the

CPSS. (From now on: we s h d skip model reduction and content with the si-xth-order

s t abilizers generated from R, op t imization.)

Our purpose is to implement stabilizers via microprocessors. Thus we discretize

the andog robust stabilizer using bilinear transformation (Tustin's approximation)

to get a sixth-order digital stabilizer. the sampling period used being 10 ms. The

Page 75: Optimal design of multirate systems

60

closed-loop responses with this digital stabilizer for the standard tests axe almost

indistinguishable fiom those with the anôlog r;'(s), and are hence omitted. However,

if larger sampling periods are used for the discretization, the performance of the digital

stabilizer is noticeably worse than the d o g performance: For fistration, take the

samphg period to be 20 ?O; the ground fadt test for the resultant digital stabilizer

is showvn in Figure 4.11 (dashed curve), from which we see a much higher overshoot

for the fault recoverq- than those using the andog stabilizer (solid curve) and even the

CPSS (dotted curve). This shows the disadvantage of discretizing an analog design

- it requires relatively fast sampling rates. If slower sampling is desired without

compromising performance, design should be done directly based on discrete-time

4.3.2 Discrete-The Design of SIS0 Stabilizers

The setup in this case is shown in Figure 4.12, where P is again the generaiized

Figure 4.12. The single-rate digit al control set up

power system with a state-space mode1 in (4.3) but the stabilizer is a sarnpled-data

one: The measured signal y (t) is first sampled by the ideal periodic sampler Sh with

sampling period h to get the sampled signal yd(k); then yd(k) is processed by the

Page 76: Optimal design of multirate systems

61

digital stabilizer Ka to generate the control sequence ud(k), which is then transformed

to a piecewise continuous signal u(t) by the zero-order hold Hh.

In this subsection, we take y to be the speed signal and so the stabilizer is SISO.

The setup invdved is a single-rate control system with only one rate ( I / h ) . For an

implementation advantage, we take a sampling rate which is a factor of four slower

than the one used in the analog design, Le., h = 40 ms. We remark here that the

sampling rate cannot be made arbitrady slow: If low-fiequency oscillations of up to

j Hz are to be attenuated by the digital stabilizer, roughly speaking, the samphg

rate should be at least twice as fast as the maximum oscillation fiequene- (Shannon's

sampling theorem) and so the sampling period has an upper limit: h < 100 ms.

To perform a discrete-time design, the first step is to discretize the generalized

systern P to get Pd [l'il. The discrete-time system Pd and the digital stabilizer fi

f o m a discrete-time feedback system shown in Figure 4.13- -4 state-space model for

Figure 4.13. The discrete-tirne setup

Pd can be cdcdated fiom that for P in (4.3) based on sample and hold equivalence:

Here,

Page 77: Optimal design of multirate systems

Then the associated disaete-time 3L, design problem is as foI.lows:

Given Pd, design an LTI digital stabilizer Ka to provide closed-bop stability

and mirrimize the 9& nom of the cIosed-loop transfer matrix T~~~ : ZQ rt zd:

an optimization problem c m be solved purely in discrete time aaalogously via

two dgebraic Riccati equations [37, 691. However, we will use the bilinear transfor-

mation to convert it to an equivalent contipuous-time problem. see, e.g.. [17] and

Section 3.2. and then solve by applyuig the M-4TL-4B funaion hinfsyn.

Xote that the weighting functions we used for digital design are still given in con-

tinuous time. However, the parameters are not necessarily the same as in Section 4.3.1

in the d o g case and are slightly re-tuned in the design and simulation process for

good performance. W e have

The corresponding digital robust stabilizer is computed to be of order six:

Xotiote that the numerator of &(z) again contains the derivative factor ( r - 1): which

guaantees that the final value of ua(k) due to step disturbances is zero. But this

factor can be eventually côncelled out by the factor (z - 0.9999) in the denominator.

yielding a f%h-order stabilizer. The test resdts are given in Figures 4.14420 (solid

cuves) . To compare, the test resdts for the CPSS are redrawn as dotted curves. It

is dear that this digital stabilizer is comparable in performance to the analog robust

stabilizer, outperforms the CPSS a great ded, yet requires a sampling period of ody

40 ms.

Page 78: Optimal design of multirate systems

tifne (sec)

Figure 4.14. Torque disturbance with normal load: CPSS (dot)? digital SISO stabi- lizer (solid), digital M I S 0 single-rate stabilizer (dash-dot)

Figure 4.15. Torque disturbance with light load: CPSS (dot), digital SISO stabilizer (solid): digital MIS0 single-rate stabilizer (dash-dot)

Page 79: Optimal design of multirate systems

time (sec)

Figure 4.16. Torque disturbance with lead load: CPSS (dot), digital SISO stabilizer (solid) digital MIS0 singlerate stabilizer (dash-dot )

Figure 1.17. Voltage disturbance with normal load: CPSS (dot), digital SISO stabi- lizer (solid), digital MIS0 singlerate stabilizer (dash-dot)

Page 80: Optimal design of multirate systems

Figure 4.18. Voltage disturbance with light load: CPSS (dot), digital SISO stabilizer (solid), digital MIS0 single-rate stabilizer (dash-dot )

o n ; L I 1 2 3 4 5 6 7 8 9 1 0

tirne (sec)

Figure 4.19. Voltage disturbance with lead load: CPSS (dot), digital SISO stabilizer (solid), digital MlSO singlerate stabilizer (dash-dot)

Page 81: Optimal design of multirate systems

0.45~ 1 O t 2 3 4 5 6 7 8 9 1 0

izrne (sec)

Figure 4.20. Ground fault test with normal load: CPSS (dot), digital SISO stabilizer (solid). digital MIS0 singlerate stabilizer (dash-dot)

4.3.3 Discrete-Tme Design of MIS0 Stabilizers

The MIS0 control setup is similar to the SISO control setup in Figure 4.12 except

that the signal y now is a twdiruensional vector,

y=[ ; ]

m e a s h g both the speed and power. The digital stabilizer has two inputs and one

output, ail sampled at the same rate. With this additional power signal, we hope to

further reduce the sampling rate with Little expense in performance.

The disaete-the design process is almost the same as in Subsection 4.3.2. But

in this case, it is helpfid to bring in a weighting function at the measured output

s; (Figure 4.12). For 31- design, the weighting function we use for this measured

output is a constant, denoted by cd. Let us take the sampling period h = 80 ms:

which is twice as much a . that used for the SIS0 stabilizer in Subsection 4.3.2. For

Page 82: Optimal design of multirate systems

the following design, the weighting b c t i o n s are

The digital stabilizer designed again via discrete-time R, optimization has the 1 x 2

transfer rnatrk

with

Again: the factor (z - 1) in the numerators can be cancelIed out by the factor ( z -

0.9938) in the denominatoc the stabilizer is actudy mh-order. The performance

test results are given in Figures 4.14-4.20 (dot-dashed curves). As can be seen, this

'rlISO digital stabilizer still outperforms the analog CPSS a great deal, although the

sampling rate is quite slow.

4.3.4 Multirate Design of MISO Stabilizers

The MISO digital stabilizer studied in Section 4.3.3 in the single-rate setting can

admit a multirate sampling scheme. To see why multirate control is appropriate. we

plot the magnitude fiequency responses of the two transfer functions in Figure 4.2

from, e-g., the voltage input Kef, to the measured signds, w and p (both in p-u.),

as are shown in Figure 4.21. From this figure, the transfer function from to p

has a wider band of significant fiequency components than the transfer function from

to w. Hence it is reasonable to sample p fast and w slowl. Let us consider

two sampling schedules. First, we use a sampling period of h = 60 ms for the power

Page 83: Optimal design of multirate systems

Figure 4.21. Magnitude frequency responses from to p (solid) and w (dot); p and a are both in p.u.

signal p and a samphg period of 2h = 120 rns for the speed signal sr: the hold (DIA

conversion) also has a period of 60 ms. Second. we inaease the sampling period at

the speed signal further to 3h = 180 ms in order to see the performance degradation.

The associated design setup is shown in Figure 4.22, where the signals w and u

are as before, and

Figure 4.22. The multirate digital control setup

Page 84: Optimal design of multirate systems

69

(Recd that as in Section 4.3.3, we have weighted m e m e d output w by 6.) Let

us focus on the b t sampling schedule. The multirate sampler S corresponds to two

and the multhte hold 3L reduces to the single-rate one: 'fl = Rh. Let T be the least

common period for all skmpling and hold channels - in this case T = Sh = 120 ms-

Discretizing the s jstem and applq-ing lifting, we amive at as LTI system of Figure 4-23

with a t hree-input , two output controuer:

Figure 4.23. The lifted LTI setup

which corresponds to grouping the inputs and outputs properiy over the period T .

Then it follows that & has a state-space mode1 given by

Xote that D13 = O &ses fiom the causality constraint on the multirate controller.

The 3L, design for this constrained problem is as follows:

Page 85: Optimal design of multirate systems

70

Design an LTI & of the form in (4.4) satisfying the causality constraint Dl, =

O to provide dosed-loop stability and min;mize the ?& n o m of the transfer

rnatrix fiom a to a in Figure 4.23.

The rnultûate stabilizer can be obtained by running the M.4TLAB programs given

in Section 3.4. The following weighting functions are used for our subsequent design:

The robust multirate stabilizer Kd, in the lifted state-space form of (4.4, is computed

Note that the order of this multirate stabilizer is again six and that the zero entry

in the D-matrix indicates causal i . The test results for this multirate stabilizer are

given in Figures 4.24-4.30 (solid m e s ) in cornparison with those for CPSS. From

these plots, we conclude that dthough the sarnpling periods for w and p are as l q e

as 120 and 60 ms? the rnultirate stabilizer is considerably better than the CPSS.

Finally, let us replace the sampling period of 120 ms for the speed signai with 180

rns and keep the remaining rates unchanged. In this case, T = 32 and the controller

& becomes a fout-input, th.e+~utput one: -

Page 86: Optimal design of multirate systems

Its state-space mode1 is

wheie, again, the zero entries &se fiom the causality constra.int. Ushg the following

set of weighting funaions

we can design the multirate stabilizer &, whose test results are shown in Figures 4.24-

4.30 (dot-dashed m e s ) . It is seen that the performance of the mdtirate stabilizer

starts to deteriorate; this is particularly reflected in the ground fadt test of the

stabiliw (Figure 4.30): It is worse than the solid curve? but still no worse than

the dotted curve (CPSS). This suggests that if we use the CPSS performance as the

benchmark, we have almost corne to the limit in using slow-rate digital stabilizea via

Figure 4.24. Torque disturbance with normal load: CPSS (dot), digital MIS0 multi- rate stabilizer with sampling period 120 ms for w (solid), and digital MIS0 multirate stabilizer with sampling period 180 ms for w (dash-dot)

Page 87: Optimal design of multirate systems

Figure 4.25. Torque disturbance with light load: CPSS (dot), digital MISO multirate stabilizer with samphg period 120 ms for w (solid), and digital MIS0 multirate stabilizer with sampling perïod 180 ms for w (dasidot)

tim (sec)

Figure 4.26. Torque disturbance with lead load: CPSS (dot), digital MIS0 multirate stabilizer with sampling period 120 ms for w (solid), and digital MISO multirate stabilizer with sampling period 180 ms for w (dash-dot)

Page 88: Optimal design of multirate systems

time (sec)

Figure 4.27. Voltage disturbance with normal load: CPSS (dot), digital MISO multi- rate stabilizer with sampling period 120 rns for w (solid), and digital MIS0 mdtirate stabilizer with sampling period 180 ms for w (dash-dot)

Figure 4.28. Voltage disturbance with light load: CPSS (dot), digital MISO multirate stabilizer with sampling period 120 ms for w (solidj, and digitd MISO multirate stabilizer with sampling period 180 ms for w (dash-dot)

Page 89: Optimal design of multirate systems

Figure 1.29. Voltage disturbance with lead load: CPSS (dot), digital MISO multirate stabiEzer with sarnpling period 120 ms for w (solid), and digital MISO multirate stabilizer with sampling period 180 ms for w (dash-dot )

0.450 , . , 1 1 2 3 4 5 6 7 8 9 1 0

time (sec)

Figure 4.30. Ground fault test with normal load: CPSS (dot), digital MISO multirate stabilizer with sampling period 120 ms for w (solid). and digital MISO multirate stabilizer with sampling period 180 ms for w (dash-dot)

Page 90: Optimal design of multirate systems

the optimization.

4.4 Conclusions

In this chapter, we have studied stabilizer design for a single-machine, infinite-

bus power system using the robust ?& optimization. Several designs have been

accomplished includïng asdog and digital, single-rate and multiate. .AH stabiIizers

designed have low complexity - they are LTL and of low orders - and outperfom

the analog, conventional power system stabilizer a great deal on a series of nonlinear

d ~ o m i c tests; the digital stabilizers designed also enjoy the additional property t hat

they use considerably iow samphg rates in implementation.

Sorne other concluding remarks are:

In 'H, control design, it is important to use appropriate weighting functions.

Weighting functions are chosen to reflect the operational requKements of the

system and are sometimes based on experience. But it is desirable to use simple

weighting functions, for they introduce minimum order increase in controllers.

Power systems are generdy described by high order nonlinear equations. In

some previous work (se, e.g., [71]). high order linearized models are used to

approximate the nonlinear equations. In our study, we showed that the low

order linearized model, which has been used in design of the conventional power

system stabilizer, is still usehl in designing high performance stabilizers.

a Mdtirate control systems offer a cost advantage in implementation. Design of

multirate control systems involves handling a new causality constraint and it is

feasible to accompli& %-optimal designs based on recent advances in the the-

ory [74, 191. To implement this theory in MATLAB, the bilinear transformation

approach proposed in Chapter 3 offers a computationd admntage.

Page 91: Optimal design of multirate systems

76

a For reiatively low sampling rates, digital design based on discretized models

works better than discretizing analog designs. Poor intersample rippIes can be

avoided by properly selecting weighting functions in the design process. To

further reduce samphg rate, the direct sampled-data design methods (se,

e.g., [l?]) would be a good choice, because they take into account intersample

information.

Page 92: Optimal design of multirate systems

-- l I

CHAPTER 5

Na DESIGN OF HYBRID MULTLRATE FILTER BANKS

The focus of the previous chapters was placed on control systems. In th% chapter.

we tuni to the study of a signal processing system, in particular? a hybrïd multirate

filter bank. Our purpose is to design the multirate filter bank using 3L, optimization.

thus providinp a new application of RI, control theory'.

5.1 Motivation and Introduction

The subject of this chapter is the hybrid multirate filter bank shown in Figure 5.LI

where Go and G1 are LTI andog analysis fikers, SZh the slow sampler with period

Figure 5.1. The hybrid multirate filter b d .

2h, E the ezpander by the factor of 2, and Fo and FI LTI digital synthesis filters-

The expander E is a new component whose function is to inaease sampling rate by

inserting zeros between sarnples:

3ote that we have denoted expanders by E instead of the traditional symbol t 2

because the former is more appropriate for equations.

lThe r d t s in this chapter have been published in [67].

Page 93: Optimal design of multirate systems

78

Such systems are potentidy usefil in digital trammÏssion of andog audio signds.

Suppose the input analog signal r(t) is stnctij- bandlimited. To f a i t M y represent

x ( t ) ? we need to sample at least as fast as its Nyquist fiequency w ~ ; this leads to

the constraint on the sampling period: h 5 Zir/wN. Using the two-channel filter

bank. one codd design Go to be lowpass and G1 highpass, splitting the band into

two subbaads; this way one codd effectively reduce the samphg fiequency in each

channel by a factor of two and at the same time achieve overall bit-rate reduction in

transmission by exploiting fiequency characteristics in coding [72]. For such a system

to work, the two synthesis filters must be designed so that the output + approximates

the desired output 4 obtained by fast sampling the input: t$ = Shx. By standard

practice, sorne time deiay wil l be tolerable and so we should compare tb(k) with

d(k - m): where rn is the timedelay integer.

Let U be the unit time-delay system in discrete t h e with trollder function z-'.

The error system is P S h - Ko as is depicted kom z(t) to the error signal, r ( t ) =

&(k - m) - tb(k), in Figure 5.2. However, this error system presents a difficulty for our

Figure 5.2. The pre-filtered error system.

design (see below) because of the bandlimited assumption of x ( t ) . Aence we introduce

Page 94: Optimal design of multirate systems

a pre-filter G which is stable, causal, and lowpass with passband compatible with the

bandwidth of x ( t ) . In t 6 i s way we get the ptefiltered error system in Figure 5.2 hom

v ( t ) to ~ ( k ) , namely, (UmSh -K)G. From now on, this WU be called the error system.

We remark that G shodd be thought of as fictitious: introduced ody for design,

mhose role is to capture the fiequency components of x ( t ) of interest; then the input

to G, v ( t ) , is no longer bandlimited. In this sense, G can be regarded as a fiequency-

selective weighting filter. In practice, one could &O think of G as the anti-aliasing

filter before A/D conversion (with period h) .

The performance criterion we w is the hybrid worst-case energy gain of the error

system, namely,

where the supremum is taken over d nonzero v in L2- Recall that for a fixed finite

energy signal v(t) , the energy gain of the error system is ( I E [($/ llv 11;. Hence J relates

to the worst-case energy gain: If J is smd, then the energy gain of the error systern

is unifonnly s m d for all possible inputs.

It is worth mentioning that without the loqass pre-filter G, J becomes infinite

(the error system is unbounded from L2 to 12) no matter what synthesis filters are

used. This is due to the presence of severe aliasing; see, e.g., [E] and [l ï? Section 9.31.

It can be verified that we get perfect reconstruction ( J = O) with no time delay

(m = O) if ideal flters are used. However, if stable and causal flters are to be used.

one often has to tolerate some time delay to get 3 smd; see the design example in

Section 5.3.

In this study, we take the viewpoint that the andysis filters have been designed

already, and the synthesis filters aze now to be designed to minimize J. The precise

design problem is as follows:

Given causal. stable, finite-dimensional analog fiters G: Go, Gi with G lowpass

Page 95: Optimal design of multirate systems

80

and given the fast sampling period h and a tolerable the-delay integer m.

design causal: stable, digital synthesis filters Fo and Fi to minimize J.

The optimal synthesis fiIteis designed will be IIR in geneal. The optimal perf'ormance

achieved is Jqt(rn), which caa be shown to be a nonincreasing tunction of the time

deIay m.

The optimization problem involved can be written as follows:

11412 JWt(m) = inf J = inf sup -. Fo FI 625 4 0 1127112

This is a minimaz design problem, which cannot be solved directly by known met hods.

because of the hybrid and multirate nature of the error system. The main contribution

of this chapter is therefore the development of a procedure to convert this problem

into an equivalent %-optimal model-matching problem involving only discrete-the

LTI systems, which is then solvable.

ho ther possible application of the problem studied is to achieve fast A/D conver-

sion by slow A/D conversion in the twcxhannel filter bank in Figure 5.1. Properly

desiped filter banks can increase the conversion rate by a factor of two. An>- integer-

factor increase is possible if one uses multiple channels - a motivation to e--end the

results of this paper to the multiple-charnel case. For such an application, practi-

cal issues such as filter complexity and jitter effect need to be addressed; these are

not discussed in this study. Some practicd issues were studied in [53] for fast A/D

conversion using digital filter bank th-

Digit al filter banlrs have been studied extensively in signal processing; Vaidyanathanos

book ['i-] provides a detailed historical survey of the literature up to 1992. Optimal

design using l2 induced n o m was advocated and studied by Shenoy et al. [63' 641

in digital multirate filter design and by Chen and Francis [18] in digital filter bank

design for perfect reconstruction. Though possible applications of hybrid filter banks

Page 96: Optimal design of multirate systems

81

were mentioned in [72], to our best knowledge, no work was reported on design of such

hybrid filter banks. The work here seems to be the first attempt in this direction.

Hybrid filter b d s may be more appealing than digital filter b d s in transmitting

analog signals. First. digital filter b a h require A/D conversion (prior to processing)

which is twice higher in speed than hybrid filter b a h . Second, practical speech and

video signals are mdog and hybrid filter banks treat them directly. Third, because no

analog signds are strictly bandlimïted in practice, aliasing caused by .4/D conversion

always exists; this aliasing is out of consideration in digitd filter banks but wiU be

under control in hybrid filter b d s since analog filters are incorporated in the design.

5.2 Conversion to a Problem of 3C, Optirnization

The design problem stated in the precedhg section is hybrîd and multirate. to

which no current methods are applicable directly. In this section we s h d reduce it

fUst to a discrete-the rnultirôte problem and then M e r to a discrete-time single-

rate problem of ?lm optimization, which can be solved readily by existing software.

The techniques used relate to both sarnpled-data control and polyphase decomposition

in signal processing.

5.2.1 Conversion ta a Discrete-The Problem

The first step is to convert the hybrid problem into an equivalent discrete-tirne

one. Let the error system frorn v to E in Figure 5.2 be T, namely,

This involves samplers Sh and SZh with diRetent periods. Let S2 (usudy represented

by the symboIJ 2) be the dismete-time dounisampler defined via

Page 97: Optimal design of multirate systems

It can be i r d e d that is the cascade of Sh and S2:

Substituting this into (5.2). we have

Csing (5.1) and (5.3), it foUows that the cascade connection ESÎ (downsample. then

upsarnple) is &en by

that is, it zeros the input at odd times. Thus ES2 is a 2-periodic system (but S2E = 1,

the identity system). The overall system T is 2h-periodic in continuous time. Define

the continuous- and discretetime systems P, and Ld, respectivelqi as follows:

Thus P, is a l-input, %output system, Ld is Sinput, 1-output, and the error system

the cascade of three systems: the continuous-time system P,, the %input, 3-output

sampler, also denoted by Sh . and the discrete-time system Ld. In our case, P, is stable

and causal with a strict- proper transfer matrix. Thus ShPc is a bounded operator

from L* to e2 [El.

Now we wish to relate the performance meanue, J = llTll, to the nom of a purely

discrete-time system. For this we need two lemmas. The first lemma is from operator

theory [9]:

Page 98: Optimal design of multirate systems

S 3

Lemma 5.1 Suppose X and Y a n iwo Hilbert spaces and T is a linear bounded

operator from X to y. Then

whete Tn is the adjoint operator mapping Y to X.

The second is from sampled-data control [17: Section 10.51 :

Lemma 5.2 For a stable, causal continuous-time system Pc unth strictly pn

fer rnatrtx

theie e=Çists a stable, causal discrete-time system Pd satisfying

and

where Ad = eu and Bd iS any mat* satisfying

h Bd B; = et" B ~'e'"' dt.

Ipplying the tmTo lemmas to the error system, we get

oper tran

The latter nom is &-induced and is on a purely discrete-time system Ld Pd. Thus we

have reduced the hybrid performance measure to a discrete-time measure.

Page 99: Optimal design of multirate systems

Suppose P, has a realization

1

GL (s)

Since P,(s) is 3 x 1, the output mat& is partitioned into three rom. From Lemma 5.2,

the state-space model for Pd is given by

This Bd given by (5.4): is not unique: for example, we can take Bd to be a ~ y Choleshc

factor or even the square root of the right-hand mat&. .Also, the r a d of Bd is in

general greater than one, meaning that although P, is singleinput. Pd is in general

multi-input In fact, if (A, B) is controllable, the matrix integrai in (-5.4) is nonsingular

[Hl and hence Bd must be a square matrix.

Partition Pd as below

Then the equivalent error system becomes

which corresponds to the discrete-time system shown in Figure 5.3, and the equivalent

design problem is:

Given discrete-time systems Ho, Ki, IV, ail stable and causal: and given the

integer rn, design stable, causal, discrete-time Fo and FI to minimize 11 Tdll : the

&induced n o m of Td.

Page 100: Optimal design of multirate systems

' ----. -.*{-q.. *-. . - .{Y[--- ----. {y- - - . {FI. .-- - . --.'

Figure 5.3. The equivalent discrete-the error system.

This is a discrete-time multirate design problem. It looks like the problem one

would get from design of muhirate filter banks for perfect reconstniction [la]; but

in fact it is different: First, the system to be matched is PN, not a pure time

delay; second, the three systems Ho, Hl, Ai are d multi-input. The second fact

leads to a limitation on achievable performance levels in design; see Section 5.4 for a

rigorous discussion. Next, we s h d reduce the multirate problem further to a single-

rate problem of R, optimization.

5.2.2 Conversion to an % Problem

From Figure 5.3, it follows that

At tthis point we need the polyphase matrices for the four LTI systems Hot Hl: Fo,

and FI [72]: For the analysis bank,

Page 101: Optimal design of multirate systems

and for the spthesis b d ,

The synthesk polyphase mat& R is as usud 2 x 2. However, note that since Ho

and Hi have multiple inputs in general, i ) ( z2 ) is a generalization. -4ssume Ho(-.) and ( z ) are both 1 x n; then the two identity matrices in (5.7) are both n x n and ~ ( i ' )

is '5 x 2n.

With the two polyphase matrices, we hope to convert design of Fô and t to that

of R. Substituting equations (5.7) and (5.8) into (5.6) and using the noble identities

Here, O-: again, corresponds to the time delay z-' and Q and R are the LTI systems

with tiaasfer mat& &(z ) and ~ ( r ) respectively.

To proceed further, we need the following lemma which can be proven readily [BI.

Lemma 5.3 The operator and its right inverse

are both nom-preseming on &, where V-l, the inverse of U* on &: is the unit time

advance.

Thus the nom of Td remains unchanged if we pre- and post-multiply Td by

respectivek This leads to

Page 102: Optimal design of multirate systems

This system W c m be shown to be LTI. In this way we arrive at the final equivalent

s-stem, shown in Figure 5.4 back in the fiequency domain. The sizes of the three

matrices &(z). ~ ( r ) , and ~ ( z ) are 2 x 273, 2 x 2, and 2 x 272, respectivel. (The

integer n is the dimension of the input in Figure 5.3.) Hence, the P2-induced nom of

Figure 3.4. Final equivalent system.

W - RQ equals the 'fl, n o m of its transfer matrix, 11 w - RQ 11,. Uë condude that the original design problem, minimizing the Cî-to-Q-induced

n o m of the hybrid, multirate system v E in Figure 5.2, is equivalent to the following

Ha optimization problem:

Given stable and causal transfer matrices ~ ( z ) and ~ ( t ) , design stable and

causal R ( t ) to minimize 11 w - 8~11,-

Equidently, we can write:

Page 103: Optimal design of multirate systems

The latter optimization is a standard ?&,-optimal model-matching problem.

Uë s h d condude this subsection by denving explicit expressions for Q(z) and

~ ( z ) . The key lemma is the following:

Lemma 5.4 Let F 6e a causal, LTI discrete-tirne system m3h a transfer matriz

and Po and its polypbase components:

Then Fo(z) and &(z) are the tramfer matrices of the causal, LTI systems Sz F E and

S2L7-' F E respectively and

Proof The state-space formulas can be easily obtained by noting the power series

expansion of &). To see that S2FE has a transfer matrk Fo(r), we need to use

noble identities and Lemma 5.3. Wnte

S, F (z) E = s2 (F, (zZ) + z-' Pl (z2)) E

= S~E&(,(Z) +s2z-'~Fi(z)

- î, - A O*

In the second equality, we have applied noble identities, and in the last one, we have

applied Lemma 5.3. The daim then follows. Similady, we can prove that S&-'F E

has t rader matrix &). O

R e d from (5.5) the state-space realization of Ho and Hi are given by

Page 104: Optimal design of multirate systems

Based on this, it is immediate from Lemma 5.4 that &) takes the fomi

Dependhg on whether the the-delay integer rn is even or odd, the mat& W ( Z )

takes on different forms. From (5.5) the state-space realization of N:

.(il = [Y] C O *

Let us first assume rn is even, w, m = 2d for some d 3 O. Then fiom (5.10)

hoking the identity = VdS2 [72], the (1, 1) and (2,2) blocks equd to Lrd(S&-1 .\-E):

this has a transfer matrix

by Lemma 5 . 4 Similady, the (2 , l ) block L 1

is Cld(LINE), which has a transfer matria

and the (1,2) block has a t r d e r matrix

Based on the four transfer matrices, we have

If m is odd, n = 2d + 1 (d 2 O ) ? the expression for &(z) can be derived in the

Page 105: Optimal design of multirate systems

5.2.3 A Design Procedure

Much of the computation involved is based on statespace data. At this point, let

us summarize the steps into a design procedure:

Input: fast period h, deIay integer m, analysis Hters Go(s) and G&). and

pre-filter G ( s ) ;

Output: optimal performance J&(rn) and synthesis f i tes Fo(z) and Fi(;).

Step 1 Compute a state-space realization for the 3 x 1 matrix:

Step 2 Compute the matrices 4 = eu and Bd satis%ng

h &Bi = e t A ~ ~ Y ' ' d t ,

via. e-g., Choleslcy factorkat ion. The matrix integral involved can be computed

using matrix exponential htnctions [73].

Step 3 Compute the tR.o matrices Q(Z) and ~ ( 3 ) :

Step 4 Solve the following %-optimal model-matching problem

The optimal ~ ( r ) is the 2 x 2 polyphase matnx for the spthesis flters.

Page 106: Optimal design of multirate systems

Step 5 Obtain & and via

It should be noted that here we do not need to haadle the causality constra.int due

to the particular structure of the filter bank. Thus we can carry out the standard

?& optimization procedure in Step 4 and this can be solved by the existing MAT-

LAB function hinfsyn, for which one needs to convert the discrete-time &, problem

equivaZent1y to a continuous-time one using bilinear transformation.

5.3 Example

In this section we study in detail an example to bring forward some points that we

feel important, e-g., performmce limitation, tradeoff between the time-delay integer

(m) and reconstruction performance, and model reduction by FIR tnuication to the

optimal IIR sqnthesis filters. It is emphasized that the example is chosen to illustrate

the new design procedure in the preceding section; the fdters are not necessarily meant

for some realistic application.

For simplicity, we assume that the hybrid filter bank in Figure 5.1 is used to

process analog signals r (t ) with fiequency spectrum mostly limited to Iw 1 5 1. Hence the corresponding Xyquist sampling period is a seconds; this is taken to be the fast

sampling period h. The design method proposed in this paper allows us to design the

andog ana&& filters &st based on considerations in coding and transmission and

then design the digital synthesis flters for best reconstmction.

In this example, the andysis filter Go should be lowpass with passband Iwl 5 0.5.

t V e t h Go to be a third-order elliptic filter with trader function

The magnitude Bode plot of Go is showm in Figure 5.5 (dotted cuve). The other

Page 107: Optimal design of multirate systems

Figure 3.5. IGI (solid), jGol (dot), and lGl 1 (dash) in dB versus W.

analysis filter Gi should be bandpass with passband 0.5 5 1 ~ 1 5 1. This is designed

via the Y.&TLAB function Ip26p ushg the prototype (lowpass) third-order elliptic

So we get a si-xth-order bandpass Gl with the transfer fimction

whose magnitude Bode plot is given again in Figure 5.5 (dashed line) .

The role of the fictitious pre-ater G is to retain the desired fiequency components

of input signals and so G is lowpass with passband Iwl 4 1. This is realized by

cascading a second-order Chebyshev type1 filter and a fourth-order elliptic filter:

The Bode plot of G is also shown in Figure 5.5 (solid curve).

With these analog filters, we can proceed to design optimal synthesis filters given

the time-delay integer m. The procedure in the preceding section is implemented in

Page 108: Optimal design of multirate systems

93

MATLAB. To carry out Step 4? we tmnsform the discrete-time 3L, problem into a

continuous-time 3L, problem via bilinear transformation and then apply the function

hinfsyn. The optimal performance Jm is in general a nonincreasing function of m. In

this Jqt versus rn is given in Figure 5.6, where Jwt drops relativel~ apidy

Figure 3.6. Jet versus m.

for 2 $ m 5 8 but stays at about the same level for m 2 10. This illustrates the per-

formance limitation: B y increasing m, one cannot get arbitrarily good reconstruction.

More discussion on this will be given in the next section.

The optimal synthesis Hters computed are in general IIR with orders that increase

linearly with the integer m and with the orders of G: Go, and Gi. In this example, for

s m d JWt we t&e rn = 15; the correspondhg Jopt is 0.0633 and the optimal synthesis

filtea Fo and FI are computed to be IIR of order 89. The magnitude Bode plots of Fo

and FI are given in Figure 5.7, fkom which we see t hat Fo is lowpass and FI highpass.

To test the two optimal s ~ t h e s i s filters and also to get some sense of what the

performance ( Jmt = 0.0633) means. we apply two analog signals at the input x( t ) of

Page 109: Optimal design of multirate systems

Figure 5.7. 1 F* 1 (solid) and 1 fil 1 (dash) in dB versus w/a.

Figure -5.1 for the desi@ hybrid bank, simulate the output $(le), and compare with

the desired output ~ ( k - 15) = x[(k - 15)hI. (Recall that the pre-Mter is introduced

O* for design; for testing one uses Figure 5.1.) The first input is bandlimited.

consisting of high- and low-fkequency components (chosen arbitrarily):

The desired output 4(k - 13) and the steady-state reconstruction error [e(k - 15) -

&(k)] are shown in Figure 5.8. It is seen that the maximum error (0.035) in mae~tude

is quite s m d relative to the size of the desired output. The second test input is a

continuous-the unit step, x ( t ) = l(t). The dynamic response of the hybrid f3ter

bank is plotted in Figure 5.9. Although this input is not bandlimited to Iwl 5 the

output of the filter bank tracks the desired output, a discrete-time unit step delayed

by 15 samples, fairly wek The steady-state error is 0.0110. Note that the system is

not particul& designed for step-tracking.

To design high-quality hybrid filter banks, one would tend to use more sophisti-

Page 110: Optimal design of multirate systems

-21 1 1 I I

O 20 40 60 80 1 O0

Figure -5-8. Steady-state error (solid) and desired output (dash) versus k.

cated analog filters and, perhaps, tolerate relat ively large time delay S. These neces-

sarily result in high-order synthesis filters. In the example, if rn = 15, the method

generates optimal Fo and FI which are IIR and of order 89. However? the proposed

method is still useful in that starting from here one could approximate the optimal Fo

and Fi by FIR ones via tmcation or by lower-order IIR ones using mode1 reduction.

Yow let us truncate the designed optimal Fo and Fi to get FZR approximations. The

impulse responses of Fo and Fi are IIR and are shown in Figure 5.10 for the initial

40 samples. W e see that both responses settle to zero rough-y in 30 samples. Thus

we tnincate the impulse responses to get FIR filters of length 30. The truncated

synthesis filters PoFrR and f iFIR achieve a performance JrIR = 0.0633, accurate to

three significant digits (compare with JW = 0.0633). The Bode plots for Fo - FoFrR

and FI - FiFIR are given in Figure 5.11. The maximum errors occur at fiequencies

around the transition bands of Fo and FI:

Page 111: Optimal design of multirate systems

12 i 1 ?

-020 I I I

IO 20 30 40 so 60 70 WI 90 rm

Figure 5.9. Cinit step response @ venus k.

Hence the 31, approximation errors are

Finally, we mention that the order issue of the synthesis filters is attacked by

model-reduct ion techniques in [Ml.

5.4 Performance Limitation

In the preceding example, we saw that as m increases, JO@ converges to some

nonzero value, indicating a performance limitation. In contrast, in the 7&, design

of digital multirate filter banks in [BI, it is proven that under some mild regularity

condition, one can always get arbitrarily close to perfect reconstniction (Jqt = O) by

tolerating large enough time delays. In this section, we will give an exphnation using

R, model-matching theory.

The problem at hand is as follows:

Page 112: Optimal design of multirate systems

Figue 5.10. Impulse responses versus k for Fo (upper) and FI (lower) .

Here, R is 2 x 2 and w and Q are both 2 x 2n, n being the number of inputs in

Figure -5.3. The nonsquareness of this Q implies that the associated 36, problem is

a tweblodr one; whereas in [la], the corresponding matrix is squaze, leading to an

one-blodr problem only.

For sorne rigomus andysis, assume that ~ ( r ) always has full row rank on the unit

circle. Then Q has a CO-inner-outer factorization [;?7]

where the outer Q,, the inverse 0;' and Q~ are ail in Go and Qn, being dl-pass,

satides Q ~ ( ~ ) Q ~ ( z ) = I . It foUows then that the matnx fimction

Page 113: Optimal design of multirate systems

= @ 11 [ &QX - RQ- W(I - Q ~ Q ~ ~ ) ] 11- (5.12) R

This is a *O-block ?&, problem: The fist block in the matrix in (3.12) now involves

a square Q,: the second block is independent of R. From here? we conclude that J,,

R e c d that ~ ( r ) can be expressed as (m = 2d or rn = 2d + 1) multiplying

some transfer matrix which is independent of d. This delay factor does not affect the

infinity nom in (5.13). This means that increasing the time delay m cannot improve

the lower bound on Jqt and so performance limitation exists.

A condition on perfect reconstmction (JOpt = 0) can be obtained too. For perfect

reconstruction, it is necessary that

In thîs case,

Page 114: Optimal design of multirate systems

99

which is a one-block problem. Sim.ïlarl,t* as in [18] me can show that this JW converges

to zero as rn tends to idhity. In other words, under the condition in (3.14), aibitrarily

perfect reconstruction is possible if one is w i b g to tolerate d c i e n t l y Iarge time

delay

The above analysis also shows that the lower bound on Jm in (5.13) is asymptot-

i c d y tight: Jm converges to this Iower bound if the time delay rn tends to infini%

In the example in the precedïng section, the limiting value JWt = 0.0633 corresponds

to this lower bound, which can be pre-computed without going through the 'fl, op-

timization, if a factor Q, cm be computed for Q. Some interesthg questions &se:

How to select andysis filters to make the lower bound smd? Wlat is the tradeoff in

terms of filter complexity?

In the case of perfect reconstruction, we have

In view of the factorkation in (5.11), postmultiply by to get

Since QX is anticausal, & and accordingly the synthesis filters Fi and F2 are non-

causal in general (see Step 5 in Subsection 5.2.3). In the next chapteq we s h d give

çome verifiable conditions for RQ to be causal.

5.5 Conclusions

In this chapter, we proposed an 31,-based procedure for minimax design of hybrid

filter banlis involving analog and digital Nters. The work forms an extension of

31, optimization to multirate systems in signal processing. The filters designed are

robust in the sense that uniformly good performance is achieved for the class of all

finite-energy signals. -4 design example was included for illustration.

Page 115: Optimal design of multirate systems

100

Our study treats only the twcdannel case. Extension of the r d t s to muitipk-

channel case is possible dong the hes in Section 5.2. However, the complexity of the

optimization problem grows with the number of channels and may become an issue

at some point-

In signai processing systems, FIR mers me ofien used to achieve a cost admage.

In the design example, the optimal IIR synthesis flters were tmcated to get FIR

approximations. -4 better approach for design of FIR synthesis filters of a given length

is by optimizing the filter coefficients direct1y. This leads to finitdknensiond convex

optimization which is studied for digital filter banks in [58]. We remark here that the

method used in [58] cm be applied to design of hybrid filter banks with FIR synthesis

a t e n using the fkmework proposed in this chapter.

Page 116: Optimal design of multirate systems

CHAPTER 6

ON CAUSALITY AND ANTICAUSALITY OF CASCADED LINEAR DISCRETETIME SYSTEMS

This brief chapter treats a fundamental question arising fiom both filter bank

design and UÎ/'fl, control: When two FDLTI systems, causd or noncausal, are

cascaded, what can be said about the causality properties of the overd system?

W e gïve conditions under which the cascaded system is causal or anticausd; these

conditions are related to Sylvester equations obtained fiom state-space data of the

sys t ems. For brevity, ody discrete-time syst ems are considered' .

6.1 Motivation

Consider cascading two FDLTI systems, Gl and G2, together as in Figure 6.1.

where both Gi and G2 can be MBlO. It is weIl-known that G2G1 is causal if both G1

Figure 6.1. The cascade of two systems

and G2 are causd. U l a t about if one or both Gi and G2 is not causai? For example,

suppose Gi and G2 are both single-input, single-output in d i ~ r e t e time with transfer

fimctions

Gl(z) = r-l: &(z) = ri, k,l > O;

it follows that G1 is causal, Gz is not, but GzGl may be either causal or noncausal.

depending on the two integers k and 1. The goal in this chapter is to answer such

lThe results in this chapter have been published in [66].

Page 117: Optimal design of multirate systems

10'2

questions in g e n d cases, in partidar, we give state-space conditions for testing

causality properties of G2Gi in the MIMO setup. To our best knowledge, such results

were not available in the literature,

This question can be motivated fiom both filter bank design and optimal control

t h e o i In the preceding chapter, the hybrid filter bank design was translated to

mode1 matchhg problem of a discrete-time system and the perfect reconstmction

condition is that a cascade connection equals a causal system

We have shonm that in general the synthesis polyphase rnatrix R (and so synthesiç

filters) must be non-causal to satisfjr this condition. Similar results arise in digitai

filter bank design, where, in some cases, a cascade connection are required to equd an

aaticausal transfer rnatriv [72,66]. The results in this chapter will give some verifiable

conditions for the cascaded system RQ to be causal or anticausal. In (361 a design

scheme was proposed using causal analysis filters and anticausal synthesis filters with

irnplementation issues discussed in [36, 461.

In control, though all practical controilers are causal for implementabiliw non-

causal systems appear in deriving various optimal controllers- For example. in deriva-

tion of Ra optimal controllers, we have cascaded an anticausd system with a causal

system to get an anticausa overall system (see Lemma 2.4). This is a special case of

Proposition 6.1 in this Chapter. -4nother case where noncausd control systems arise

is in discrete-time 31, controi [69] in which it is shown that it is possible to attain

optirnaiity via non-causal controllers.

In order to treat noncausd systems in discrete time, we need to consider signals

and systems dehed on the (time) set of aJ integers, both negative and positive. With

this in mind. a discretettirne LTI, MIMO system G is completely characterized by its

Page 118: Optimal design of multirate systems

impulse response: a rnatrix sequence,

This is in general a two-sided sequence. To make things precise, we define the fouowing

terms: G is causal if g(k) = O whenever k 5 -1; G is strictly causal if g ( k ) = O

whenever k 0; G is anticBusal ifg(k) = O whenever k 2 1; G is strictiy anticausa1 if

g(k) = O for k > 0; and G is noncausal i f g ( k ) # O for some k -1. It is clear from

the definitions that a noncausal systern G can be decomposed into G = G, + Ga with

G, causal (stnctly causai) and Ga strictly anticausal (anticausd).

m e n tw-O discrete-the systems are cascaded as in Figure 1, is the resultant system

causal or anticausal? Wk s h d Grst look at some specid cases involving a causal

system and an anticausd system and then tadde the generd case. Throughout the

chapter we çhall assume systems in discussion are finite-dimensional.

6.2 Special Cases

In this section, we assume Gl is causal and G2 is anticausal. What c m be said

about causality properties of GzGl and GiGz? The analysis to follow is based on

state-space data.

Since Gl is LTI, finitedimensional and causal, it has a state-space realization (with

input ull output ~ 6 ~ : and state 6):

The transfer matrix is written

(with the subscript c denoting a causal system), or as a power series in z-l :

Page 119: Optimal design of multirate systems

104

Since Gz is anticausal, there exist matrices BZ, c2 and i)2 such that G2. mith

input v2 and output is describeci by the following kckwutd statespace equations:

The corresponding t rader matrk is

where subscript a means anticausality. In terms of power series: we have

The causality properties of G2Gl will relate to the Sylvester equation

ahich is more general than the Lyapunov equation we saw in (2.9). Assume that

both Al and .a2 are stable, i.e., the eigendues of Al and L& are in D. (6.3) admits

a solution of infinite series:

If t his equation has a unique solution, or equitdentiy [30], no eigenvalue of At is

L ow we c m reciprocal of some eigenvalue of A*: then this solution equals X in (6.4). Y'

state our fist result-

Proposition 6.1 Assume X in (6.4) ezists.

(a) G2G1 is causal if B2L4 + Â 2 ~ ~ ~ 1 = O and is strictly causal if, in addition, B2& + = O. Moreooer, the converses hold i f the pair (&: Â2) iS o bsentable.

(6) 4 G 1 is anticausal if D ~ C ~ + C?XA~ = O and is strictly anticausal if, in addition.

D ~ D ~ + & X B ~ = O. Moreover, the converses hold if(&, B I ) is controllable.

Page 120: Optimal design of multirate systems

Proof The trader rnatriv for G2Gi is

The coefficient matrices g(k) can be obtained by multiplying (formally) the two series

in (6.1) and (6.2) together and using (6.4) to simpiify the expressions; then we geto

Thus the d c i e n c y proof follows easily from respective definitions. For necessi@

note that causality of G2G1 implies g(k) = O for k E. -1: which implies in tuni

where nt is the dimension of the square matrix Â2. Now observability of (&A2) implies that the observability matrk, appeared on the right-hand side of (6.5): is

injective and hence we get B ~ D ~ + $x& = O. The rest of the necessin- proof

-lote from the proof that if B& + Â2xl?i = 0, the causal system G2Gi has a

st ate-space reahzation

This means that all the anticausal poles introduced by G2 are cancelled by zeros of

Gt for the resuitant system to be causal. Similar interpretations can be given for the

ot her cases.

The dual of Proposition 6.1, which concerns causality properties of Gi G2 (G1 and

Gz as in Proposition 6.l), dl involve a dinerent rnatrix

Page 121: Optimal design of multirate systems

with the assotiated Sylvester equation

The following r d t s can be proven andogously.

Proposition 6.2 Assume Y in (6.6) &ts.

(a) G1G2 is causai if D ~ C ~ + C~Y-& = O and is strictly causai if, in addition, D ~ & + C~Y& = O . Moreover, the converses hold if the pair (&: $) is controllabk.

(b) G1 Gz is anticattsal if B~ D~ + A ~ Y B~ = O and is strictly anticausa1 if, in addition,

~~b~ + C ~ Y & = O. ~ M o ~ ~ o v ~ T , the converses hold if (Cl: .4i) is obsentable.

6.3 General Case

Consider G2Gi, where Gi and G2 are LTI, finite-dimensional, but general non-

causal systems. Decompose them into causal and strictly anticausa parts:

Bring in state-space matrices:

&te that the direct feedthrough terms in G&) are zero. Then G2G1 is the sum of

four cascaded systems

W e wish to h d the impulse response of each of the four systems.

First? given the state-space realizations of the causal Gi, and G2,, we get a real-

ization of G2,G1, via a formula in: e.g., [27],

Page 122: Optimal design of multirate systems

107

To simplify this, introduce the Sylvester equation

and assume it has a solution O-. (It has a unique solution in A2 and Al have no

common eigenvalue [30] .) Performing a similarity transformation, with

Hence the impulse response g&) of GacGiC is given by: g&) = O for k 5 -1:

g c m = 9D1: and

Similady, introducing the Sylvester equation

and assuming this equation has a solution V : we can obtain the impulse response of

G2.G1.: g,,(k) = O for k 2 O and

g , , (k)=-~2 .4;k-1~Êl+&~A;kœ1~17 k s -1.

The other t~vo cascaded systems in (6.7), &Glc and G2,GI,, have been studied

in Propositions 6.1 and 6.2; so we define

Again, if such X and Y exist, they must satidy the associated Sylvester equations:

Page 123: Optimal design of multirate systems

108

Then as in the derivations of Propositions 6.1 and 6.2, the impulse responses of GZa Glc

and G2,Gia are respectively (note that the d w c t feedthrough terms of bl, and

Xow putting things together, fiom (6.7) we get the impulse response g for the

overd system G2G1:

Based on this. the next results follow easily.

Proposition 6.3 Assume U and V are solutions of equations (6.8) and (6.9) respec-

tiüely; and X and Y defined in (6.10) ezist.

(a) G2G1 iS causal if

and is stn'ctly causal if, in addition,

(b) G2G1 is anticausal if

and is strictly anticausal if, in addition, (6.13) holds.

Page 124: Optimal design of multirate systems

109

Partial necessity conditions can be given too. For example, if (6.11) holds. then

causai^. of G2Gi and controllability of (A1, BI) imply (6.12); if (6.12) holds, then

causality of G2G1 and obseftability of (&A) imply (6.11). Similar staternent can

be made about anticausa3ity and conditions (6.14) and (6.15)-

Finally, we mention that the resdts in this chapter can dso be extended to

continuous-time systems; for details, see [66].

Page 125: Optimal design of multirate systems

CHAPTER 7

CONCLUDING REMARKS

This thesis has been devoted to both theory and applications of multirate systems.

Xenr dgorithms have been developed for multirate optimal control with X2 and 7&

criteria. C o m p d to the previous solutions [45, 74, 191, the resdts obtained have

the advantage of computational efficiency and ease of implementation on cornputers.

A software package based on the multirate 'H, dgorithms has ben developed and

applied to a power system stabilizer design. The multirate 3tz atgonthms were applied

to case studies in -6,571. .Uso. new MATLAB softwxe for multirate controllers based

on the work in Qiu and Chen [54: 19, 581 and the algorïthms in this thesis is under

development [56]. In addition, we have provided a study for hq-brid multirate filter

banks via Xm optimization.

It is no doubt that multirate systems have many advantages .A lot of research

efforts are required in this area. In what follows, we s h d suggest some future work

relative to the topics in the thesis.

Multirate Optimal Control

In recent study of multirate optimal control, the focus was on LQR, LQG, 3C2, R,

and el criteria (2; 43,20; 74,54,55,19: 651, all with a single optimal criterion. Further

research should tadde more sophisticated specs such as mixed H2/7frn, mixed l1 /'Mm :

ceneralized ;H2 , lfCoc with time-domain constraints, and multiple objective ?&/?Lw. Of b

course. the studies of those subjects depend on the corresponding results for single-

rate problems. For example, for the single-rate mixed ?&/Ra problem? there are

many forms of expressions so far (see, e.g., [IO, 61, 23, 70, 621): and the solutions

to the general problern remain, to a large extent. u n h o m . In spite of this. i t is

Page 126: Optimal design of multirate systems

111

still worthwhile to look at the multirate setting in terms of the available single-rate

solutions. Chen and Qiu's work (19, 551 provides an effective h e w o r k based on

nest matrices for the problems with the mixed/multiple specs. Future study shouid

focus on how to use the tool to handle causality constaints. Wë &O mention that

the solution to the mdtirate minimum entropy probIem, a specid case of the mixed

X2 /lz/'flo problem [49], has been given in [55].

Another interest ing topic on multirate control is finitehorizon optimization prob-

lems wïth LQG or 36, type of performance meastues [35]- Controllers for finite-

horizon measures are dways timevarying. Bence, it may be possible to look at

time-va.ryZng systems directly. For the multirate LQG problem, we believe that the

idea to do the lifting in [54] may be borrowed to reduce the multirate problem into

a single-rate LQG problem subject to a causality constraint. However, it should be

recognized that finitehorizon problems require solving diffeerential Riccati equations

and the techniques involved to handle causality constrkints will be essentially dinerent

from the infinite-horizon cases.

Power System Stabilizers

Today's power -stems feature mdti-machines, multiple areas and htercomected

complex networks. Design of power system stabilizers in such a saphisticated environ-

ment becomes extremely dïfEcuIt in the sense of controller coordination and parameter

tuning. Hence, it is desirable to use simple weighting hct ions to reduce the amount

of trial and error involved. However, weighting functions must be chosen to reflect

the basic operational requirements of the system, as mentioned in Section 4.2.2. The

method for choosing weighting functions presented in Chapter 4 lays the groundwork

for syn t hesising mult i-machine systems. In addit ion, since mult i-machine systems

involve more sipals with diflerent bandwidths, we beüeve that multirate sampling

schemes provide more freedom of design. But the coordination of controllers may

Page 127: Optimal design of multirate systems

become more difncult than singIemachine systems.

Hybrid Filter Banks

Compared to digital fdter banks, hybrid filter b d s , though more attractive con-

cept~*, are more diaicult to design due to non-ideal analog filters. At ppresent,

selection of the ânafog filters is stiU based on trial and error. W e feel that a sys-

tematic procedure is needed to help choosi~g the analog filters. For this, one would

need to look into the connedion between the state-space representations of the analog

filt ers and the performknce limit derived in Section 5.4.

S tudy of hybrid filter banks is relatively new in signal processing. More research

efforts are required to analyze hybrid filter banks in the frrquency domain. Some

related questions are: How to define for hybrid filter banks the f d a r concepts in

digitd filter banks such as phase distortion and magnitude distortion? How to relate

these distortions to the 3t, performance masure proposed in Chapter 5?

Page 128: Optimal design of multirate systems

[1] S. S. .r\hmed, L. Chen and A. Petroimu, "Design of suboptimal X- excitation

controllers", IEEE Trans. on Power Systems, vol. 11, no. 1, 312-328, 1996.

[-21 H. M. -41-Rahmani and G. F- Franklin, "A new optimal mdtirate control of

h e a r periodic and tirneinvariant systems", IEEE Trans. ilutomat. Control, vol.

35,406415, 1990.

[3] H. M. AT-Rahmani and G. F. F r a . , "Techniques in multirate digital control".

Digital Control Systems Impiementation Techniques, Academic Press: 1995.

[4] P. M. Anderson and .4. A. Fouad, Power System Control and StabiLity, Iowa

University Press, Iowa 1977.

[5] R. hsgharianl 'A robust R, power system stabilizer with no adverse effect on

shaft torsional modes", IEEE Trans. on Energy Conversion, vol. 9 , no. 3, 475-

481, 1994.

[6] K. J. Astr6m and B. Wittenmark, Cornputer ControlZed Systems: Theory and

Design, Pretice-Hd, Englewood Clins, N.J., 1984.

[ f ] G. J. Balas, J. C. Doyle, K. Glover, A. Packd, and R. Smith, User's itlannai

for p -Analysis and Synthesis Toolboz, Mathworks: 1991.

[8] B. Bamieh and J. B. Pearson, "A general framework for hear periodic systems

with application to 'fl, sampled-data control" , IEEE Trans. Automat. Contr01~

vol. 3'7, 418-435, 1992.

[9] S . K . Berberian, Introduction to Hilbed Space, Chelsea New York, 1976.

Page 129: Optimal design of multirate systems

114

[IO] D. S. Bernstein and W. H. Eaddad, %QG control with an 36, performance

bond: a Riccati equation approachn, IEEE Tram Automat. ControZ, vol. :H.

293-305, 1989-

S. P. Boyd and C. H . Basratt, Linear Controllet DmeSEgn: Limits of Pe~fonnance~

PrenticeHall: 1991.

C.-T. Chen, Linear System Theory and Design, Holt, Rinehart and Winston.

Xew York, 1984.

S. Chen, Power System Ilncertainty Analysis and Robust StabiZizer Designo Ph.D.

Dissertation, Department of Electncal and Cornputer Engineering, University of

Calgary, 1995.

S. Chen and O.P. Malik, =Power system stabilizer design using p spthesis".

IEEE Trans. on Energy Conversion, vol. 10, no. 1, 175-181. 1996.

[15] T. Chen and B. A. Francis,"Input-output stability of sampled-data systems'?

IEEE Trans. Automatic ControZ, vol. 36, 50-58, 1991.

[16] T. Chen and B. -4. Francis, "State-space solutions to discrete-time and sampled-

data X2 control problems" : Proc. CDC. 1111-1116, 1992.

[17] T. Chen and B. A. Francis, Optimal Sampled-Data Control Systems, Springer-

Verlag, 1995.

[18] T. Chen and B. A. Francis, &Design of rnultirate filter banks by 31, optimiza-

tion" , IEEE Trans. Signal Processing, vol. 43, 2822-2830, 1995.

[19] T. Chen and L. Qiu, "X, design of general rnultirate sampled-data control sys-

tems"? Automatica. vol. 30, no. 7, 1139-11521 1994. \

Page 130: Optimal design of multirate systems

11.5

@O] M. A- Dahleh, P. G. Vodgaxis, and L. S. Valavani, "Optimal and robust con-

trollers for periodic and multirate systems" , IEEE Trans. Automat. ControZ? vol.

37, 90-99, 1992.

@] K. R Davidson, Nest Algebm, Pitman Research Notes in Mathematics Series.

vol. 191, Longman ScientSc k Technical, 1988.

[22] J. C. DoyIe. K. Glover, P. P. Khargonekar, and B. A. Francis, uState-space

solutions to standard 'flz and 31, problems", IEEE Truns. Automaf. Control.

vol. 34, 851-847, 1989.

@3] J. C. Doyle, K. Zhou, K. Glover, and B. Bodenheimer, "Mixed Ra and Rfl,

performance objectives II: optimal control" , IEEE Trans. A utomat. Control, vol.

39, 15ïb158?, 1994.

@4] F. P. deMeII0 and C. Concordia, "Concepts of spchronous machine stability as

dected by excitation control", IEEE Tmns. on Power .4pparatus and Systems.

vol. PAS-88, no. 4, 316-329, 1969-

r25] A. T. Fam. "MFIR filters: properties and applications", IEEE Trans. Acoust.,

Speech, Signal Process., vol. ASSP-29, 1128-1136: 1981.

[-61 A. Feintuch? P. P. Khatgonek, and A. Tannenbaum, the sensitivity min-

imization problem for linear time-varying periodic systern~"~ SL4M J. Control

and Optirnizution, vol. 24, 1076-1083, 1986.

[27] B. A. Frasùs, -4 Course in 'fl, Control Theory, Springer-Verlag, Kew York,

1987.

[.BI B. A. Francis and T. T. Georgiou, "Stability theory for linear time-inkariant

plants with periodic digital controIlers", IEEE Tram. Automat. Contmlr vol. 33,

820-532, 19S8.

Page 131: Optimal design of multirate systems

116

p91 G. F. Franklin, J. D. Power, and M. L. Workman? Digital Control Dynarnic

Systems, Addison-Wesley, 1990.

[30] F. R. Gantmacher, The Theory of Matrices, Chelsea New York, 1960.

[dl] T. T. Georgiou and P. P. Khargonekar, 'A constructive algorithm for sensitivity

optimization of periodic systems" , SIAM J. Contrd and Optirniration, vol. 2.5.

334340, 1987.

[32] D. P. Glasson, 'Development and applications of multirate digital control", IEEE

Control Systems Magazine, wl. 3, 2-8, 1983.

[33] K. Glover and J. C. Doyle, 'State-space formulae for aU stabiiizing controllers

that s a t i e an Ra-nom bound and relations to risk sensitivity" , Systems and

Control Letters. vol. 11, 167-1'72, 1988.

[34] R. A. Gopinath, J. E. Odegard, and C. S. Buinis, "Optimal wavelet represen-

tation of signds and the wavelet sampling theoremZl IEEE Tram. Cirniits and

Systems II: Analog and Digital Signal Processing, vol. 41, 262-277, 1994.

[35] M. Green and D. J. N. Limebeer, Linear Robwt Control, Prentice-Hd, Engle-

wood Cliffs, 1993.

[36] J. Huso? and T. Ramstad, YApplication of an efficient paallel IIR fdter bank to

image subband codingR . Signal Processing, vol. 20, 279-292, 1990.

[37] P. A. Iglesias and K. Glover, "State-space approach to discrete-time R, control" . Int. Journal of Control, vol. 54, 1031-1073, 1991.

[38] P. P. Iihargonekar, K. Poolla, and A. Tannenbaum, "Robust control of linear time

invariant plants using periodic compensation", IEEE Tram. Automat. ControL

vol. 30, 1088-1096, 1985.

Page 132: Optimal design of multirate systems

117

[39] P. P. Khargonekar and 31. A. Rotea, 'Multiple objective optimal control of linex

systems: the quadratic nom cas$, IEEE Trans. Automat. Control, vol. 36, 14

24, 1991.

[a] A. Khayatian and D. G. Taylor, %Idtirate modeling and control design for

switched-mode power converters", IEEE Trans. Automat. Control, vol. 39- 1848

185'3, 1994.

[41] B. C. Kuo, Digital Controi Systevq Holt, Rinehart and Winston, New York,

1980.

[$2] E. V. Larsen and D . A. S wannt 'Applying power system stabilizers" , IEEE Trans.

on Power dpparatus and Systems, vol. PAS100, 301'7-3046, 1992.

[43] W. M. Lu, K. Zhou, and J. C. Doyle, 'Stabilization of LFT systems", Proc.

CDC, 1239-1244, 1991.

[44] D. G . Meyer, %.A parametrization of stabilizing controllers for miiltirate sampled-

data systems", IEEE Trans. Automat. Control, vol. 35, 233-236, 1990.

[45] D. G. Meyer, "Cost translation and a lifting approach to the multirate LQG

problem" IEEE Trans. Automat. Control, vol. 37, 1411-1415, 1992.

[16] S. K. ;Mitra, C. D. Creusere, and H. Basic, "A novel implementation of perfect

reconstruction QMF banks using IIR fdters for infinite length signalsn , IEEE fnt.

Symp. on Circuits and Systems, vol. 5, 2312-2315: 1992.

[47] M. Mohamed, D. Thorn, and E. Hill, "Contrast of power system stabilizer per-

formance on hydro and thennal unitsn , IEEE Trans. on Power: Apparatus and

Systems. vol. PAS-99, 1522-1533, 1980.

Page 133: Optimal design of multirate systems

118

[48] S. Monaco and D. Normand-Cyrot, "An introduction to motion planning under

mdtirate digital controln , Proc. CDC, 178780-1785, 1992.

[49] D. Mustafa, -Relations between maximum entropy/% control and combined

31,JLQG control" , System C o n t d Letters, vol. 12, 193-203, 1989.

[50] P. Nagel, i'rYon-causal complementary filters" , Signal Processing, vol. 29, 151-164.

199'2.

[jl] K. Ohtsuka, T. Taniguchi, T. Sato, S. Yokokawa, and Y. Ueki, "An 'fl, optimal

t heory-based generator cont rol system" , IEEE Trans. on Energy Conuersionl vol.

7, no. 1, 108-ll5? 1992.

[Z] T. Pappas, A. J. Laub, and S. R. SandeU, Jr., "On the numerical solution of the

discrete-time dgebraic Riccati equation". IEEE Tmns. Automat. Control, vol.

-25: 631-641, 1980.

[53] -4. Petraglia and S. K. Mit- %Hi&-speed A/D conversion incorporating a QMF

bank", IEEE Trans. Instrumentation and Measurement, vol. 41, 427-431: 1992.

[541 L. Qiu and T. Chen, '7i2-optimal design of multirate sampled-data systems" ,

IEEE Trans. -4utomat. Control, vol. 39, 2506-2511, 1993.

[55] L. Qiu and T. Chen* "Multirate sampled-data systems: al l31, suboptimal con-

trollers and the minimum entropy controllern , Proc. CDC. 3W2-3712: 1991.

[56] L. Qiu. T. Chen, A. Potvin, J. Gu, K. Tan, Y. Sun and H. Shu, "Development of

M.41LAB toois and case study for multirate control design", Proc. CDC, 1996.

[Si] N. Rafee, T. Chen, and 0. P. Malik, "Multirate discretization of analog con-

trollers", Proc. Canadian Conf. on Electrical and Cornputer Engg, 554-557? 1996.

Page 134: Optimal design of multirate systems

119

[SI S. Ratzlaff, An Optimal Model-Matching dpproach to Multirate FIR Filter Bank

Design, MSc. Thesis, Department of Electncat and Compter Engineering, C'ni-

versity of Calgary, 1993.

[59] R. Ravi, P. P. Khargonekar, K. D. - W o , and C. K. Xett, YController param-

eterization for time-varying multkte plantsn, IEEE Tram Automat. C o n t d

vol. 35. 1259-1262: 1990.

[60] R Y. Redheffer? "On a certain linear fiactional transformation', J. Math. Ph ys.,

vol. 39, 269-286, 1960.

[61] M. -4. Rotea and P. P. Khargonelar, Wz optimal control wïth an R, constraint:

the state feedback case" Proc. -4 CC, 2380-2384, 1990.

[62] G. W. Scherer, "5I:dti-objective R2/& control", IEEE Tmns. Automat. Con-

tro6 vol. 40, 10541062,1993.

[63] R. G . Shenoy, " . b d y s i s of multiate components and application to multirate

filter designz, Proc. ICASSP: vol. 3, 121-124,1994.

[641 R. G. Shenoy: D. Burnside and T. W. Parks, "Linear perîodic systems and mul-

tirate filter designn, IEEE Tram. Signal proeessing, vol. 42, 2242-2256: 1994.

[65] H. Shu and T. Chen, "State-space approach to discretettirne R2-optimal control

with a causality constraint", Systems Control Letters, vol. 26, 69-77, 1995.

[66] B. Shu and T. Chen, "On causality and anticausality of cascaded discrete-time

systems", IEEE Trans. Circuits and Systems 1: Fundamental Theory and Appli-

catio. vol. 43, 240-242, 1996.

Page 135: Optimal design of multirate systems

120

[67] H. Shu, T. Chen, and B. A. Francis, "Minimax design of hybrid mdtirate filter

banksn, IEEE Tmns. Circuits and Systems II: Analog and Digital Signal Pro-

cessing, vol. 44. no. 2, 120-128, 1997.

[68] H. Shu and T. Chen, "Robust digital design of power system stabilizers", to

appear in Pmc. ACC, 1997.

[69] A. A. Stoorvogel, The 71, Control Problem: A State-Space Approach? Prentice

Ha& 1992.

[?'O] M. Sznaier? 'An exact solution to general SIS0 mixed 3tz/'?& problems via

convex optimization", IEEE Trans. Automat. Contrd: vol. 39, 2511-2517, 1994.

[71] G. N. Taranto, J. H. Chow, and H. -4. Othman, 'Robust design of power s-stem

damping controllers", IEEE Truns. on Control Systems Technology, vol. 3: no. 3,

290-298, 1993.

[72] P. P. Vaidpathm, Multirate Systems and Filter Banks, Prentice Hall, 1993.

[7:3] C. F. Van Loan, "Computing integrals involving the matrix exponentialz, IEEE

Trans. Automatic Control, vol. 23, 395-404, 1978.

[ï4] P. G. Voulgaris, M. A. Dahleh, and L. S. Valavani, "fl, and X2 optimal con-

trollers for periodic and multirate systems", Automatica, vol. 30, 251-263, 1994.

[75] P. G. Voulgaris and B. Bamieh: "Optimal 31, and 3Lz control of hybrid multirate

systems", Systems Control Lefters? vol. 20, 249-261, 1993.

[76] K. Zhou, J. Doyle, and K. Glover, Robust end Optimal Control, Prentice-Hall,

1995.

[Ti] K. Y. Zhu and K. V. Ling, "Multirate control for a gaification pro ces^"^ Proc.

Am 3268-32R1 1994.

Page 136: Optimal design of multirate systems

APPENDIX A

P O m R SYSTEM MODEL AND PARAMETERS

First defme the fouowing signak involved in modeling of the singlemachine infinite-

bus pomr system: 6: power angle

machine rotor slip speed mechanical torque electricd torque quadrature-Oms transient voltage generator field voltage armature m e n t , direct component armature m e n t : quadrature component armature voltage, direct component mature voltage: quadrature component bus voltage (us = 1 p-u. for idhite bus) reference (disturbance) voltage terminal voltage control voltage from the stabilizer (if any).

The Park's two-axis representation [24,4] is used in the thesis for the generator.

and for the transmission lines,

Page 137: Optimal design of multirate systems

1 22

These are obtained by neggecting tansients in the stator cirait and the effect of

the rotor amortisseur. The AVR and exciter are modeled by a simpMed first-order

system:

Taif = -ef + K,(Ker - ut tu).

The parameters involved in the above equations axe given in Table A.1, which are

Table Al. Parameters for the power system model.

based on a Iaboratory setup at the University of Calgayy [13]. In this table, aIl the

time constants are given in seconds, ;J. is in radfsec, and the other quantities are in

p-u-