You are on page 1of 225

Diss. ETHNo.

10666

20. JUll

KWH

Ma,

Robust Control of Industrial

an

High-Purity

Distillation Column

A dissertation submitted to the

SWISS FEDERAL INSTITUTE OF TECHNOLOGY


ZURICH
for the

degree

of

Doctor of Technical Sciences

presented by
HANS-EUGEN MUSCH

Dipl. Chem.-Ing. ETH


born June 19,1965
citizen of Germany

accepted

on

the recommendation of

Prof. M.

Steiner, examiner

Prof. Dr. D. W. T.

Rippin,

co-examiner

1994

Leer

Vide

Empty

To my

grandparents

Acknowledgments

This Ph. D. thesis

was

written

during my years

as a

research and educa

tional assistant of the Measurement and Control

Laboratory
persons

at the

Swiss Federal Institute of Technology (ETH) at Zurich. I would like to take


this

opportunity
this project.

to

thank the

numerous

who

have

supported

First of all I express my

gratitude
are

to Prof. M. Steiner. He

arranged this

project and helped


environment.

to

overcome

many difficulties with the industrial

Many
clarity.

thanks

also due to him and to Prof. D. W. T.

Rippin for improved


The

the critical examination of this thesis, which

essentially

its

numerous

discussions with my colleagues and their uncountable

suggestions
mentioned.

gave rise to

important contributions

to this work. In this

context, E. Baumann, U. Christen, and S. Menzi must be specially

Last but not least I should

emphasize the support of B. Rohrbach. She

never lost her patience with my never ending questions concerning the English language. Without her willingness to correct the manuscript,

the choice of the English language for this thesis would have been

impossible.

Content

Symbols
Abstract

13

15

Kurzfassung

17

Introduction
"Modern Control:

19

1.1 1.2

Why Don't We Use It?"

19 21
21 21

Scope and significance of this thesis


1.2.1 1.2.2
1.2.3 1.2.4

Distillation

as a

unit

operation example

Earlier research
Robust control and nonlinear

plants

22 22

Contributions of this thesis

1.3 1.4

Structure of the dissertation

23 26

References

The Distillation Process

An Industrial
2.1

Example

29 29

Introduction Column

2.2

design and operation

29
32

2.3
2.4

Steady-state behavior

Composition dynamics
Control
2.5.1

35 37 39 40

2.5

objectives and configurations

The 5x5 control

problem

2.5.2

Control design steps

2.6

Tray temperatures
2.6.1

as

controlled outputs

41 42

Pressure-compensated temperatures

2.6.2
2.7

Temperature

measurement

placement

44 45

References

Rigorous Dynamic Model of


47 47
48

Distillation Columns
3.1

Introduction
Conventions

3.2 3.3
3.4

The

objective

of modelling

48
48

Simplifying assumptions
Balance
3.5.1
3.5.2

3.5

equations

51
51
52

Material balances

Energy balance equations

3.6

Fluid
3.6.1

dynamics

55
rates

Liquid flow
Pressure

55 57
59 59 60

3.6.2
3.7

drop

Phase
3.7.1

equilibrium
Vapor phase composition Boiling points

3.7.2
3.8

Volumetric
3.8.1 3.8.2

properties

60
61 61

PVT relations

Density

3.9

Enthalpies
3.10.1 The

62 63
63 66

3.10 Numerical solution

dependent variables and the equation system...

3.10.2 Formal

representation

of the DAE

3.10.3 The index 3.10.4 Solution methods and software


3.11 3.12

66
67 71 74

Notation

References

Linear Models

77 77

4.1

Introduction
How to linearize the
4.2.1 4.2.2

4.2

rigorous model?

78 78 80
80

The state,

input, and output vectors

Handling of the algebraic equation system simplified nonlinear model

4.3

Linearization of a
4.3.1
4.3.2

The

simplified model

80
84

Analytical linearization

4.4

Linearization of the rigorous model


4.4.1 4.4.2

86 86 88
89

Model modifications Numerical linearization


linear models

4.5

Comparison of the
4.5.1

Open loop simulations

89
92

4.5.2 4.6
4.7

Singular values

Order reduction

94 96
97 101 101
102

Summary

4.8
4.9

Appendix: Model coefficients


Notation
4.9.1 4.9.2

Matrices and Vectors

Scalar values

4.10 References

103

A Structured

Uncertainty Model

105
105 106

5.1 5.2 5.3


5.4

Introduction
Limits of

uncertainty models

Input uncertainty
Model
5.4.1

107
110 110

uncertainty
Column

nonlinearity
dynamics

5.4.2

Unmodelled

117

5.5 5.6 5.7 5.8

Measurement uncertainty

118

Specification of the

controller

performance

119 120 122

Summary
References

|0,-Optimal

Controller

Design

123
123

6.1

Introduction The structured singular value


6.2.1 6.2.2
6.2.3

6.2

124 124 126 128

Representation of structured uncertainties


Definition of the structured Robustness of stability and

singular value performance

6.3

The

design model

130
133 134 137 137
148

6.4

Controller design with u-synthesis


6.4.1

Synthesis algorithms Applying the DK-Iteration Applying the uK-Iteration


structure

6.4.2 6.4.3

6.5

Design of controllers with fixed


6.5.1

Diagonal PI(D) control

structures
...

149

6.5.2
6.5.3

PI(D) control structures with two-way decoupling


PID control structures with one-way

156 161
164

decoupling

6.6

Summary
References

6.7

166

Controller

Design

for

Unstructured Uncertainty
A

Comparison

169

7.1 7.2

Introduction

169 170 170

Diagonal Pl-control
7.2.1

The BLT method

11

7.2.2

Sequential loop closing Optimized robust diagonal Pi-control

172 174
177 182 187 187

7.2.3 7.3 7.4 7.5 7.6

Pi-control with

decoupling

H optimal design

Summary
References

8 8.1

Feedforward Controller

Design

189 189 190 190


190

Introduction
The

8.2

design problem
The

8.2.1 8.2.2

design objective
or

One-step

two-step design?

8.3 8.4
8.5 8.6

Hro-minimization

192 196 199


200

Optimization approach

Summary
References

9 9.1 9.2

Practical

Experiences

203
203
204

Introduction
Controller implementation

9.3 9.4 9.5 9.6

Composition
Controller

estimators

207

performance

208
214

Economic aspects

Summary

214

10

Conclusions and
Recommendations
217

10.1

Introduction

217

12

10.2 10.3
10.4

Controller

synthesis
or

218 219
220 221

State-space

PID control?

How many temperature measurements? Column models

10.5 10.6

Recommendations
10.6.1 Academic research

221 221 222 223

10.6.2 Decentralized control systems 10.6.3

Cooperation industryuniversity

Curriculum vitae

225

Symbols

8
A

Uncertainty scalar value Uncertainty matrix


Parameter vector
or

deviation from nominal

operating point

8
k

Condition number,

ov
ind.x

/o_.
nun

X (j, p
a

Eigenvalue
Structured

singular value

Spectral radius Singular value


Bottom

B D d
e

product

stream

(mol/s)
or

Distillate stream (mol/s)


Disturbance

diagonal scaling

matrix

signals

Control

error

Feed flow rate

(mol/s)

7t
G(s)

Lower fractional transformation


Transfer function

Gu
I

Transfer function from control

signals

to

output signals

Identity matrix
Controller Reflux (mol/s) Joint

K(s)

L0
M
P
p

weighted plant and controller,

(P, K)

^(P, K)

Weighted plant
Pressure

(N/m2)

Reference

signals
=

Se(s)

Sensitivity function at e, Se (s) Sensitivity function Temperature (C)


at u,

[I [I

G (s) K (s) ]

-1

Su(s)
T

Su(s)

K(s)G(s)]_1

Tr
u

Transfer function from reference

signals

to

output signals

Control

signals

V51
W(s)
w(s)
xrj xg xF
y

Boilup (mol/s)

Diagonal matrix of weighting transfer functions


Weighting transfer function
Top product composition (mol/mol)
Bottom
Feed

product composition (mol/mol)

composition (mol/mol)

Output signals

15

Abstract

It is well known that

high-purity distillation columns

are

difficult to

control due to their ill-conditioned and

strongly
within

nonlinear behavior.
a

Usually distillation columns


difficult.

are

operated

wide range of feed

compositions and flow rates, which makes


Nevertheless,
a

a control design even more tight control of both product compositions is

necessary to

guarantee the smallest possible energy consumption,


and uniform

as

well

as

high

product qualities. approach


for the dual
range of
a

This thesis discusses

a new

composition control
distillation column

design,
column,

which takes the entire

operating
an

into account. With the


a

example of

industrial binary distillation

structured uncertainty model is

developed which describes

quite well the nonlinear column dynamics with several simultaneous


model uncertainties. This uncertainty model forms the basis for feed
back controller

designs by |x-synthesis

or

u-optimization. The resulting

controllers robustness

are

distinguished by

high

controller

performance and high


well
as

guaranteed for

the entire operating range. This method


as

enables the

synthesis of state-space controllers

the u-optimal

tuning
The

of advanced PID control structures. of feed flow disturbances

already satisfactory compensation

can

be

improved
design

even further by use of feedforward control. Even for the

design design
yields

of the feedforward controllers the basic ideas of the feedback controller


can

be

employed.

A simultaneous feedforward controller


extreme column loads

for two column models

representing the

outstanding

results. Similar to the feedback controller

of state-space controllers

by Hm -minimization

or an

design, a design optimal tuning of


is

simple feedforward possible.


Control engineers
of the

control structures

by parameter optimization

working

in

an

industrial environment

are

conscious

high

effort needed for the

implementation of state-space control-

16

lers in
on

distributed control system. Therefore

controller

PID

or

advanced PID control structures is of

design based high relevance for the


PID control struc

industrial
tures is

practice. Usually, the performance of these


to

expected

lag significantly

behind the

performance of highperformances of
controllers,
detected.
are

order state-space controllers. However, comparing the

the state-space controllers with those of the advanced PID

merely slight advantages


This

of the state-space controllers


an

surprising result, achieved with


a

unconventional tuning of the


of advanced
a

PID control structures, allows the PID control structures in

simple implementation

decentralized control system without

significant
The

loss of controller

performance.

good robustness properties and the high performance of the control


are

schemes

confirmed by the
on a

implementation

of

an

advanced PID

control scheme

real industrial distillation column. An estimation of

the economic benefits made


effort

by this project much

more

than justifies the

expended.

17

Kurzfassung

Bekanntermafien sind Rektifikationskolonnen mit hohen Produktreinheiten wegen ihres schlecht konditionierten und stark nichtlinearen Verhaltens
Bereich

schwierig
was

zu

regeln. Haufig werden


von

sie in einem weiten

unterschiedlicher den Entwurf

Zulaufkonzentrationen

und

-mengen

betrieben,
Dennoch

Regelungen zusatzlich erschwert.

ist

eine

gute

Regelung beider Produktkonzentrationen


moglichst kleinen Energieverbrauch

notwendig,
zustellen.

um

einerseits einen

und andererseits hohe und einheitliche

Produktqualitaten sicher-

Diese Arbeit beschreibt einen

neuen

Ansatz fur den Entwurf

von

Konzentrationsregelungen,
Rektifikationskolonne

der den gesamten Arbeitsbereich

einer

berucksichtigt.

Am

Beispiel

einer industriellen

binaren Rektifikationskolonne wird ein strukturiertes Unsicherheitsmodell entwickelt, welches das nichtlineare dynamische Verhalten der Rektifikationskolonne
durch mehrere

Modell-Unsicherheiten
bildet
die
Basis

gut
den

beschreibt.
Entwurf von

Dieses

Unsicherheitsmodell

fur

Reglern mittels u-Synthese oder u-Optimierung. Die resulzeichnen sich durch eine
-

tierenden

Regler
aus.

iiber den gesamten


grosser
von

Betriebsbereich garantierte Robustheit


Dieses

hohe

Regelqualitat bei sehr

Vorgehen erlaubt sowohl den Entwurf Berechnung u-optimaler

Zustandsregelungen

als auch die

Einstel-

lungen
Die

fur erweiterte

PID-Regelstrukturen.

bereits

zufriedenstellende

Unterdriickung

von

Storungen

der

Zulaufmenge

wird durch den Einsatz einer

Storgrofienaufschaltung
Konzepte
bei

noch verbessert. Auch fur ihren Entwurf kdnnen ahnliche

verwendet werden. Ein Entwurf

von

Storgrossenaufschaltungen,

dem

gleichzeitig zwei Modelle der Rektifikationskolonne berucksichtigt


extremen

werden, welche die


zu

Kolonnenbelastungen wiedergeben, fuhrt hervorragenden Ergebnissen. Vergleichbar mit dem Regelungs-

entwurf konnen sowohl

Storgrossenaufschaltungen

mit der Struktur

18

von

Zustandsregelungen (durch Minimierung der H^-Norm) als auch


mit einfacher Struktur (durch Parameter-

Storgroflenaufschaltungen

optimierung im Zeitbereich) berechnet werden.


In der industriellen Praxis

tatige Regelungstechniker sind sich der Realisierung


von

Schwierigkeiten, die
dezentralen
Daher ist der
terten

mit der

Zustandsregelungen auf

ProzelJleitsystemen
von

verbunden sind, sicherlich bewufit.

Regelungsentwurf auf der Grundlage von PID- oder erweihoher

PID-Regelstrukturen

praktischer Relevanz. Meist

bleibt die mit solchen Strukturen erzielbare

Regelgiite hinter

der

von

Zustandsregelungen
entworfenen
schrittenen

deutlich zuriick.

In dieser Arbeit werden die

Zustandsregelungen Regelstrukturen,

und die

optimal eingestellten fortge-

PID-Regelstrukturen verglichen. Dabei zeigt sich, dafi auch


die

mit einfachen

entsprechenden unkonventionellen
eine

Regler-Einstellungen vorausgesetzt,
der
von

Regelqualitat erzielt wird,


Dieses
von

die

Zustandsregelungen
in dezentralen

nahekommt.

iiberraschende

Resultat erlaubt die einfache

Implementierung

erweiterten PID-

Regelstrukturen
Verlust
Die
an

ProzelJleitsystemen

ohne wesentlichen

Regelgiite.
eines

Erprobung

Regelungsentwurfs
an

auf der

Grundlage fortRektifikations
in der

geschrittener
kolonne

PID-Strukturen

der

industriellen

bestatigt die grofie Robustheit und die hohe Regelgiite Abschatzung


der

Praxis. Dabei zeigt eine


einem solchen

Wirtschaftlichkeit,

dafi der bei

Projekt notwendige

Aufwand mehr als

gerechtfertigt ist.

1.1 "Modern Control:

Why

Don't We Use It?"

19

Chapter

Introduction

1.1 "Modern Control:


"Modern Control:

Why

Don't We Use It?"


is the title of
a

Why Don't We Use It?"

paper written

by R.
based

K. Pearson in 1984 [1.4]. In the first section of that paper Pearson

states: "Advanced control


on

systems utilizing multivariable strategies

process models

can

outperform traditional designs


spite
of market forces

in broad

classes of
process ments

application. Yet,

in

demanding

better

performance and ample evidence showing that the improve


be achieved, the gap between

can

theory and practice

in the

industrial sector is not narrowing

appreciably."
changed. The modern control theo

Ten years later the situation has not


ries

provide
a

tools for

increasingly sophisticated robust, model-based controller design. The advantages of


the process control engineer with
over

these controllers
tuned

the PID control structures which


numerous

are

usually
use

on-line, have been shown in


than 90% of all control
a

publications.

Neverthe PID
can

less,
be

more

loops

in the process

industry

control, while only

few

applications

of the modern control theories

reported [1.10]. Therefore the mismatch between theory and practice


reasons

is still evident. Some of the

for this situation

are

discussed

below.

20

1 Introduction

Distributed Control Systems For


a

control engineer in the process


is
a

industry, process control in the first


perspective
is the installation and

place

hardware
a

problem.

His

configuration of
modern DCS
are

Distributed Control System (DCS) [1.1]. Even the


of modern state space controllers

often limited to PID and advanced PID control. For the

DCS,

an

implementation
coupling with

requires

either the

an

external computer
are

or

the programming of

software modules. Both ways

troublesome and
to this

university research pays little attention


robust controllers with fixed structures
a

expensive. The situation. The design of

(e.g.,

PID control structures) is

largely unexplored field.

Dynamic Models
Linear

dynamic models
design. However,
on

are
no

the

foundation

of

modern, robust
are

controller unit

general dynamic

models

available for

operations. For each plant linear dynamic models


based
are

must be devel
or on

oped,

either linearization of nonlinear models


often

system

identification methods. Both ways

expensive and

very time-

consuming ([1.5], [1.6]). Furthermore,


industry
a strongly factorily described by a single

most

plants

in

the process

show

nonlinear

dynamic

behavior, which is unsatis

linear model.

Economic benefits The economic benefits of

improved control tend


study by

to be

significantly
than
one

underestimated. A benchmark
use

ICI "indicated that the effective


more

of

improved

process control

technology could add


[1.2].

third to the worldwide ICI

Group's profits"

[1.1]. Another

study

shows

smaller, but still


Of
course

massive benefits

it is not necessary to

replace all PID-controllers by modern

advanced control structures. Most control

problems

in the process

industry

are

handled well with simple PID control. However, strongly

nonlinear or/and ill-conditioned

plants require advanced control

tech

niques for

high controller performance.

1.2

Scope

and

significance

of this thesis

21

1.2

Scope

and

significance
as a

of this thesis

1.2.1

Distillation
one

unit

operation example

Distillation is

of the most

widely used unit operations in the process


a

industry. In the simplest


two

case,

distillation column separates


stream

feed of

components into

top product
and
a

(with

high

fraction of the
a

low-boiling component)
tion of the flow rate and the feed

bottom

product
In
an

stream

(with

high frac

high-boiling component).

industrial setting, the feed

composition may vary within a wide range of oper

ating conditions.
This

separation
a

consumes a

huge
an

amount of energy. A minimization of

the energy consumption and

economic

optimal operation usually


impurities
in the

require (1)
streams

tight control of both product compositions (dual composi

tion control) and (2) often small fractions of

product
and

(high purity distillation). However, the strongly nonlinear high-purity

ill-conditioned behavior makes


to control. Therefore

high-purity distillation columns difficult


distillation columns have become
an

interesting
1.2.2

test

case

for robust control

design methods.

Earlier research

Without any doubt the distillation process is the most studied unit oper
ation in terms of control.

Skogestad

estimates that

new

papers in this

field appear at

rate of at least 50 each year


a

[1.7]. It is

practically

impossible
Waller

to

give

review of all these

publications.

The interested

reader is advised to consult the reviews of Tolliver and

Waggoner [1.8],
review of

[1.9], MacAvoy and Wang [1.3], and the

recent

Skogestad [1.7].
If
we

focus
we

our

interest

on

the

design of linear,

time-invariant control

lers,

must state that all the well-known model-based and robust

control

design

methods

(LQG/LTR, H^, Normalized Coprime Factoriza


to

tion, u-synthesis, etc.) have been applied

distillation

columns.
one

However, all these

publications discuss the controller


a

design

for just

operating point. The problem designing

robust controller which maxi-

22

1 Introduction

mizes the controller been addressed


as

performance for the entire operating

range has not

yet.

1.2.3

Robust control and nonlinear

plants
or

The well-known robust control

design methods like HM -minimization


an

LQG/LTR
tured

are

based

on

the assumption of
at
one a

unstructured, frequency

dependent uncertainty
output,
or an

location in the

plant. Such
at

an

unstruc
or

uncertainty may be
additive

multiplicative uncertainty

plant input

uncertainty.
operating
range of
a

A controller

design
one

for the entire

distillation

column using lems:

of these well-known methods has two inherent

prob

Due to the

high nonlinearities uncertainty bounds will lead to


any

an

estimation of unstructured

very

large bounds, prohibiting

acceptable

controller

design.
any arbitrary, smaller uncertainty performance (RP) and robust stability not for the entire oper

A controller

design using

bound guarantees robust

(RS) for the actual operating point, but

ating
1.2.4

range.

Contributions of this thesis


a new

This thesis presents of a


on a

approach for the composition control design

binary distillation column (Figure 1.1). The design concept is based


structured

uncertainty model which describes the column dynamics


operating
range

for the entire

quite well. The resulting controller


or

designs using u-synthesis (for state-space controller)

u-optimization

(for controllers with fixed structure), respectively, lead to results which

guarantee robust performance and robust stability for the entire oper

ating

range of the distillation column.

optimal tuning
shown that with these

of

Special emphasis is placed on the easy-to-realize PID-control structures. It will be


can

extraordinary controller performance relatively simple controller structures.

be achieved

even

1.3 Structure of the dissertation

23

Standard

approaches

Improved approach

Uncertainty
Linear model for
a

model

single operating point

describing
for entire

column

dynamics
range

operating

Robust control

design

(i-synthesis
(X-optimization

IL LQG/LTR,

Weak

point:

Advantage:
RS & RP for whole

guaranteed

operating range

Figure

1.1: Robust control

design approaches

1.3
A

Structure of the dissertation


a

robust, model-based controller design for

distillation column

consists of several steps. A

typical

course

is illustrated in

Figure

1.2.

The results and methods of each step influence all the The consideration of just
one

following steps.

of these

design steps, disengaged from all

others, neglects the conceptional coherence. Therefore all of the design

steps

are

discussed within this thesis. The sequence orients itself to the


of the controller design.

natural

course

24

1 Introduction

Nonlinear Model

Uncertainty

structure

Controller

synthesis

Nonlinear simulations

Tests

on

plant

Implementation

in DCS

Figure

1.2:

Steps of a model based controller design

The

following chapter consists of three parts: The first part describes the design and operating data of the distillation column, followed by an
overview of the

steady-state and dynamic

column behavior. The second

for this part discusses the control objectives and control configuration

column, while the third part describes the

use

of pressure-compensated

temperatures

as

controlled outputs.
are

Rigorous nonlinear dynamic models


and for linearization.

the basis for simulation studies

They

are

discussed in

Chapter

3.

1.3 Structure of the dissertation

25

The main

subject of Chapter
are

4 is the derivation of linear models. Two


to linear models which

different methods

presented which lead

neglect

and include flow

dynamics, respectively.

A structured

uncertainty model which describes the nonhnear behavior

of the distillation column for the entire

operating

range is

developed

in

Chapter
Based

5.

on

that structured

uncertainty model, controllers

can

be

designed

within the framework of the structured


of Chapter 6 the theoretical

singular values. In the first part

background of the structured singular value

\i is summarized. While the second part of that

chapter presents

the

u-

optimal design of state-space controllers, the third part


the

is dedicated to

u-optimal design

of PID control structures. Simulation studies

confirm the theoretical results.

In

Chapter 7 the results of the (i-optimal controller design

are

compared
on an

with results obtained

by

more common

design

methods, based

unstructured uncertainty

description.
a

Usually the feed

flow rate is

measured disturbance input to


can

distilla

tion column. Therefore, feedforward control

significantly improve

the compensation of feed flow disturbances, which is discussed in

Chapter

8.

A controller

design should yield


are

satisfactory control quality not only in


in

dynamic simulations but also in the real plant. The results of the prac
tical

implementation

presented

Chapter

9.

The conclusions

and the recommendation for further research in


this thesis.

Chapter

10

complete

The literature references and, if necessary, the

special

notations

are

given

at the end of each

chapter.

26

1 Introduction

1.4 References
[1.1]

Brisk, MX.: "Process Control: Theories and Profits," Preprints of


the 12th World matic Control,

Congress of the International Federation of Auto

Sydney, July 18-23, 7,


W.

241-250 (1993)
M. L. Brisk: "Ben

[1.2]

Marlin,

T.

E., J. D. Perkins, G.

Barton, and
a

efits from process control: results of

joint industry-university

study,"
[1.3]

J. Proc.

Cont, 1,

68-83

(1991)
of Recent Distillation
5-21

McAvoy,

T. J. and Y. H.

Wang, "Survey

Control Results," ISA

Transactions, 25,1,

(1986)

[1.4]

Pearson, R. K: "Modern Control: Why Don't We Use It?," InTech,

34, 47-49 (1984)


[1.5]

Schuler, H., F. Algower, and E. D. Gilles: "Chemical Process


Control: Present Status and Future Needs
ropean

The View from Eu

Industry," Proceedings
on

of the Fourth International Con

ference

Chemical Process Control, South Padre Island, Texas,

February 17-22, 29-52 (1991)


[1.6]

Schuler, H.: "Was behindert den praktischen Einsatz moderner

regelungstechnischer Methoden in der Prozess-Industrie," atp,


34, 3, 116-123 (1992)
[1.7]

Skogestad,
a

S.:

"Dynamics and Control

of Distillation Columns
on

Critical

Survey," Preprints of the 3rd IFAC Symposium

Dy
1-25

namics and Control

of Chemical Reactors, Distillation Columns


April 26-29, College Park, Maryland,

and Batch Processes,

(1992)
[1.8]

Tolliver, T. L. and R. C. Waggoner: "Distillation Column Control;


a

Review and

Perspective from the CPI," Advances


83-106 (1980)
on

in Instrumen

tation, 35, 1,
[1.9]

Waller, K. V.: "University Research


trol of Distillation: A

Dual

Composition

Con

Review", Chemical Process Control 2, Sea

Island, Georgia, January 18-23, 395-412 (1981)

1.4 References

27

[1.10] Yamamoto, S. and I. Hashimoto: "Present Status And Future


Needs: The View from Japanese Fourth International
South Padre

Industry," Proceedings of the


on

Conference

Chemical Process Control,


1-28 (1991)

Island, Texas, February 17-22,

28

1 Introduction

2.1 Introduction

29

Chapter

The Distillation Process


An Industrial

Example

2.1 Introduction
A distillation column is not
toaster process
or

just any mass-produced article such as a washing-machine. Each distillation column is a unique

unit, specially designed for the separation of


are

substance mixture.

particular Nevertheless, the thermodynamic principles and


a

basic dynamics

always the

same.

Therefore it is possible to demon


one

strate ideas for the controller

design by the example of

column

without extensive loss of generality.

First in this

chapter,

the

design

and

operating data of the industrial


a

distillation column

are

outlined, followed by

brief

description

of the

composition dynamics.

The further two sections outline the control


use

objectives,

the control structures, and the

of tray temperatures

as

controlled outputs. The literature references terminate the chapter.

2.2 Column

design

and

operation
an

The distillation column described in this thesis is

industrial

binary

distillation column. A

synopsis

of the most

important

data for this distil-

30

2 The Distillation Process

An Industrial

Example

lation column is

given

in Table 2.1. The distillation column


a

(Fig. 2.1)

is

equipped
from the

with 50 sieve trays,

total

condenser, and
on

steam-heated

reboiler. The subcooled feed F enters the column

tray 20 (counted
a

top) and for the greater part consists of

mixture of two
are

substances. Because of the small fraction of

impurities, these
a

neglected and the distillation column boiling component)


bottom

is considered to be
are

binary distil

lation column. The desired product compositions for the top

0.99 mol/mol (low

product

D and 0.015 mol/mol for the


are

product

B. As these
can

product purities
as a

relatively high,

this

distillation column

be classified

"high purity distillation

column."

Table 2.1:

Steady-state

data

Column data
No. of trays
50

Column diameter (m)


Feed tray

0.8
20

Murphree tray efficiency


Relative volatility
a

=0.4

1.61

Operating

data
0.99

Top composition x-q (mol/mol)


Bottom composition xg (mol/mol)
Feed composition xp (mol/mol) Feed flow rate F (mol/min)

0.015 0.7-0.9 20-46


60

Top
Nominal

pressure

(mbar)

operating point
0.8

Feed composition (mol/mol)

Feed flow rate (mol/min)


Reflux

33
65 104

L0 (mol/min)

Boilup V51 (mol/min)

2.2 Column

design and operation

31

Vacuum

Condenser

Top product (Distillate)

D,xD
Reflux accumulator Reflux

Feed

20

F,xp

47 48

49 50

Boilup

Bottom

product

Figure 2.1: The industrial distillation column

32

2 The Distillation Process

An Industrial

Example

Feed disturbances The distillation column is connected in series lation columns, which operate in of these two columns
are

following two other distil


streams

parallel. The bottom product by


a

buffered

tank and fed into the column

considered here. The level of the buffer tank is measured

periodically

(typical period:

hours) and the feed of the column is set to keep the

tank level within

specified bounds. Therefore,


In contrast to
a

the feed flow is varied not

continuously but stepwise.


columns cannot
tion.
cause a

that, the variations of the feed

composition are always smooth. Even

shutdown of one of the other two

sudden increase of the buffer tank's

composi

Top pressure control


The

boiling points of the entering substances


pressure.

are

high at standard

atmo

spheric
vacuum.

Because of

thermal

decomposition of the light

component
is

at

higher

temperatures, the column is


the

operated
a vacuum

under

Correspondingly,

cooling

water flow rate for the condenser

kept

constant and the

top

pressure is controlled

by

pump.

Top level control


The reflux accumulator level is controlled

by overflow. manipulated

Hence the

top
a

product flow
composition

rate D is not available

as

variable for

control system.

2.3
Let
e.g.,

Steady-state
us assume a one

behavior

composition control scheme with integrating behavior,


one

PI controller which controls the

lating
tions

the reflux and

which controls the bottom composition


in

top composition by manipu by


an

manipulating the boilup. Then,


are

steady-state, the product composi

kept perfectly

profile

is

S-shaped composition developed within the distillation column. Figure 2.2 shows the
at their

set-points, and

simulated

composition profiles

for different feed flow rates and compo


are

sitions. While these

steady-state profiles

nearly independent of the

2.3

Steady-state behavior

33

1i

1ir

xp xp xp

0.7 mol/mol 0.8 mol/mol

0.9 mol/mol

20 mol/min

F F

33 mol/min
46 mol/min

0.0

0.2

0.4

0.6

0.8

1.0

Composition (mol/mol)
2.2: Simulated

Figure

composition profiles for the industrial distillation column

34

2 The Distillation Process

An Industrial

Example

feed flow rate, they


a

depend essentially
setpoints,

on

the feed

composition. This has

high significance for a controller design:


to their
we

If we want to

keep the product


variations in

compositions close
tion in the middle

must allow
we

profile

the middle of the column.

Consequently,

cannot control any

composi

of the column.
can

The internal flow rates

be illustrated in

similar

manner.

Figure

2.3 shows the simulated

liquid and

vapor flow rates for the nominal


as

operating
are

point. As

previously mentioned, the reflux


enter the column at
a

well

as

the feed

subcooled, i.e. they

temperature below the

boiling point.

A fraction of the vapor flow is condensed at the trays


are

where these two streams

fed into the distillation column. The two

discontinuities of the vapor flow

profile

at

trays 1/2 and 20/21 result

Liquid Vapor

flow

flow

Figure

2.3: Simulated vapor and

liquid

flow rates at nominal

operating point

60

80

100

120

Flow rate (mol/min)

2.4

Composition dynamics

35

therefrom. The

reason

for the

slopes of the

two

profiles within the strip

ping

and

rectifying

section of the column is the different heat of evapo

ration of the two substances.

2.4
The

Composition dynamics
composition dynamics within
a

distillation column is

effectively

described

by

movements

and

shape

alterations of the composition


control the reboiler level of the
rate

profile.

In order to illustrate

this, let

us

distillation column
the reboiler heat

by
to

the bottom

product flow
a

B, and let

us

keep

duty
a

constant. The simulated

step

responses of the

composition profile
rate
are

5% increase and

5% decrease of the reflux flow 2.4 a) raises

shown

by Figure 2.4. An increase of the reflux (Fig.


in the column bottom.
moves

the fraction of the the

light component composition profile of the light component


than 30%

Consequently,
towards the
to

column
more

bottom, degrading the bottom product composition from 1.5%

impurity. The opposite effect is observed for (Fig.


2.4 b): The

decrease
moves

of the reflux flow rate

composition profile

towards the column top, which

improves the bottom product composi

tion and debases the top

product composition.
important properties of the composition

These

plots illustrate

two

dynamics:

Column

nonlinearity: The product compositions

are a

nonlinear

function of the

reflux, boilup, and the feed condition: A 5%-

increase of the reflux flow rate

improves the top product


5% decrease

compo
more

sition

by

0.007

mol/mol, but

degrades

it

by

than 0.2 mol/mol.

Strong interactions:

change of

reflux

or

boilup alters both

product compositions.
The interaction between both

product compositions and reflux and composition dynamics, usually

boilup
called

has

a severe

consequence for the

36

2 The Distillation Process

An Industrial

Example

0.0

0.2

0.4

0.6

0.8

1.0

0.0

0.2

0.4

0.6

0.8

1.0

Composition (mol/mol)
a)
Figure
2.4: Simulated

Composition (mol/mol)

b)
composition profiles (light component) duty, feed
flow rate and for
a

step change oi

the reflux. Reboiler heat

composition

are

kept

at their

nominal values (see Table 2.1)

a)

L0=1.05*L0>nom

b)

L0=0.95*L0inom

2.5 Control

objectives and configurations

37

Ill-conditioned behavior.

This is best
both

explained by another
exact

two

examples. If we like
we

to increase

product purities simultaneously,


an

have to increase reflux and


+26.5% and the

boilup by

quantity, for example the reflux by


2.5 a). This the

boilup by

+19%

(Figure

tion constant, but it

keeps composition profile's posi slowly intensifies the S-shape of profile. However a
size for the reflux

slightly smaller step


behavior

completely alters the dynamic


response is much vector

(Fig.

2.5 b): The

purity of the top product decreases, the purity


input
[L, V]T is

of the bottom

product increases, and the dynamic


an

faster. Therefore

exact direction of the


a a

required

in order to achieve
even

simultaneous increase of both

product

purities. Consequently,
[L, V]T may lead

small uncertainty of the input vector

to undesired results.

High condition numbers


(2.

K.

-.tq> <Jmi{G(jo)))
a

of the

plant model G indicate such

behavior.

2.5 Control

objectives and configurations

The control of distillation columns has three objectives [2.2]:

Control of the material balance Product

(inventory control)

quality control

Satisfaction of constraints

The first

objective includes the control of the vapor holdup (top pres sure), the reflux accumulator level, and the reboiler level. Generally, objectives
are

these control
The second

easily achieved by simple

PI controllers.

objective

is the most

related to the economic and


column.

important objective. It is strongly ecological optimal operation of a distillation product qualities minimizes the
energy

Tight

control of both

consumption and the


is not
a

amount of products
to

being off the specifications. It


to their

simple task

keep both product compositions close

38

2 The Distillation Process

An Industrial

Example

0.0

0.2

0.4

0.6

0.8

1.0

0.0

0.2

0.4

0.6

0.8

1.0

Composition (mol/mol)

Composition (mol/mol) b)

a)
Figure
2.5: Simulated

composition profiles (light component) for duty. The feed


is

step change of
(see

the reflux and the reboiler heat

kept

at nominal condition

Table 2.1).

a)

Lo=1.265*L0>nora

b)

L0=1.260*L0,nom

V51=1.19*V51inom

V51=1.19*V51>noln

2.5 Control

objectives

and

configurations

39

setpoints, especially in the presence of disturbances such


of feed flow rate and feed composition.

as

variations control

Tight composition
main

requires sophisticated control schemes. Their design is the


this thesis.

topic of

Reflux, boilup, and predefined


range may range.
cause

pressure

drop

are

allowed to vary within

Any operation of a distillation column outside of this


insufficient separation
or even

damage

of the column.

Each control system must handle such constraints to enable safe opera
tion. This

topic

is well discussed

by Buckley

et al.

[2.2] and Shinskey

[2.4].

2.5.1 A

The 5x5 control

problem
as

simple distillation column, such


a

the industrial

example discussed objectives

here, presents

control

problem

with the five control

Top composition
Bottom

composition

Reflux accumulator level

Reboiler level

Top

pressure

and the five

manipulated variables

Reflux

Boilup (indirectly controlled by reboiler duty)

Top product flow rate


Bottom

product

flow rate
rate

Cooling water flow

(or vapor flow rate to vacuum)

This

problem

is often called the 5x5 control


pressure is controlled

problem. As

mentioned

above, the top


to

by

a vacuum

pump and the reflux

accumulator level
a

by overflow.

Thus the 5x5 control


are

problem

is reduced

3x3 control

problem. These relations

illustrated in Figure 2.6.

40

2 The Distillation Process

An Industrial

Example

Controlled outputs

Manipulated inputs

3x3 control problem

Top product
Bottom

xp xB

Reflux L

product

Boilup V (Reboiler duty Q)


Bottom

Reboiler level

Mb

product flow

rate B

Condenser level

MD

Top product flow


Overhead vapor

rate D

Top

pressure p

Vp
rate,

(Cooling

water flow

vacuum

pump)

5x5 control

problem
problems

Figure

2.6: The distillation control

2.5.2 In

Control the

design steps design of a MIMO controller


for the 5x5
or

principle,

in this

case

the 3x3 control

problem

does not
one

cause

any

particular difficulties.
disables all control
sensor or

However, the failure of just

actuator

or sensor

loops. Due
ator
are

to the

high sensitivity

of MIMO controllers to

actu

failure,

the inventory control and the

composition

control

usually

independently designed, thus improving the robustness of the control system and simplifying the controller design. The corresponding
design approach
1.

consists of three steps [2.5]:

Choosing the control configuration


a

In

first step the two

manipulated variables
names

for the

composition control configuration. For


L and the

are

to be chosen. This choice

the control

example,

if the top

composition

xrj is controlled

by reflux

2.6

Tray temperatures

as

controlled outputs

41

bottom

composition
variables

xjg is controlled

by boilup V, the control configura


control,
the

tion is called

L,V control configuration. After the choice of the manipu


for

lated

composition
are

remaining

three

manipulated variables

available for level and pressure control.


on

The choice of the control


selection methods such

configuration is often based


as

configuration

Relative Gain

Array (RGA), Niederiinski


[2.1], [2.6]), and the reli

Index,

or

Singular Value Decomposition (SVD). The application of these


to be low. One
reason

indices may lead to very different results (see

ability

seems

for the limited


et. al.

reliability may be the


to the substantial

neglect

of

inventory control:

Yang
on

[2.9] point

influence of

inventory control

the

composition control dynamics.


are

Most indices for control


state

steadygains. Consequently, perfect inventory control is assumed and


on

configuration selection

based

dynamic effects due


control
are

to the interaction of

inventory and composition

neglected.
control

The most

common

configuration in the chemical industry is the


an

L,V configuration [2.7]. This control structure is rather independent of

inventory control dynamics [2.9] and has shown good results within

experimental comparison
2.

of different control structures [2.8].

Inventory control design


general, tight inventory control
an can

In

be achieved with three


an

simple PI

controllers. Some distillation columns show reboiler level to with


3.

inverse response of the

increase of

boilup.

In this case, tight level control

boilup

as

manipulated

variable may be difficult.

Composition control design


composition control is
to be

A 2x2 controller for of the

designed
5-8.

as a

third step

design.

This step is discussed in

chapters

2.6

Tray temperatures

as

controlled outputs

On-line composition

analyzers are frequently used to determine product compositions. However, their investment and maintenance costs are

42

2 The Distillation Process

An Industrial

Example

prohibitive for distillation columns below


substances with
and that the
a

certain size. Provided that


are

boiling point difference


temperatures

of at least 10 C
are

separated

product purity specifications


([2.2], [2.4]).

not

extremely stringent,
composition

pressure-compensated
measurements

may

substitute

2.6.1

Pressure-compensated temperatures binary


mixtures
a

For

definite correlation exists between

boiling

temperature, pressure, and composition


T
=

f(p,x)

(2.1)
2.7 for the two components

This correlation is illustrated in

Figure

entering
sition

the industrial distillation column. A substitution of the compo

measurements

by

temperature

measurements

requires

compensation

for the effect of pressure variations.


are

If the pressure variations be

small, the temperature


N.

measurement

can

compensated by
are

linear function. The nominal pressure and compo

sition

denoted

by the index
=

Compensated

T +

gp

(P-PN)
N

(2.2)

In

case

of

larger

pressure

variations,

second-order term has to be

supplied:
T

Compensated

Qp

N<p-PN>+5aprT

(p-pN)2
N

(2.3)

Estimation
It is

of tray composition
to infer the

possible

tray composition directly. By regression of


a

{x, T, p}

data, the coefficients of

simple polynomial expression

can

be

calculated. An

example
x
=

is

given by
+

e]

Q2(T: + TCon)+e3p

Q4p2

(2.4)

2.6

Tray temperatures

as

controlled outputs

43

Figure

2.7:

Boiling points of the two-component-mixture

44

2 The Distillation Process

An Industrial

Example

Such

an

equation
to

in terms of the absolute

temperature and

pressure is

simpler

implement

in

distributed control system than

an

equation

in terms of deviations from reference values

e1

e2(T-TN)+e3(p-pN)+e4(p-pN)2
estimate is
a

(2.5)

One

problem of the tray composition


expected due

potential bias of the


a

temperature

measurements. Practical

experience has shown that

bias

of up to 2 C is to be

to heat

transport phenomena. In (2.4)

the bias is corrected

by

the parameter

TCoTT

In

practice, however, this


it would be

correction is difficult to estimate. In

principle,

possible

to

include
errors

cross

terms such

as

0Tp

in the

regression

model. However,

in the absolute
cross

temperature

may lead to incorrect numerical


cross

values of these

terms.

Therefore, in the regression model,

terms should be avoided.

Pressure compensation

as

well

as

the estimation of tray

composition

are

easily implemented in a process control system. Without a pressure compensation, it is impossible to use tray temperatures in a vacuum
column
as

controlled variables and expensive composition

analyzers

are a

necessary. For

temperature

measurements close to the column

top,

linear eompensation is usually sufficient. For trays close to the column

bottom,

we a

have to

expect higher

pressure

variations, and

compensa

tion with

second-order

polynomial

is recommended.

2.6.2 The

Temperature

measurement

placement
near

sensitivity of the tray temperatures


of the

the ends of the column to

changes

product compositions

is very small. To make the temper it has to be located at


some

ature measurement sensitive

enough,

distance from the column ends. Figure 2.2 shows simulated steady-state

composition profiles for the industrial distillation column. These profiles


illustrate the fact that the effect of increases with
a

change of operating conditions product composition

growing distance from the column ends. On the other


results from
an

hand,

deterioration of the correlation between

and tray

temperature

increasing distance from the

2.7 References

45

column ends. A
tion and

compromise between correlation with product composi


must thus be found. Eister discusses the most

sensitivity

important
In the

rules and tools in [2.3].


of the industrial distillation
are

case

column, the temperatures

on

trays
tray,

10 and 44
on

chosen

as

controlled outputs. Additionally, the


to any

temperature
it is

tray 24 is measured. Since tray 24 is close to the feed

expected to be sensitive

change of feed composition and,

dynamically,

to the feed flow rate.

2.7 References
[2.1]

Ariburnu, D., C. Ozge, and T. Gurkan: "Selection of the Best Control Configuration for an Industrial Distillation Column,"
Preprints of 3rd IFAC Symposium
on

Dynamics and Control of

Chemical Reactors, Distillation Columns and Batch Processes,

April 26-29,1992, College Park, MD,


[2.2]

387-392 (1992)

Buckley,
Research

P.

S., W. L. Luyben, and J. P. Shunta: Design of Distil Systems,


Instrument

lation Column Control

Society of America,

Triangle Park,

NC (1985)

[2.3]

Kister, H. Z., Distillation Operation, McGraw-Hill, New York


(1990)

[2.4]

Shinskey,

F.

G., Distillation control for productivity and energy

conservation, 2. ed., McGraw Hill, New York (1984)


[2.5]

Skogestad, S., and


for Distillation

M. Morari: "Control

Configuration

Selection

Control," AIChE J., 33,10,1620-1635 (1987)


E. W. Jacobsen:

[2.6]

Skogestad, S., P. Lundstrbm, and


Best Distillation Control
764 (1990)

"Selecting

the

Configuration,"

AIChE J., 36, 5, 753-

[2.7]

Skogestad, S.: "Dynamics and Control of Distillation Columns


A Critical

Survey," 3rd IFAC Symposium

on

Dynamics and Con-

46

2 The Distillation Process

An Industrial

Example

trol

of Chemical Reactors, Distillation Columns and Batch Pro April 26-29, 1992, College Park, MD,
1-25 (1992)

cesses,

[2.8]

Waller, K. V., D. H. Finnerman, P. M. Sandelin, K. E. Haggblom,


and S. E. Gustafsson, "An

Experimental Comparison

of Four

Control Structures for Two-Point Control of Distillation," Ind.

Eng. Chem. Res., 27, 624-630 (1988)


[2.9]

Yang, D. R., D. E. Seborg, and D. A. MeUichamp: "The Influence


of

Inventory Control Dynamics

on

Distillation

Composition Con

trol," Preprints of the 12th World Congress of the International


Federation

of Automatic Control, Sydney, 18-23 July 1993,1,

71-

76(1993)

3.1 Introduction

47

Chapter
A

Rigorous Dynamic Model

of

Distillation Columns

3.1 Introduction
The rigorous

dynamic

process simulation has become


even more so

an

accepted and
design [3.11].

widespread
Increasing
require
same
an

tool in process and

in controller

competition

and

environmental

optimization
a

of process and control

protection provisions structures, which can be

obtained

only by

substantial

knowledge
on a

of process

dynamics. At the
are can

time, dynamic experiments

running plant

less and less

desired.

Rigorous dynamic modelling

and simulation

replace long

such

expensive and time-consuming measurements. This has special signifi


cance

for

high-purity distillation
are

columns. Due to their

time

constants and

varying feed flow rates and feed compositions,

reproduc
new

ible operating conditions controllers


are

difficult to guarantee. Therefore,

usually tested thoroughly by dynamic simulation reality


to

for the

full operating range of the distillation column. The rigorous models of

distillation columns used for that purpose match the


extent

large

[3.17].

In this

chapter,

rigorous dynamic model for distillation columns

is

discussed. This model is used in all nonlinear dynamic simulations

48

3 A

Rigorous Dynamic Model

of Distillation Columns

within this thesis. In

special section, the numerical

treatment of the

resulting algebraic-differential equations is outlined. The modelling and control fields use very different notations. Therefore the
system of
notation used within this

chapter

is

explained

in section 3.11.

3.2 Conventions

Figure

3.1 shows
nt

schematic

representation of

distillation column
accumu

equipped with

trays. The column top (condenser and reflux

lator) is denoted by the index 0, the trays with the indices 1, 2,... nt, and
the column bottom

(including the reboiler) with the

index nt+1. To

simplify

the formal mathematical


as

description

the reflux stream R is

designated

liquid flow (L0).


as

The feed of the industrial distillation column,


is in

described in

Chapter 2,
by
a

liquid phase and subcooled. The top


pump and the condenser is

pressure is controlled with


a

vacuum

operated

constant
as

cooling
well
as

water flow rate. Flash calculations for the feed stream

dynamic models for the top

pressure of the column

are

therefore not
is

considered here. For other applications, the model extended with appropriate model equations.

presented

easily

3.3 The
The control

objective
or

of

modelling
(e.g., tray temperatures, product

process

engineer is interested in the dynamic behavior of

various

important

process variables

compositions) as a function of the time-varying column inputs. The objective of a dynamic model is an approximation of the real process input/output behavior by a system of differential and algebraic equa
tions. These model equations
as

are

based

on

material and energy balances

well

as on

thermodynamic

and fluid

dynamic correlations.

3.4

Simplifying assumptions
a

Within

distillation column many different


be

physical phenomena

occur.
on

Although it would

possible

to include models for the fluid streams

3.4

Simplifying assumptions

49

Qo
1

Condenser

Reflux accumulator

R
.1.
(=L0)

.2. .3.
4

Si,

V;

nt-2

nt-2

5v,nt-l

nt.:!
nt

Vnt+1

Qnt+1
B

Reboiler

&
nt+1

Figure

3.1: Distillation column

50

3 A

Rigorous Dynamic Model of Distillation

Columns

from

the trays, for the dead time caused by the transport time of vapor flow one tray to the next one above, or for the heat exchange with the
very high order. As earlier, the aim of modelling the distillation column
a

environment, the resulting model would be of


mentioned

dynamics
This

is

sufficient description of the real macroscopic behavior.


we are

means

that

interested

primarily in the dynamics of tray


no

compositions, temperatures, and


streams
ment
on

pressures etc. rather than in the fluid

the trays.

Experience shows that


the
a

substantial improve
with
more
are

can

be achieved with models

including effects

micro

scopic characteristics. Hence


introduced in order to achieve and order ([3.3], [3.13], [3.17]):

following assumptions

usually

compromise between model

accuracy

The

holdup of
of the

the vapor

phase is negligible compared

to the

holdup

liquid phase.
vapor

Liquid phase and

phase

are

each well mixed


vapor

on

all trays,
are

i.e., the composition of the liquid and of the

phase

inde

pendent of the position

on

the tray.

The residence time of the

liquid

in the downcomer is
a

neglected.
on

The variation of the


all trays. (This

liquid enthalpy on
is not

tray can be neglected


to the

assumption

applicable

evaporator.)
for all

In the literature

so

far, uniform liquid flows and

constant

holdups

trays have often been assumed (equimolar overflow). This assumption


is

problematic

because it
may

dynamic effects

implies a neglect of flow dynamics. Essential remain unmodelled, e.g., a non-minimum phase
an

behavior (inverse response) of the reboiler level and the tray composi
tions in the lower section of the column to
increase in reboiler heat

supply.

3.5 Balance

equations

51

3.5 Balance

equations

3.5.1

Material balances

The differential

equations describing the dynamics of the holdup for


on a

each component

tray

are

derived from

material balance for each


and

component. The balance border is the

single tray with its ingoing

outgoing

streams

(Figure 3.2).

Figure

3.2: Balance border for the material balances

Material balance for component k

on

trayj (k=l,

...,

nc;j=l,

...,

nt)

dnVi
"dT
=

d(n-xt-)

dT1^

pixF,kj +Vi*kj-i- (VSy)^


+

(3.

(Vj 1-SVij 1)yk)j 1-Vjyk>j


+ + +

52

3 A

Rigorous Dynamic Model of Distillation Columns

In the

same

way, the balance


are

equations for the column top and the

column bottom

formulated:

Material balance

for component k in condenser (k=l,

...,

nc)

dnk0
dt

d(n0xk0)
dt

(Vi-Sv,.)yk,i-(Lo + D)xk,o
...,

(3.2)

Material balance

for component k in the evaporator (k=l,

nc)
are

Usually the liquid phases


mixed either

in the column bottom and the reboiler


or

by

natural convection

by

pump.

Assuming perfect

mixing

we

obtain

dnk,nt+1
dt

d(nnt+lxk,nt+l>
dt

(3.3)
~~

*-'ntXk, nt "Xk, nt
~

%t

^k,

nt

The total holdup

on

tray j equals the

sum

of the

holdups of the indi

vidual substances:
nc

nj

X
k= 1

nk,

(3.4)

3.5.2

Energy

balance
a

equations
distillation column
same are

The vapor flows within

calculated

by

an

energy

balance. The balance border is the


which
was

to the border in

Figure 3.2,

used for the material balance equations.

Energy balance for tray j:

SW=F^ V.hH+(VJ -Sv,] ,)h"]


+
+
+

(35)

-(S^ L^-V^
+

For the left-hand side of this equation the

following

holds

3.5 Balance

equations

53

dni

dh'i

rt(nih'i)=h'jdF+nniF
If in

(3-6)

(3.6)

we

substitute the

expression for the differential

term dn-/dt

according

to

^
the

VLj-i+vj+i-svj+i-si,rLrvi
equation holds

following
A "h'

energy balance

W=
Usually,
or a

tFi<hVj-h'P+LJ-i^j_1-h-j)
+

(3g)

(Vj 1-SVfj 1)(h"j 1-h'j)-Vj(hj-h'j)]


+
+

the

assumption

n-

(dh'./dt)

0 is

permissible, except for cases


a

with large temperature variations

on

the trays,

large heat of mixing,


rewrite

large tray holdup.


as an

With this

assumption

we can

equation

(3.8)

algebraic expression for

the vapor flow rate V-

h-i^[Fi(hF,rhi)+Li-1(hj-1-hj)
equation
is formed for the
n-

(39)

A similar balance

evaporator. Because of the


cannot be

large inventory, the derivative


an

(dh'./dt)

neglected.

Since

increase in vapor flow


an

causes an

increase in bottom pressure and


in the

consequently

increase of

boiling temperature
in reboiler heat

evaporator, the
a

vapor flow follows any

change

supply

with

time

lag.

Hajdu et al. [3.9] present a model for this vapor flow lag. We
that
an

can

imagine

energy stream

Q supplied
causes

to the

evaporator is subdivided into


as a

two fractions: One

part

an

evaporation of liquid, the other


differential equation
we

increases the bottom temperature. Written obtain the energy balance equation

54

3 A

Rigorous Dynamic

Model of Distillation Columns

AQ
^

AH

,AV
, v, nt

, nt+.

dAT
+ n
nt
+

lVnt+lPnt+lCp,nt+l

nt+1

(3.10)

Jt

To achieve
term

first-order differential equation in has to be substituted by


a

AVnt+1,

the differential

dATnt+1/dt
a

differential term in
a

AVnt+1.

The increase of the pressure

drop

due to the

changing
can

vapor flow rate

(assuming

constant total

holdup

on

tray)

be estimated with

A(APj)

K J
+

AV.

(3.11)

Hence the pressure

change

in the evaporator

can

be

approximated for

distillation column with nt trays

by

A<Pnt+l)

=nt

UVj J
+

8APj

)
AV,
nt+1

(3.12)

The increase in bottom pressure

boiling point temperature


can

caused
to

by the increase

in

be calculated

according

AT

nt+1

3Pnt+l
Mi,
nt
+

A(Pnt+l)
1

(3.13)

Substituting (3.13)
tion is obtained:

in

equation (3.10), the following differential

equa

AQ-AHVjnt+1AVnt+1
^nt+l

nnt+lVnt+lPnt+lCp,

nt+1
V

3p

nt

r9APj^dAVnt UVi
+

Hnt+17

(3.14)

dt

Therefore, the
be described

vapor flow lag at an by the first-order lag

increase in reboiler heat

supply

can

3.6 Fluid

dynamics

55

^-g
with the time constant

T3(Q-Qlag)
lag

(3.15)

nnt+lVnt+lPnt+lcp,nt+l 9Pnt+l
lag

nt

UVj J
+

AH^
holdup balance equation in

(3'16)

If we substitute the total bottom

the energy

balance

equation
dn+,
,
=

"nt+l-ir1
the

Lnth'nt + Qlag-Bh'nt+I-Vnt+1h"nt+1

(3.17)

following equation holds:

Energy balance for the evaporator


V Vnt+1

Lnt(nnt-h'nt+l>
_

^lag

V5

Iv
n

1Sx

"nt+1

nt

The parameters

(e>Tnt+))/(9pnt+1)
or

and

(3Ap)/(3V-

1)

canbeeval-

uated

numerically

analytically from the appropriate equations (see

sections 3.6.2 and 3.7.2)

3.6 Fluid
In the

dynamics
as

previous sections, the equations describing composition and total


well
as

holdup dynamics,
discussed.

the vapor flow rates have been derived.


flow rates and of the pressure

Here the calculation of the

liquid

drop is

3.6.1

Liquid flow

rates

The volumetric

liquid

flow rate

over

the weir

on

tray j

can

be calculated

according

to the Francis weir formula

([3.16], [3.10]):

56

3 A

Rigorous Dynamic Model

of Distillation Columns

LV;j
For

u^2i|bwh^
0.64

;j

(3.19)

sharp-edged
the

weirs

jo.

holds. Perfect

mixing

on

the trays,

including

liquid

in the

downcomers,

is assumed.
,

Nevertheless, if we

calculate the effective


to take the
comer
sum

liquid head

hLW

above the weir

edge,

we

have

liquid phase fraction ej and the liquid volume in the downinto account (Figure 3.3). The liquid level in the downpipe is the
liquid head
on

of the

the tray and of the

hydrostatic level due

to the

pressure

drop according

to

Hydrostatic liquid level

in downcomer

p- -p=

(3.20)

Pjg
liquid
on a

The

liquid head hL

of the pure

tray (without
on

vapor

phase

fraction) is equal to the total liquid volume

the tray n-v'- minus the

o o o o o
Pj
"LWJ

Pj-Pj -l
1

Pj*

thLJ

Figure

3.3:

Liquid

levels

on a

tray

3.6 Fluid

dynamics

57

liquid

volume
-

in
>

the

downcomer

due

to
area

pressure

drop

AB (P:

Pj

j) / (Pjg)

both divided by the total

AA + AB:

Vj

Pj-Pj-1,

Ki
For the

AA + AB
formula, we

(3.21)

application

of the Francis weir

have to evaluate the

liquid level of the


and second the

pure

liquid (liquid without vapor phase fraction).

For

that purpose, first the height of the two-phase layer is to be evaluated

liquid phase fraction j


level becomes

must be taken into account. The

effective

liquid

VjTi.W.j

Pi-Pi_L-i

-h

Pjg
+

AA AB

-jhw
(3.19),
we

(3.22)

Substituting (3.22)
volumetric

into the Francis weir formula

obtain the

liquid flow rate of the two-phase


by:

mixture. The flow rate from

tray j

in molal units is calculated

Vr

Pj-p'izi.

3/2

u-v^tv
Lj
=

pjg

AA + AB

jhw
(3.23)

In many industrial distillation the weir. For this case,


e=

columns, calming
on

zones

exist in front of
we

holds at the weir edge. Otherwise,

have to estimate the

liquid

phase fraction

the trays. The Stichlmair

correlation is well suited for that purpose [3.18].

3.6.2

Pressure

drop
a

A vapor flow

through

tray in
on

distillation column suffers

pressure

drop.

Its amount

depends

the vapor flow rate, the tray

holdup, and

58

3 A

Rigorous Dynamic

Model of Distillation Columns

the geometry of the tray.

Usually, the

pressure

drop

is assumed to

consist of three different parts

([3.7], [3.12]):

Dry pressure drop occurring at the flow through the tray without

liquid (Aptr

j)
pressure

Hydrostatic

drop due

to

liquid head and liquid density

(ApLJ)

Pressure

drop by bubble-forming due

to surface tension of liquid

(APa;i>
The pressure

drop by bubble-forming usually is insignificant and can be

neglected.
Dry pressure drop
With sufficient accuracy, the the

dry

pressure

drop

can

be

approximated by

following well-known expression:

AptrJ
The orifice coefficient
on

^(Re)^
either

Ao

(3.24)
J

(Re)
it
can

can

be evaluated

by

measurement

comparable trays, During

or

be estimated with

experimentally verified

correlations.

the

simulations, the following correlation for sieve

trays is used [3.19]:

Ap tr,j

1-

aaJ

0.211

Ao

(3.25)
;

Hydrostatic pressure drop


The

hydrostatic

pressure

drop results from the liquid head and the

liquid density.

We have to take the

liquid

volume in the downcomer into

account (see 3.6.1).

3.7 Phase

equilibrium

59

ApL,i'

A.+L
sum

p>
of the two parts

(3'26)

The total pressure drop consists of the

dry pressure

drop

and

hydrostatic

pressure
=

drop:
=

APj
3.7 Phase
All
or

Pj + i-Pj

Aptr>j + ApLj

(3.27)

equilibrium
have discussed in the

equations

we

previous

implicitly

interrelated with the vapor

sections are explicitly phase composition. In this vapor

section, the most important correlations concerning the compositions and boiling points are presented.
3.7.1 The

phase

Vapor phase composition


on

liquid

each tray and in the evaporator is at


can

boiling-point.

Phase
few
can

equilibrium
be obtained

thus

be assumed. At moderate pressures up to in the vapor flow

some

bar, the concentration of a substance

leaving tray j

according

to

yEquilibrium

M*Jx

Kk

.xk j

(3.28)

If the substance mixture exhibits ideal


y becomes one, and the vapor

behavior, the activity coefficient


are

phase compositions

equal

to the ratios
on

of the

partial

pressures of the substances and the absolute pressure

the tray.

The vapor pressures of the pure substances pk can be calculated with a high level of accuracy by the Antoine equation (3.29). The parameters A,

B, and C

are

listed in many tables of substance properties (e.g., [3.5]).

^M^tTC

(3"29)

60

3 A

Rigorous Dynamic

Model of Distillation Columns

The calculation of the

effected

by

one

of the

liquid phase activity coefficients yk can be well known correlations (Wilson, NRTL,
.

UNIQUAC etc.). Murphree tray efficiency


In
a

distillation column

only little liquid

contact time exists

on

each tray for


no

the

mass

transfer between
can

and vapor

phase.

Therefore

perfect

phase equilibrium
efficiency for the

be achieved, and the tray


can

efficiency

will deviate

from the unit value. This effect


vapor

be modelled

by the Murphree tray

phase.

-^Equilibrium ,. yk,j ~yk,j

3.7.2

Boiling points

The vapor

phase composition according to (3.28)

is

function of the tray

temperature

Tj.

At

boiling point, the


one.

sum

of the vapor
a

phase mole

frac

tions calculated becomes

Hence for

tray j, the following boiling

point equation holds:

X
k=l

yEquilibrium

^ ^ p., ^ ^
.

(3.31)

k=l

The

Murphree tray efficiency is


to the

not considered for the


mass

boiling point

calculation, because

it relates to the

transfer between vapor and

liquid phase rather than

equilibrium composition.

3.8 Volumetric

properties
are

The fluid dynamic models discussed

interrelated with the molar


with the
corre

volumes of the vapor

phase and of the liquid phase, and


subject

sponding

densities. Their calculation is the

of this section.

3.8 Volumetric

properties

61

3.8.1

PVT relations

The molar volumes of the


on
or

liquid phase

v'-

or

the vapor

phase v". depends

the pressure pj, the temperature Tj, and the actual compositions x^j A great number of different equations of state has been developed ykj.

to describe this behavior.

They

[3.6]), and

discussion of their

are extremely well documented ([3.5], properties is not repeated here.

The PVT behavior is described here tion (SRK correction

by

the

Soave-Redlich-Kwong equa

equation, [3.15], [3.6]) with the Peneloux correction. This improves the estimate of the molar volumes of the liquid
is overrated

phase, which

by 10-15% using the SRK equation.

If measurement data of the PVT behavior of the pure substances exist

and their

mixing behavior is nearly ideal,


results for the

different

possibility has

shown We
a

good

liquid phase:
by

can

correlate the molar volumes measured with the temperature

polynomial regression.
be

The molar volume v'- of the substance mixture


a

can

approximated

as

weighted

sum

of the individual molar

volumes:
nc

v'j
3.8.2

I xk,/k,j
k=l

(3-32)

Density phase
can

The densities of liquid and vapor

be

computed from the molar

volume, the molar

mass, and the mole fractions.


nc

I xk,jMk
Liquid phase density:
o'=

k= ]
,

(3.33)

nc

I yk>JM k
Vapor phase density: p"

(3.34)

Vj

62

3 A

Rigorous Dynamic Model

of Distillation Columns

3.9
The
in

Enthalpies
quantity
or

not discussed

so

far is the

enthalpy of a
a

substance mixture
is estimated

liquid
of

vapor

phase.

The

enthalpy of a real fluid enthalpy from


the

by the

sum

an

ideal part and the value of


of the

departure

function

Ah^ap

describing the deviation


gas state:

enthalpy of the ideal

h +

j cjfdT Ahp
+

(3.35)

The ideal part

can

be calculated

by summing the ideal parts for each

component:
(
T

KddT= I xkHdkdT
T
k=l
Tn

(3.36)

"_iV

*0

The ideal heat

capacities

are

often

approximated by

third-order

polynomial

for each component:

cj,dk
The parameters for
substance

Ak + BkT + CkT2 + DkT3


are

(3.37)

equation (3.37)
or

listed in many tables of

properties,

they
p.

can

be estimated with very

high

accuracy

by Joback's method ([3.15],


The real part
behavior. It

154-156).

Ah^ p

describes the

departure of a mixture from the ideal


one

can

be evaluated

using

of the well-known

equations of

state, e.g., the SRK equation ([3.15], [3.6]).


If measurement data for the heat capacities and for the heat of vapor

ization

are

available,

simple solution

is

possible

in

a manner

similar

to that mentioned in section 3.8.1:

3.10 Numerical solution

63

T,

Liquid phase enthalpy:

ti

V
k=ll

k,JcP,J,kdT
Tft

(3.38)

Vapor phase enthalpy: h".

V Yk
k
=

h (3.39)

VTn

3.10 Numerical solution


The

complete rigorous dynamic model

for distillation

columns,

as

intro

duced above, consists of a system of differential and


trates the interconnection of the model

algebraic equations

(DAE). The complexity of the model is illustrated by Figure 3.4. It illus

equations for three adjoining


on an

trays. The solution of the differential equations obviously depends


the solution of the numerical

algebraic equation system. Therefore

efficient

integration using standard integration methods is not possible. This requires special adapted integration algorithms, as
outlined in section 3.10.4.

3.10.1 The

dependent

variables and the

equation system
we

As

first step for the numerical treatment,

have to decide which


vector

variables should form the vector of the

dependent variables. This

of
of

dependent variables must at any time


a

completely

describe the state


exces

distillation column and should be of minimum size to avoid

sive

computation

times.

The vapor

phase composition is

an

illustrative example for the complete


we

description
composition

of the distillation's state: If

know the tray composition,

the tray temperature, and the tray pressure, then the vapor
in

phase phase

equilibrium

is

easily calculated by

an

explicit algebraic

equation. Consequently, composition

it is not necessary to insert the vapor

into the vector of the

dependent variables.

CO

3 crc?

3.10 Numerical solution

65

As

one

vector which satisfies the

requirements of a complete description


vector is

and of minimum
tion of the vector

order, the

following

proposed (as

modifica

proposed by Holland & Liapis [3.10]):


>

[QlCond> D> nl,0

nnc,0> T0> P0> L0>

(Vj,

nxj,..., nncJ,

(Sy), (SvJ), Tj,


...,

pj(

Lj}j=1> 2>..., nt

Qlag

Q> Vnt+1, nlnt+1,

nncnt+1,

B, Tnt+1, pnt+i,
States of the control system]

(): Value is inserted only if it physically exists


The Jacobian matrix of the equation system (as described below)
corre

sponding
block

to these

dependent variables has

numerically advantageous

diagonal

dominant structure.

For the calculation of these


tions
are

dependent variables

y, the

following

equa

to be solved

Differential equations
nc

material balance

equations (3.1)

Algebraic equations
1 1

equation for equation for

vapor flow rate (3.9)

liquid flow
pressure

rate

(3.23)

1
1

equation for boiling point (3.31) equation for


nc

drop (3.27)

Total:

+4

equations per tray

and in addition the equations for the evaporator, the condenser, and the control system. often
ential

Considering industrial distillation columns which


with
more

are

equipped

than 50 trays, the

resulting algebraic

differ

equations

the industrial

equations. The model for binary distillation equipped with 50 trays gives an
consists of
a

amount to several hundred

impression of these numbers: It


and 210

system of 107 differential

algebraic equations.

66

3 A

Rigorous Dynamic

Model of Distillation Columns

3.10.2 Formal representation of the DAE We


can

formally represent the

entire

dynamic

model

by the

semi-

explicit equation system

^
0

f (t,

(t),

(t))

(t0)

n0

(3.40a)

g(t,n(t),z(t))

z(t0)=z0

(3.40b)

The vector
vector
z

consists of all tray

holdups (for all components), while


A different but

the

contains all other

dependent variables.
implicit form:

equiva

lent formal representation is the

F(t,y(t),y'(t))

=0

y(t0)

y0

(3.41)

Here the vector y contains all the


the dynamic behavior requires
a

dependent variables.

A simulation of

simultaneous solution of the whole

equation system.
3.10.3 The index The index of
izes the
a a

set of

differential-algebraic equations (DAE) character


more

integration problem. The higher the index, the

difficult is
defini

solution of the DAE. The

differential

index is the most

common

tion: The

differential index
m

of the system

F (t, y (t),
y

y' (t))

0 is the min

imal number

such that the system

ofF (t,

(t), y' (t)) =0 and of the

analytical differentiations d(F(t,y(t),y'(t)))


dt be
_

-U'""

dm(F(t,y (t), y'(t)))


dt into

can

transformed by algebraic manipulations

an

explicit ordinary

differential system [3.8].

Consequently,
m=0.

system of ordinary differential equation has

an

index of

3.10 Numerical solution

67

3.10.4 Solution methods and software The first

general method for the numerical solution of semi-explicit


was

DAE with index 1

proposed by C.

W Gear in 1971 [3.4] and


1

was soon

extended to the solution of


based
on a

implicit index

problems. The method


standard

is

special

class of the linear

multistep methods entitled the


are

backward differentiation formulas, which


the

algorithms for

integration

of stiff systems. The most important convergence results

may be found in

[3.1]. In theory, it is possible to solve problems of higher


neces

indices with the backward differentiation formulas. However, the


sary software is not available
as

yet. The apparently


are

very

frequently

used integrators DASSL and LSODI


methods
are

based

on

Gear's method. These

distinguished

for their effectiveness in

solving continuous

problems.

However, the

computational

effort grows

significantly

for

systems with discontinuities arising, for example, during the simulation of the response to several feed flow or feed composition step changes. For
such cases, the one-step methods find The one-step methods
are
more

and

more

interest [3.11].

extensions of the well-known

Runge-Kutta,

Rosenbrock,

or

extrapolation methods. An extensive discussion of the


is found in

properties of these methods


of the integrators mentations
are

[3.8]. However, the development

(RADATJ5, LIMEX) is

found in any of the

in an early stage, and no imple widespread Fortran libraries.

For the simulation studies the DASSL

integrator,
success. an

as

implemented in the
differential-alge according
to

NAG Fortran
braic

Library

is used with
are

good

The

equations (DAE)

solved in

implicit

manner

(3.41).
The calculation sequence

During the integration, the right-hand sides of the differential and alge
braic

equations repeatedly have

to be evaluated for

given

vector y of

the

dependent variables and for a given time


as

t. The
are

algebraic equations,
an

and often the differential equations


manner.

well

solved in

implicit integra

The equation errors, which have to be


the difference between the

supplied

to the

tion,

are

right-hand

sides of the

equations

68

3 A

Rigorous Dynamic Model of Distillation Columns

(that
the

means

the calculated vapor flow rates,


the vector y.

liquid flow rates, etc.) and

corresponding values within

A correct calculation sequence

evaluating these

terms is

stringent: If,

for

example,
(3.31),

we

calculate the

right-hand

tion

we

must know the vapor

boiling point equa phase composition. Therefore we


If this basic idea is

side of the

first have to calculate the vapor the


error

phase composition and subsequently applied


to the

of the

boiling point equation.

whole

model,

the calculation sequence illustrated in

Figure

3.5 results.

The vector y, which is


contains the component

supplied by the integration holdups


as

routine

(Step a),

tray

pressures,

as

well

liquid phase, liquid and vapor flow


in

the tray temperatures,


rates. With the data
can

supplied, (Step b).


are

all vapor
In
a

phase compositions
step, the

in

equilibrium
of the

be calculated

next

step (Step c), using the distribution coefficients


errors

Kk

obtained in the

previous

calculated. Since all the feed data


are

are

boiling point equations known, its enthalpy, molar


vapor

volumes and densities


sitions

computed

in

step d. The
vapor

phase

compo

deviate

from

the

equilibrium compositions. Applying the


phase compositions
are

Murphree tray efficiency, the effective

computed (Step e). Since for the computation of the effective vapor

phase composition for a tray the


next lower

effective vapor the

phase composition of the


starts at the column
are

tray

must be
n
=

known,

computation

bottom, assuming

for the reboiler. Now all necessary data

known to calculate the

enthapies,
are

molar volumes, and densities for all


g, the energy

trays, the condenser and the evaporator (Step f). In step


balance equations for the trays
the
tion routine
rator.

applied, resulting vapor flow rates and the flow rates supplied by the integra
are

and the differences between

calculated. In step h, the


to the
error

same

is done for the evapo


errors

Similarly

of the vapor flow rates, the


are

of the

liquid

flow rates and the tray pressures


rates and

computed

in

steps i and j.

Using the flow


terms

rather than the data

previous steps supplied by the integration routine, the differential


the substances in liquid phase (Step 1) are (Step m), all differential terms, the errors

compositions calculated

in the

(left-hand sides) of the equations describing the vapor flow lag


last step

(Step k) and the holdup of


calculated. In
a

3.10 Numerical solution

69

a)

Vector of dependent variables y

b)

Vapor phase composition for evaporator and trays (Equation (3.28))

Error for

boiling point
k

at

evaporator and trays

c)

Calculation of the

thermodynamic

states

d)

h', v', v", p', p"


for the feed

Murphree tray efficiency


e)

for trays nt, nt-1,..., 1

yk,J

<yk^"lbnum-yk,J+iJ+yk,]+,

Calculation of the thermodynamic states

hVhVv'rv>>"j
for condenser, all trays, and evaporator

Error for vapor flows

hVh*j

Figure

3.5: Calculation sequence


see

Explanation:

text

70

3 A

Rigorous Dynamic

Model of Distillation Columns

Error for vapor flow

leaving evaporator

h)

Lnt(hnt-hnt+l>+Qlag
h"
"nt+1

-h'

nt+1
"nt+1

Error for

liquid

streams
3/2

i)

13

p,g
A
+

B
,
J

AA+AB

If

Error for pressure

drop

j)

P]+1-P3-AP]
(Equation (3.27))

' '

Differential

k)

equation for vapor (Equation (3.15))

flow

lag

< '

1)

Differential equations for holdup of substances

(Equations (3.1), (3.2)

and (3.3))

'

'

m)

Vect or of differeiitials and

errors

Figure

3.5 continued

3.11

Notation

71

between
errors

supplied and calculated flow


boiling point equations
back to the integration routine.

rates and pressures, and the

of the

are

combined in

one

vector and

supplied

3.11

Notation

A0 AA
Ab

[m2] [m2] [m2]


[m]

Hole

area

in

tray
area

Tray

area

without downcomer
area

Downcomer

bw
pid LP

Length

of weir

[J/mol-K]

Ideal gas heat

capacity

CP,1

[J/kg-K]
[m]
[mol/s] [J/mol]

Liquid heat capacity


Diameter of holes of sieve tray
Feed flow rate to tray j Molar

do

Fj
h

enthalpy

h'j

[J/mol]
[J/mol]

Molar Molar

enthalpy of liquid phase

h"j
hL hw

enthalpy

of vapor

phase edge of weir

[m] [m]

Liquid level
Weir

above upper

height
on

AHv,k,j
AHvj

[J/mol]
[J/mol]

Heat of evaporation of component k Heat of evaporation of liquid


on

tray j

tray j
on

Kkj
LJ
Wj
Mk
nt

[mol/mol]
[mol/s]

Distribution coefficient for comp. k

tray j

Liquid flow leaving tray j


Volumetric flow from tray j
Molar
mass

[m3/s]
[g/mol]
H

of component k

Number of trays in column

72

3 A

Rigorous Dynamic

Model of Distillation Columns

nj

[mol]

Total

holdup

on

tray j
k
on

nkj
nc

[mol] [-]

Holdup of substance

tray j

Number of components
Pressure Pressure
on

Pj

[N/m2]

tray j
over

APj

[N/m2]
[N/m2] [N/m2]
[J/s] [J/s] [m]

drop

tray j

K
P

Steam pressure of pure component k


Pressure

Heat

supply to evaporator
supply

Qlag
s

"active" heat

Thickness of sieve tray Side

SU
Jvj

[mol/s]

product flow

rate from

tray j,

liquid phase [mol/s]


Side

product flow

rate from

tray j,

vapor
t

phase

[s] [K]

Time

Temperature

TJ
Vj

[K]

Temperature
Vapor

on

tray j tray j

[mol/s]

stream from

VVj
xkj

[m3/s]
[mol/mol]

Volumetric vapor stream from tray j

Liquid phase mole fraction


component k
on

of

tray j

XF,ko

[mol/mol]

Mole fraction of component k


in feed to

tray j

ykj

[mol/mol]

Vapor phase mole fraction of


component k above tray j

Yk

[-]

Liquid phase activity coefficient


of component k

3.11

Notation

j
Tl
V

[m3/m3]
[mol/mol]

Liquid phase fraction

on

tray j

Murphree tray efficiency for vapor phase


Molar volume Molar volume of liquid Molar volume of vapor
Orifice coefficient

[m3/mol] [m3/mol] [m3/moI]


H

V'j V"j
%

phase
phase

on

tray j tray j

on

P'j

[kg/m3]
[kg/m3]

Liquid density
Vapor density

on

tray j

P"j

on

tray j

74

3 A

Rigorous Dynamic

Model of Distillation Columns

3.12 References
[3.1]

Brenan, K. E., S. L. Campbell, and L. R. Petzold, Numerical


lution

so

of initial-value problems

in

differential-algebraic

equa

tions, North-Holland, New York (1989) [3.2]

Byrne, G. D., P.

R.

Ponzi, Differential-Algebraic Systems, Their

Application
(1988)

and Solution,

Comp. Chem. Eng., 12, 5, 377-382

[3.3]

Gani, R., C. A. Ruiz, and I. T. Cameron: "A Generalized Model for


Distillation Columns," Comp. Chem. Eng., 10, 3, 181-198 (1986)

[3.4]

Gear, C. W.: "Simultaneous Numerical Solution of Differential-

Algebraic Equations," IEEE


89-95 (1971)

Trans,

on

Circuit

Theory, CT-18, 1,

[3.5]

Gmehling, J. Collection;' 1,

and U. Onken: Part

"Vapor-Liquid Equilibrium

Data

1, XI-XXII, DECHEMA, Frankfurt (1977)


Ver-

[3.6]

Gmehling, J. and B. Kolbe: Thermodynamik, Georg Thieme

lag, Stuttgart (1988)


[3.7] Grassmann, P. and F. Widmer, Einfiihrung in die thermische

Verfahrenstechnik,
[3.8]

2nd ed., de

Gruyter, Berlin (1974)

Hairer, E. and G. Wanner: Solving Ordinary Differential Equa


tions II

Stiff and Differential-Algebraic Problems, Springer

Verlag, Berlin (1991)


[3.9]

Hajdu, H., A. Borus, and


tion

P. Foldes:

"Vapor Flow Lag


1-8 (1978)

in Distilla

Columns," Chem. Eng. Sc, 33,

[3.10] Holland, C. D. and A. I. Liapis, Computer Methods for Solving

Dynamic Separation Problems, Chapter 8, McGraw-Hill,


York(1983)

New

3.12

References

75

[3.11] Marquardt,

W.:

"Dynamic

Process

Simulation

Recent

Progress and Future Challenges," Fourth International Confer


ence

on

Chemical Process Control, South Padre Island, Texas

(1991) [3.12] McCabe, W. L., J. C. Smith, and P. Harriott: Unit Operations of


Chemical

Engineering,

4th

ed., McGraw-Hill, New York (1985)

[3.13] Najim, K. (Editor): Process Modeling and Control in Chemical

Engineering, Marcel Dekker, New York (1989), Chapter III,


211, S. Domenech, L. Pibouleau, "Distillation" [3.14] Petzold, L.: "Differential/Algebraic Equations
SIAMJ. Sci. Stat.
are

145-

not

ODE,"

Comput, 3, 3, 367-384 (1982)

[3.15] Reid, R. C, J. M. Prausnitz, and B. E. Poling: The Properties of


Gases and

Liquids,

4th ed., McGraw-Hill, New York (1988)


von

[3.16] Retzbach, B.: "Mathematische Modelle


nen zur

Destillationskolon-

Synthese

von

Regelungskonzepten," Fortschritt-Berichte

VDI, Reihe 8: Mess-, Steuerungs- und Regelungstechnik, Nr. 126,


VDI

Verlag (1986)

[3.17] Rovaglio, M., E. Ranzi, G. Biardi, and T. Faravelli: "Rigorous Dy


namics and Control of Continuous Distillation

Systems

Simu
871-

lation and
887 (1990)

Experimental Results," Comp. Chem. Eng., 14, 8,

[3.18] Stichlmair, J.: Grundlagen der Dimensionierung des GaslFliis-

sigkeit-Kontaktapparates Bodenkolonne, Verlag Chemie, Weinheim (1978)

[3.19] Weiss, S.

et.

al.:

Verfahrenstechnische Berechnungsmethoden,

Teil 2: "Thermisches
heim (1986)

Trennen", VCH Verlagsgesellschaft, Wein-

76

3 A

Rigorous Dynamic Model of Distillation Columns

4.1 Introduction

77

Chapter

Linear Models

4.1

Introduction
are

Robust controllers

designed

on

the basis of linear process models.

Therefore the elaboration of linear column is


a

dynamic

models for the distillation

central part of control system

should describe the

synthesis. These models dynamic behavior of the process within a wide


can

frequency

range.

They

be obtained in two ways:

System identification
Linearization of
a

nonlinear model
it avoids

It is

big advantage of the system identification that


drawbacks, for example:

cated and expensive nonlinear model. Nevertheless, this


some severe

compli approach has


a

The time-constants of the

composition dynamics

are

large.

recording

of

input/output

data for the real

plant

is very time-

consuming.

Due to the

high sensitivity

of distillation columns to
for small

changes

of

the internal flow rates,


variation

magnitudes input (e.g., 5% of the steady-state value) the response may


even

of the

far exceed the linear

region.

78

4 Linear Models

Each

experiment

causes

undesired disturbances of the

product

qualities.

It is

practically impossible

to obtain models for the entire oper

ating
These

range of the distillation column


some

disadvantages and
the

other fundamental
a

problems of the iden


strong recommenda
that
means

tification itself (see Jacobsen et al. [4.5]) lead to


tion

of

second

method

(Skogestad,

[4.12])

the

linearization of nonlinear column models. Two linear models


are

evaluated within this

chapter, which

are

based

on

the linearization of different nonlinear column models. The first linear


model
is

obtained

by

an

analytical linearization of

simplified
is obtained in

nonlinear model

neglecting flow dynamics. The other model


presented

by

numerical linearization of the rigorous model

Chapter

3. In further sections the accuracy of these linear models and the role of

the flow

dynamics
are

are

discussed. Different

mathematically order reduc

tion methods

compared
on

at the end of this

chapter.

The notation is
are

listed in section 4.9


in section 4.10.

page

101, the literature references

collected

4.2 How to linearize the


The state,

rigorous model?
vectors

4.2.1

input, and output

The

complete rigorous dynamic


a

model of

as

discussed in

Chapter

consists of

high-order system
the

coupled differential and algebraic large system would be possible


causes

equations
in

(DAE). A linearization of this


order
a

principle. However,
same as

resulting linear state-space model would be


a

of the

the DAE. Such

high order
for
a

high computa

tion times for

controller

design

or even

model reduction. Conse


accuracy must be
are

quently,
very

compromise between model order and


means we

sought. This

have to decide, which

dependent variables
are

important

for the composition


x

dynamics

and should be included in


the

the state vector

of the linear model. Most important, of course,


Because flow

tray compositions themselves.

dynamics have

high influ-

4.2 How to linearize the

rigorous model?

79

ence on

the
is
a

composition dynamics
candidate
as

in the

high-frequency range,
a

the

tray

holdup

reboiler and the reflux

Assuming accumulator, it is not necessary

well.

perfect level control

of the

to include their

holdup. The corresponding candidates for the


model
are

state vector of the linear

dx0 dxc
dx,
x
=

dx,

or

dx

51

(4.1)

dx

dn,
51

dn

50

The

dynamics of the distillation column


Most

are

stimulated

by the manipu

lated variables (reflux


sources.

L0 and boilup V51) and the several disturbance important disturbances are variations of the feed compo
as

sition xp and the feed flow rate F. Other disturbances such

variations

of the reflux temperature


influence and
we can

or

the feed temperature have


for the

significantly less design. Hence

be

neglected

composition control

define the input vector

according

to

dxF
d
u

dF

(4.2)

dL0 dv51

The output vector y follows


ments. It atures
on

directly

from the temperature

measure

represents the deviations of the pressure


tray 10, 44, and 24:

compensated temper

80

4 Linear Models

dTPio

dTP44 dTP24
4.2.2

(4.3)

Handling of the algebraic equation system


rigorous column model defines
flows, and liquid
x

The algebraic equation system of the

dependent

variables such
are

as

tray

pressures, vapor

flows, which
Nevertheless

not included in the state vector

of the linear model.

the

algebraic equation system represents algebraic

constraints for the composition and


can

holdup dynamics.

These

equations

be handled in two ways:

elimination

by idealizing assumptions,
during linearization
an

or

numerical solution

The first method allows

analytical linearization of the resulting


that

model. This has the


must be

advantage
can

merely

one

steady-state data
of the
common

set

supplied, which

be calculated

by

one

flow-

sheeting programs (e.g.,

PROCESS, ASPEN PLUS).


a

In contrast to

that, the second method requires

numerical linearization of the

rigorous model, which

is discussed in section 4.4.

4.3 Linearization of

simplified

nonlinear model

4.3.1

The

simplified model
we

In this section

will derive

simplified nonlinear column model dynamics.


For that purpose

neglecting
flow
the

the

holdup

and thus the flow


are

idealizing assumptions

formulated which allow to dispense with all


energy balance
a

dynamics and with most of the

subcooling

of reflux and feed have

equations. However, significant influence on the


account.

internal flow rates (see section 2.3) and

are

explicitly taken into

4.3 Linearization of

simplified nonlinear model

81

Idealizing assumptions
The

algebraic
can

constraints

of the

rigorous model and the holdup

dynamics

be eliminated

by

the

following idealizing assumptions:

constant pressure
constant and

drop
on

equal enthalpies

all trays

constant total

holdup

on

all trays

(equimolar overflow)

Of course, all these The first

assumptions do
means a

not agree with the real conditions.

assumption

neglect of the correlation between tray


rates. The second

pressures,

holdups, and boilup

assumption implies
tray holdup

uniform vapor flows within the

stripping section and within the recti assumption of


The
error
a

fying

section of the column. The


a

constant

contains

neglect

of flow

dynamics.

in the

high frequency

range introduced

by

that is discussed in section 4.5.

It has to be

emphasized here that these assumptions concern only the simplified nonlinear model as a basis for an analytical linearization.
The steady-state

operating points

must be calculated
as

using
as

model,

which includes the energy balance equations,

well

the flow

dynamic models.
The
For

composition dynamics
a

column separating

binary mixture,

it is sufficient to formulate

the material balance equation for the


mixture. If
we assume

light component of the substance

constant total

holdup,

the

following

material

balance equation holds for the tray j:

^
Similar balances
obtain the

Lj-i+Vj i+Fj-VVj
+

(4.4)

are

obtained for the reboiler and the condenser. Substi


in the material balances

tuting these balance equations

following

differential equations

(3.1)-(3.3), we describing the composition

dynamics:

82

4 Linear Models

Condenser

7>ays (Feed

is

liquid phase, j

1.. .50)

^j[V.(vi-v+Vi^-v
Evaporator

(4.6)

"aT

4[L50(x50-x5i)-V51(y51-x5])]
are

(4.7)

Effect of subcooled reflux and feed


Feed and reflux of the distillation column
vapor flow is condensed at the

subcooled. A portion of the


streams enter the

trays where these two

column. The effect of


caused

an

additional condensation of the vapor stream


rates of these streams must be considered

by increasing the flow

to avoid

large

model

errors.

The two energy balance

equations for

the

reflux tray

(tray 1) and for the feed tray (tray 20) become part of the

nonlinear model:
0

Loh'o-Ljh'j+Vjh'^-V^",

(4.8)

L19h'19~L20h'20 + V21h"2]-V20h"20 + Fh'F


as

(4.9)

Tray temperatures
The model outputs

model outputs the deviations of the pressure compensated tray


are

are

temperatures. These temperatures


sition

correlated with the tray compo

by

the

boiling point equation

4.3 Linearization of a

simplified

nonlinear model

83

X (y)
k=l

-1=0

(4.10)

Consequently
of the

the

boiling point equation for tray 10, 44, and

24

are

part

simplified

column model.

The vapor phase

composition composition dynamics


too.
are

The differential equations for the


the vapor

in terms of is

phase composition yj,


one

Usually,

the tray

efficiency
can

smaller than

and the vapor

phase compositions deviate from


Chapter
3, this
n

the be

equilibrium
modelled

compositions. As described in

by the Murphree tray efficiency

Yj
Primarily the
of the relative
vapor

O-TDyj^+Tiy'j
in

(4.11)

phase composition
x-.

equilibrium
if we

y*j

is

function
constant

liquid phase composition volatility a. on tray j,


by

For

example,

assume a

the vapor

phase composition

in

equilib

rium is calculated

ct X-

y*-

i-J

(4.12)

To

simplify the analytical linearization,

it is convenient to substitute the

y by equilibrium compositions y*. Then the calculation of the derivatives

vapor

phase composition

correlation exclusive in terms of the

3 (...) /dx for each these terms

cause no

particular problem.

Such

an

equation in

terms of the

equilibrium compositions y* is derived

by

subsequent substitutions of the vapor

phase compositions

in (4.11)

from the evaporator up to actual tray. For

example (n=l assumed for the

reboiler)

y50

(l-Ti)y*gi+r|y*50

(4.13)

84

4 Linear Models

y49

(l-Ti)y50 + 1iy*49

(i--n)V5i+Ti(1-1i)y*5o
(i-Ti)y49 + 1iy*48

7iy*49

(4.14)

y48

,A,E,

(l-*i)3y*5i+Ti(i-'ri)2y*5o

(4.15)

'rl(i-n)y*49 + riy*48

For

binary distillation column with

nt

trays the following generalized


in the

formula (with obtained:

vapor/liquid equilibrium (n=l)

evaporator)

is

y.

(l-Tl)at+1-ynt+1
nt+l-j
+

X lKl-M)nt+1-j-Vnt+,-n
n=

(4>16)

This equation demonstrates the strong influence

on

the composition of

the trays below the actual one, presuming the tray

efficiency

is

substantially
4.3.2

smaller than

one.

Analytical linearization

Let

us

formally represent

the

simplified nonlinear model

as

the vector

functions f and g:

d^l

f[x(t),u(t),d(t)]

(4.17a)

y(t)

g[x(t)]

(4.17b)

Then the matrices A, B, and C of the linear state space model

d^

Ax

Bd

(4.18a)

Cx

(4.18b)

4.3 Linearization of a

simplified nonlinear model

85

are

evaluated

as

partial

derivatives of the vector functions f and g at

steady-state operating point (OP). The OP


the

is calculated either solving equation system of the complete rigorous model for steady state, or

using a steady-state flowsheeting program. The following relations hold


for the

partial derivatives:

*0 <*0 dx0 dx,


A

dx51
(4.19)

3f
=

3x

3fj 3fj 3x0 dx.


OP
3f. 51 dx, 0

3x

51

IOP

"0
B 3f
=

3f0 df0 3f0 3xF 3F 3L0 3V51


3f, 3f, 3f, 3f, 3xF 3F 3L0 av51
(4.20)

d
u

OP

*5i 3f51 ^51 3f51 3xF 3F 3L0 IJ


9g] 3g,

[OP

3xQ 3xj

*i '3x51

L~di OP

3g2 3g2 3x0 3xj

3g? '^Sl
3g3 '^51

(4.21)

3g3 3g3 3xQ 3x,

IOP

86

4 Linear Models

Combining the idealizing assumptions, we


of reflux

can

conclude that

deviation

Lo

or

boilup V51

causes

the

same

deviation of the

liquid flow

and vapor flow rates within the whole column:


dL.
J

dLn

dVj

dv51
A, B, and C
are

(4.22)

The resulting coefficients of the matrices

listed in the

appendix
out

of this

chapter (page 97). Although important


are

interactions in
it turns

the column model

suppressed by

the

idealizing assumptions

that this model coincides

within

acceptable bounds with

the

rigorous nonlinear model. This aspect


4.5 below.

is discussed in detail in section

4.4 Linearization of the

rigorous

model

In this section the linearization of the


is discussed. This model includes the

rigorous dynamic column model


and

dynamics of the tray holdups

thus flow

dynamics.

Hence it describes the

much better than the

high-frequency dynamics simplified nonlinear model (section 4.3.1).

4.4.1

Model modifications the deviations of pressure

The desired outputs of the linear model

are

compensated tray temperatures,


sitions, but
ance

which

are

functions of the tray compo

not of

component's holdups. Therefore the material bal


are

equations (3.1)-(3.3)

replaced by the following, equivalent

differential equations: Condenser

^-iv.to,-^)

(4.23)

4.4 Linearization of the

rigorous model

87

Tray holdups:

S
dx.

Li-'-VV--VFi
is

(4-24)

Tray composition (Feed


i
=

liquid phase, j

1.. .50)

^[Lj_1(xj_1-xj)+Vj ^-ii*j-i-'y

1(yj

1-xj)

vj 1iyj 1-xj;
+ + + +

(42g)

-vj(yrxP+Fj(xF,rxP]
Evaporator

5^L (*-**> -V51(y51-x51)]


rigorous model follows from the

(4-26)

The second modification of the


tion of a

assump

perfect

level control in the reflux accumulator and in the evap

orator. This

assumption is justified by the fact that the level control


can

loops for

the condenser and the evaporator

be tuned much faster

than the composition control

loops.

With

perfect

level control and there

fore constant

holdup

for the reflux accumulator and the evaporator, the


streams
are

top and bottom product

calculated

according

to

V,-L0

(4.27)

and B
=

L50-V51

(4.28)

The

whole

equation system consists of the differential equations

describing
and the
pressure

the composition

dynamics for

holdup dynamics (4.24), and

light component (4.5)-(4.7) the algebraic equations for the


the

drop (3.27),
a

the vapor flows (3.9),

liquid flows (3.23), and

the

boiling

points (3.31). These algebraic equation form


constraint for the differential
can

whose solution is

an equation system equations. All these

equations

be

comprised by the following vector equations:

88

4 Linear Models

di|P
^P

f[x(t),n(t),u(t),z(t),v(x,n,u,z)]

(4.29a)

l[x(t),n(t),u(t),z(t),v(x,n,u,z)]

(4.29b)

y(t)
0
=

g[x(t)]
n, u,

(4.29c)

k[v(x,

z)]

(4.29d)

The vector

(x,

n, u,

z) represents the solution of the algebraic equa


the

tion system k and consists of the tray pressures, the vapor flow rates,

the

liquid flow rates, and

boiling points.

4.4.2

Numerical linearization

The matrices of the linear state space model

including

flow

dynamics

d_
dt

X
=

X n

d
u

(4.30a)

Cx

(4.30b)

can

be

numerically evaluated column by column using a finite difference


a

approach. After solution of the whole equation system for


operating point (OP), each composition
a
or

steady-state
be varied

tray holdup

can

by

small increment, the

algebraic equation

system k
can

can

be

solved, and according

each column of the state


to

dynamic matrix A

be calculated

a i=l

102,j

for

.51

(4.31a)

4.5

Comparison

of the linear models

89

(dx
dt
V

dx
a

dt

opj
for

Anj
1=1

102J

52

fdn
dt
V

1...50

(4.31b)

dn
a

dt

VAnj

opJ
can

An,
In
a

similar

manner

the

input matrix B

be evaluated.

4.5

Comparison of the linear models


Open loop simulations

4.5.1

Most important for the control

design

is

good representation of
range.

the

dynamic behavior
behavior
as

in

the

mid-frequency

The

steady-state

well

as

the

high-frequency behavior
a

Some idea of a linear model's

important. quality is obtained by a simple qualitative


are

less

comparison of the various models. Nevertheless,


requires
A
a

definitive

judgement

comparison

of control

designs based

on

the different models.

simple

method to compare the two linear models with the


is the simulation of

complete
inputs

rigorous model
are

step responses

to the model

(reflux L0, boilup V5i, feed composition xp, and feed flow rate F). These
shown

by the Figures
was

4.1- 4.4.

During the nonlinear simulations, the

bottom level

controlled by the bottom product flow rate B.


are

Except
at their
are

for the denoted

input, all other column inputs

kept constant

steady-state values. The changes of flow rates and feed composition


very small to maintain the column close to the

steady-state and

to avoid

large

nonlinearities.

The coincidence of the


is

step

responses with the

rigorous nonlinear model


a

acceptable
a

for both linear models. However, the linear model obtained

by

linearization of the

rigorous model

is

distinguished by

somewhat

better representation of the

low-frequency gains.

90

4 Linear Models

Tray 10

Tray

44

Nonhnear model

Anal linearized model

Num linearized model

Figure

4.1:

Step

response to

0.3 mol/min

(0.46%)

increase in reflux

Tray

10

Tray

44

Nonlinear model

Anal linearized model

Num linearized model

4 Time

(h)
a

Figure

4.2:

Step

response to

0.3 mol/min (0.29%) increase in

boilup

4.5

Comparison of the

linear models

91

Tray 10

Tray

44

Nonlinear model

Anal, linearized model

Num. linearized model

Figure

4.3:

Step

response to

0.005 mol/mol increase in feed

composition

Tray 10

Tray 44

Time

(h)

Time (h)

Tray 24

Nonlinear model

Anal, linearized model

Num. linearized model

Figure

4.4:

Step

response to

0.3 mol/min (0.91%) increase in feed flow rate

92

4 Linear Models

Surprising
as

is the

high coincidence for the analytically


the influence of the
is

linearized model
on

well.

Apparently

algebraic

constraints

the

composition dynamics
within the

substantially

smaller than the interactions

composition dynamics
values

themselves.

4.5.2

Singular

An important difference between the two linear models is the

high
of the

frequency behavior due

to unmodelled and modelled flow


a

dynamics,
4.5.

respectively.

Best suited for

comparison
,

are

the
,

singular values
Figure

transfer functions G,

>y(jco) andGu.

(jco)

shown in

in3

Disturbance inputs

10

Magnitude

"ii

10"

ioJ

10

10

10

10

Frequency (rad/min)
Control inputs

10

10'

10"

10

10

10'

Frequency (rad/min)

Figure

4.5:

Singular

values of the linear models

Upper plots:
Solid lines:

GA^y(i<Si)

lower

plots:

Gu->y (jo)

Analytically linearized

model

Dashed lines:

Numerically linearized model

4.5

Comparison

of the linear models

93

Both models show the

typical

course

of the

singular values for

high
are

purity

distillation column. In the low

and minimum

frequency singular values of the transfer functions Gu

range, the maximum

(jo)

very different and the condition numbers

<WGu-yfj>)}
K(jco)
=

(4.32)

WGu_>y(J)}
are

high. With increasing frequency, the maximum and minimum


never

singular values approach and the corresponding condition number

decreases, but Figure


sure

falls below 20. This is illustrated in

more

detail

by

4.6 for the transfer functions from the control


on

compensated temperatures
a

inputs tray 10 and 44. The large condition


to the direc

to the pres

numbers indicate
tion of the control

high sensitivity of the column outputs


u.

inputs

Consequently, the performance of a control


uncertainty
at the control

system

can

be very sensitive to

inputs.
as

Significant for the numerically linearized model


condition number
course

is the double and

big

in

the

low-frequency

range

the

completely

different

in the

high-frequency they

range. These differences of the


can

high-frequency

range between the models


are

be

explained from the

structure of the nonlinear models

calculated from:

As mentioned above the Thus the

analytical model neglects the flow dynamics. high frequency behavior is determined only by the first-order
0.1 rad/min)
are

equations of the composition dynamics. Therefore, the singular values


in the

higher frequency domain (above


of one

dominated

by

negative slope
The

magnitude

per decade.
account.

numerically linearized model takes the flow dynamics into


the reflux
are

Thus, considering
value
a

composition dynamics
frequency

as input, introduced, and for the minimum singular

column

additional lags for the decade for the

negative slope of several magnitudes

per

range above 1 rad/min results therefrom. The effect of the

flow dynamics

considering the boilup

as

the column

input is different:
on

An increase of the

boilup

increases the vapor fraction

the tray, which

94

4 Linear Models

10J
'

Singular values

v ,o T3 10

.___^^

'i
10

:
-

*^^^^^
V

^
\

\ \

10

10'

10

10"

10

10

10'

Frequency (rad/mm)

io3

Condition number
1
~

'i
-~
^

J.'. 1
'

a)

"3 3

io

^^^^
2
io
r

r
^
--

/
N

/
\ \
/

10

10

10

10

10"

10'

Frequency (rad/min)

Figure

4.6:

Upper plot: Singular


u

values of the transfer function from the control

inputs
Lower

to the

tray temperature

T10

and

T44
transfer function

plot: Condition

number K of the

same

Solid lines:

Analytically

linearized model

Dashes lines: Numerically linearized model

causes

higher liquid flow light

rates

leaving the trays. Because the composi

tion of the
more

light component

is

higher

in the upper

part of the column,


is

of the

component is

transported down, and the expected


retarded in the

decrease of the

light component's composition

frequency

range between 0.2 and 1 rad/min.

4.6 Order reduction


The orders of the linear models

developed above

are

52 for the

analytical
modern

model and 102 for the

numerically

evaluated model.

During

4.6 Order reduction

95

robust control

synthesis procedures such

of the model is

as H^ or (x-synthesis, the order enlarged by frequency-dependent weights for the model

inputs and outputs. Since the computation time for the controller design strongly depends
tance.
on

the model order, order reduction is of utmost

impor

Many
a

methods exist to approximate the state-space representa


a

tion of

linear system with

lower-order state-space approximation


are

[4.13]. Most of the mathematical methods which

available in

MATLAB
sponding
al. [4.5]
to

or

MATRLXX

toolboxes

are

based

on

computing the
corre

Hankel Norm

singular
the

values and

subsequent removing of states


methods,
with
a

relatively small Hankel Norm singular values. Jacobsen

et
a

compared

following

four

reduction of

column model of 82 states to 2 states. These methods


in
one

are

implemented

of the

MATLAB

toolboxes:

Balanced Truncated

Approximation [4.7] (Robust

Control Tool

box [4.2] and u-Tools [4.1])


II

Balanced Truncated

Approximation without balanced

minimal

realization [4.9] (Schur method, in Robust Control Toolbox)


III IV

Hankel Norm

Approximation [4.3] (n-Tools [4.1])

Optimal Hankel Norm Approximation without balancing [4.8]


(Robust Control Toolbox)

Jacobsen et al. conclude that the methods II and IV gave

significantly

better models than the other two methods. These results have to be

considered such
an

carefully:

It is not necessary to reduce the column models to


an

extremely

low order. Models of

order 10-15

suitable for control

synthesis and show

very

absolutely good coincidence with the


are

full-order linear model.


As
an

example,
an

numerically evaluated linear model of order

102

was

reduced to

order 10

using each of the four methods mentioned above.


inputs and calculated with the
a

All step responses to the different

different reduced-order models have shown

perfect coincidence
the

with

the full-order linear model. This fact is

supported by

singular-value

plots of the models. Figure 4.7 shows that all reduced order models

96

4 Linear Models

10J

10

I
Full order model Methods I
II

10

Method III

Method TV

10

10

10

10*

10

10

10'

Frequency (rad/min)

Figure

4.7:

Singular

value

plots

of the transfer functions G

(jco)

of the

full order model and the different reduced order models

approximate the low and medium frequency behaviors


very well.

up to 1 rad/min
are

However, in the high frequency range the singular values


the models derived with
or a

best

approximated by

Balanced Truncated

Approximation

(Method I

II).

4.7
This

Summary
chapter presented
linearization
of
a

two methods to obtain linear models for the


an analyt neglecting flow

industrial distillation column. The first model is derived by


ical

simplified

nonlinear model

dynamics and

most of the energy balance


a

equations. The second linear

model is obtained by

numerical linearization of the complete rigorous

4.8

Appendix:

Model coefficients

97

model. Both linear models exhibit


process

an

acceptable approximation
plots indicate
a

of the

dynamics.

The

singular

value

of the linear models in the


ences

mid-frequency
model,

range, but

high significant differ

coincidence

in the low and

high-frequency

range.
a

with those of the rigorous nonlinear


tion of the column

Comparing step responses slightly better representa


is

dynamics by the numerically evaluated model

demonstrated. The relatively high order of the linear model (52 for the

analytically, and 102 for the numerically linearized models) can be reduced essentially by one of the well known order reduction methods.
All tested methods

yielded
a

nearly perfect approximation of the Unear


a

model of order 102 up to

frequency of 1 rad/min by

model of order 10.

4.8

Appendix:

Model coefficients
following holds:

For all coefficients the

Analytically differentiating

the equation (4.12), the actual numerical

values of k. may be calculated.

A-Matrix

Condenser (k=l. ..50)

a, ,i

-v,
=

dx0

no

(4.34)

ai ,k

9f0
+ :

nd-rDk-iVj^
no

'"ark

(4.35)

ai ,52

_^o_.

(1-

n)50Vlk51
no

(4.36)

98

4 Linear Models

Trays (j=l...50, k=j.. 50)

^ _v.

j+,-j_3xJ-r
_

(4.37)

ni

*j
dx

i.J

^VZi^iLVZ&^V

(4.38)

(4.39)

at

aJ+,,52

axi=(1-Ti)50-Jkk
ox51

(V^.-a-TDvp

(4.40)

Evaporator

a52,51

3f51
_
-

_ _

^50
n

ox50

ax

(4.41)

n51

df

x51

-(B
-

Vksl)

51""51J

l52,52

ox51

ax

(4.42)

n51

B-Matrix We have to consider the portion of the vapor flow which is condensed by
an

increase of the feed flow rate because of the subcooled feed. For the

decrease of the vapor flow in the


column ( h'
dV
=

rectifying

section of the distillation

n20

-h'

'*

nF

dF

(4.43)

Vn 20_n20y
The

liquid flow
same

rate in the

stripping

section of the column is increased

by

the

amount.

4.8

Appendix: Model coefficients

99

Condenser

at,
b,
=

,
'

"

3x,

(4.44)

3fo
J.2
"

V,'
h

V,'
n

20 20

yi-xo
nn

3F~

(4.45)

Vn

n207
1... 19)

Rectifying

section

(j

af.

(4.46)

af.
h
DJ
+

'.2

20 20

X1

yj-n-yj

dF

(4.47)

Vn

n2oy

Feed tray (20)

"*20
21> '

*20
n
n20

d-x

(4.48)

oxF

521,2

5f20
_ "

xF

x20

n20~^F
,

,,

11

20

n20
x20

(y20

x20j
(4.49)

"9F
section

Stripping

(j

21

...

51)

(4.50)

af.
b.
J +

1.2

-i

9F

1+rrr
"

h'n-h 20 "F
20

Vl-*j

(4.51)

n2o;

nj

If the reflux is

subcooled, the

vapor stream in the column is condensed

partially

at the first

tray. The liquid stream leaving the first tray and all

100

4 Linear Models

trays below is thereby increased by the


obtain

same

amount. From

(4.8)

we

dV,

'h'i-V dL vh"i-h'iy
f,.h'i-ho
=

(4.52)

dLt

1+CV7
h"i-h'iy

dLn

(4.53)

With (4.52) and (4.53) there follows for the elements of the B matrix:

Condenser

h'j-h'g

bL3

af0 3L=

h"1-h'1(xo_yi)
n^

(4.54)

Dl>4"av"
Tray
1

9fo_yi-xo
nA

(4.55)

"

^
af,
i

x0-x1+(y1-x1)K^jl
"

h',-h'0
l

2,3

9L-

(4.56)

n,

y2-yi
n,

J2,4

av'

(4.57)

Trays (j

2...

50)
, j-"3-i

3f,

x-

x.

h'.-h'

"j

i.3

3L

1 CT

"0

(4.58)

nj
yj+i-yj

h",-h'iy

b;

9fj

'j

1,4

gy

(4.59)

4.9 Notation

101

Evaporator
f)f
X
x50

0I51

-x x51

'
1
1

u'

u'

52,3-

9Lafi

hj-h'0
l_n

(4.60)

5f

h"l-h'l7

-(yi-x,,)

b-w-^r
C-Matrix The coefficients of the measurement output matrix evaluated
are

(4-61)

numerically

by solving the boiling point equation for small

increments in

tray composition.

4.9 Notation
4.9.1 A B C G

Matrices and Vectors

State dynamic matrix


Control input matrix Measurement output matrix Transfer function matrix Transfer function matrix from
control

Gu
n

signals

to

output signals

Vector of holdup deviations from operating point (OP)


n
=

[dnpdn2, ...,dn50]T
[n,,n2,

Vector of holdups

iv
u

...,n50]T
manipulated variables (L0, V51) deviations

Vector of the
u=

[dL0,dV5,]T [L0>V5l]T

Vector of column inputs


*=

102

4 Linear Models

Vector of composition deviations from OP


x=

[dxQ, dxp ...,dx50,dx5,]T


compositions
XjjJ
[Xq, Xj,

Vector of tray
X
=

...,

Vector of the deviations of the pressure-comp. temperatures

y=[dTP10>dTP44'dTP24]T
Vector of pressure
y

compensated tray temperatures

[TPio' TP44' TP24.]


input deviations from OP
=

Vector of the disturbance


z

[dxp,dF]T
inputs

Vector of disturbance i=

[xF,F]T

4.9.2

Scalar values

B
D F

[mol/s] [mol/s]

Bottom

product flow

rate

Top product flow


Feed flow rate

rate

[mol/s] [J/mol]

hJ

Molar

enthalpy

of liquid

phase

on

tray j

h"i
Lo

[J/mol]
[mol/s] [mol/s]
[mol]
[-] [K] [K]

Molar
Reflux

enthalpy of vapor phase

on

tray j

LJ
nJ
nt

Rate of liquid flow

leaving tray j

Holdup

on

tray j

Number of trays in column

Temperature
Pressure

Tpj
V

compensated temperature

[mol/s]
[mol/s]

Boilup
Rate of vapor flow

Vi

leaving tray j

4.10 References

103

Xj
xB
xF

[mol/mol] [mol/mol]
[mol/mol]

Liquid phase composition

on

tray j

Composition

in column bottom

Feed composition

yj

[mol/mol] [mol/mol]
[-]
[-]

Vapor phase composition


Equilibrium
Relative
vapor

on

tray j
on

y
a

phase composition

tray j

volatility

Murphree tray efficiency Singular value


Condition number

[-]
[-]

4.10 References
[4.1]

Balas, G. J., J. C. Doyle, K. Glover, A. Packard, and R. Smith:

u-

Analysis and Synthesis Toolbox (\i-Tools), The MathWorks, Inc.,


Natick, MA (1991)
[4.2]

Chiang,

R.

Y., and M. G. Safonov: Robust Control Toolbox, The

Mathworks, Inc., Natick, MA (1988)


[4.3]

Glover,

K.: "All

optimal Hankel-norm approximations of linear


error

multivariable systems and their L

bounds," Int.

J. Con

trol, 36,1115-1193(1984)
[4.4]

Haggblom,

K. E.:

"Modeling

of Flow

Dynamics for Control of Dis


Control

tillation Columns," Proc.

1991 American

Conference,

Boston, USA (1991)


[4.5]

Jacobsen, E. W., P. Lundstrom, and S. Skogestad: "Modelling


and Identification for Robust Control of Ill-Conditioned Plants
a

Distillation Case

Study," Proc.

1991 American Control

Confer

ence,

Boston, USA (1991)

104

4 Linear Models

[4.6]

Kapoor, N.,

and T. J.

McAvoy:

"An

Analytical Approach

to

Ap

proximate Dynamic Modeling of Distillation Towers," IFAC Con


trol

of

Distillation

Columns

and

Chemical

Reactors,

Bournemouth, UK (1986)
[4.7]

Moore, B.C.: "Principal Component Analysis

in Linear

Systems:

Controllability,

Observability

and

Model

Reduction," IEEE

Trans. Automatic

Control, 32, 115-122 (1981)

[4.8]

Safonov, M.G., R. Y. Chiang, and D. J. N. Limebeer: "Hankel


Model Reduction without
Proc. IEEE

Balancing

Descriptor Approach,"
Dec.

Conf.

on

Decision and

Control, Los Angeles, CA,

9-11(1987)

[4.9]

Safonov, M. G. and R.Y. Chiang: "Schur Balanced Model Reduc


tion," Proc. American Control Conference, Atlanta, GA, June
17 (1988)
15-

[4.10] Skogestad, S. and M. Morari: "The Dominant Time Constant for


Distillation

Columns," Comp. Chem. Eng., 11, 6,

607-617 (1987)

[4.11] Skogestad, S. and M. Morari: "Understanding the Dynamic Be


havior of Distillation Columns," Ind. Eng. Chem. Res., 27, 18481862 (1988)

[4.12] Skogestad, S.: "Dynamics and Control of Distillation Columns


A Critical

Survey," 3rd IFAC Symp.

on

Dynamics and Control of

Chemical Reactors, Distillation Columns, and Batch Processes,

April 26-29, College Park, MD, USA (1992)


[4.13] Troch, I., P. C. Muller, and K.-H. Fasol: "Modellreduktion fur Si
mulation und

Reglerentwurf," at, 40, 2,

45-53 (1992)

5.1 Introduction

105

Chapter

A Structured

Uncertainty Model

5.1 Introduction
Each linear
or

nonlinear
a

dynamic model

can

only approximately
a

describe the behavior of model may be valid for


a

real distillation column. While

nonlinear
error

wide range of operating

conditions, the

of

linear model

rapidly

increases with the distance from its

design point due


ence

to process

steady-state nonlinearity. Since stochastic effects influ


well, the
be
error

the process behavior

as

of

linear model

compared
an

to the real process


error

can never error


a

exactly

determined.

Lacking

exact

description, the
as

between the process model and the process

itself is modelled

single frequency-dependent uncertainty bound


uncer

(unstructured uncertainty) or as several frequency-dependent tainty bounds (structured uncertainties).

Typical
ment

sources

of

uncertainty for
actuator

distillation column unmodelled


sources

are measure

errors,

limited
process

dynamics, and
occur

speed, nonlinearity. All these


can

high-frequency
of

uncertainty

simultaneously

and

be classified into three different groups:

Uncertainty of the manipulated variables (input uncertainties)


Model uncertainty due to process

nonlinearity and unmodelled

high-frequency dynamics

106

5 A Structured

Uncertainty Model

Uncertainty of the temperature


tainties)

measurements

(output

uncer

This

grouping corresponds

to the

principle that uncertainty should be

modelled where it
In this

physically

occurs.

chapter

an

uncertainty model for the industrial distillation

covers not only a single operating point but the entire operating range of the column. It is

column is

developed. The complete uncertainty model

the basis for the

analysis and synthesis of controllers using the


u\

struc

tured

singular value

5.2 Limits of
Before
we

uncertainty models
uncertainty
in the

start to model the

frequency domain,

we

must be conscious of its limits: An uncertain model in the

domain is
cients.

frequency model, which is time-invariant, but uncertain in its coeffi

This statement is best

explained by
of any

uncertainty
within the

at the

input

an example: Let us model a 10% plant and design a controller which


a

guarantees closed-loop stability and

certain

performance for all plants


time-varying plant, that

specified bounds.

Then the

stability and performance proper


a

ties of the controller


means

are not

guaranteed for

e.g. for variations of the

input

error

between -10% and +10%.

Consequently,
into account.

using uncertainty models in the


If

frequency domain,

the

excitation of the controller

by the time-variation of the plant is


are
a

not taken

time-varying uncertainties

assumed, nonlinear

simulations must be used for

validation of the robustness properties.


most
cases

However, the experience shows that for


tions with

uncertainty descrip frequency dependent and hence time-invariant uncertainty


sufficient.
our

bounds

are

This holds especially for


are

distillation column: The main disturbances

step changes of the feed flow rate. Each step change alters the
state

steady
the

operating point and defines


up to the next

a new

linear model

describing

dynamic behavior

step change. Each of these linear

5.3

Input uncertainty

107

models is defined

one

of the models within the set of all models. This set is

by

the

specified uncertainty bounds.

5.3

Input uncertainty
manipulated variables reflux
requested by
and

The actual values of the


never error

boilup will

match

exactly

the values

the control system. The the reflux and the true


causes are

between the setpoints for the

boilup

or

streams will be

frequency dependent. dynamic

The main
errors

static and

measurement

of reflux and

reboiler

duty
to pressure and

changing heat of evaporation due


temperature variations
reboiler

lags

actuator

lags sampling
errors

effects of

The bounds for the relative


modelled

of the column

inputs

can

be
the

by

multiplicative
error

uncertainty

description

with

frequency-dependent
bound wu
matrix

for the

boilup

bound wu for the reflux L and the error V These bounds are combined in the diagonal

Wu.

As illustrated in

Figure

5.1 the

following uncertainty model

holds:

u(jco)

{I

Au(jco)Wu(j<o)}u(j()

(5.1)

1 W

Au

u
1 1

Figure

5.1:

Multiplicative uncertainty description

for column input

108

5 A Structured

Uncertainty Model

with

|Au(jco)|
u

<1

loo

(5.2)

Wu(jco)

wUl(joo)
0

(5.3)

wUy(jco)
matrix

The

frequency-dependent complex
shapes only
case

Au(jco)

is limited in

magni

tude. It

the worst

error spatial during u-analysis (see Chapter 6). Therefore

the

direction of the

and is chosen to be
the

phase

behavior of the individual uncertainty bounds wu. is not significant. They should be chosen to be stable and minimum phase.
If we
assume

that the reflux and the


a

boilup
model:

errors are

independent,

the

matrix

Au

(jca) becomes

diagonal

matrix with two

single perturbations

yielding the following uncertainty

(jco)

8Ul(jo
0

5U
uv

(jco)

Wu(jco) u(jco)

(5.4)

with

|5u.(jco)|
Both models have been used for

<1

(5.5)

u-synthesis preferred

with very similar results.


in this

For two reasons, the model (5.1) is

study:
is reduced

The number of

uncertainty blocks

(Aj

or

8j)

by

one

compared

to

(5.4). This simplifies the fi-synthesis.


in reflux may
cause a versa.

Any change

change
are

of the vapor flow rate

within the column and vice

The interactions due to flow


to be considered here.

dynamics

and to the energy balance

5.3

Input uncertainty

109

Shaping the input


It has been shown

error

bounds
et al.

by Skogestad

[5.4] that the controlled system's


is very sensitive to
or

performance for
errors

high purity distillation column


should be estimated
as

in the

manipulated variables. For controller design

analysis

the

error

bounds

Wu

exactly

as

possible. This
condition
controller

holds

especially for the low-frequency high.


of
an

range, where the

number of the column models is

Otherwise

potential

performance
In the

is

given

away in

case

overestimation.

lower-frequency
are

range the

errors

of the

manipulated variables
parameter variation

at

the

plant input

strongly dominated by flow measurement errors and


an

parameter variations. As

example

for

we

consider the heat of evaporation in the reboiler. The

boilup is controlled

indirectly by
will

steam flow rate. Therefore

change

in heat of evaporation

cause an error

in vapor flow rate

leaving the reboiler.


conservative 20%
error

Skogestad and Morari [5.4]


brated the
error

assume a

state, which is fairly high. If all flow measurements


bounds should be less. An range is assumed to be
error

are

steady carefully cali


in

bound of 10% for the

lower

frequency

large enough.
measurement errors, range. The
errors

The effects of reboiler and

lags,

actuator

lags, dynamic

sampling
exceed

time

concern

the

higher frequency

caused

by these uncertainty
more

sources

increase with the

frequency

and

easily

than 100% of the nominal value for

frequencies

above 0.5-1 rad/min. The

steady-state

error,

together with the high-frequency

error, is well

approximated by

the first order

lead/lag transfer function

s/coM

G(s)=KTT^7uf
The

withcoN<coD

(5.6)

gain K represents the steady-state


chosen

error.

The cut-off frequencies

are

typically

according

to

coD

>

10coN.

110

5 A Structured

Uncertainty Model

5.4 Model

uncertainty
nonlinearity

5.4.1

Column

The

highly

nonlinear behavior of distillation columns is observed at


rate and feed

varying operating points (varying feed flow


and at transients fied composition

composition)

during dynamics (without feed

disturbance rejection. If we consider the


or

simpli side-product stream) of a

tray (see (4.6))

nj(S)
we

LJ-i(xJ-i-xJ)+vj+i^+!-V-vJ(yJ-xj)

(5-7)

recognize that the composition dynamics and thus the nonlinear

behavior

depend

on

the

varying internal flow

rates (L and

V), and

the

composition profile within the distillation column (repre by the liquid and
vapor

sented

phase compositions)

Effect of varying operating points Any control system for


the
a

distillation column must exhibit


errors

large gains

in

low-frequency
at
on no

range to achieve small control

Therefore,
tures

steady-state

both be

steady-state. product compositions (or the tempera


at at their

trays 10 and 44)

can

kept

setpoints. Thus transients


and the internal
a

have

significant influence

in the
as

low-frequency range
as

vapor and

liquid
a

flow rates

well

the

composition profile within

column become

function of feed flow rate and composition


a

only.
depends
actual
a

However, the dynamic behavior of

distillation

column
on

substantially

on

the

actual

composition profile

and

the

internal vapor and

liquid
can

flow rates.

Normally
a

the operating range of


a

distillation column
feed flow rate and

be bounded with

maximum and

minimum

composition. If we consider the whole operating range


we can

defined in this way,

observe the largest internal flow rates for the

smallest feed composition and

largest feed

flow rate

and,

vice versa,

5.4 Model

uncertainty

111

smallest internal flow rates for the


smallest feed flow rates. The

largest feed composition and


state

composition profiles for these two steady

states bound the domain of all

steady

composition profiles (see

Figure 2.2).
Hence
we

can

conclude that the

low-frequency behavior by
a

of

binary

high-purity

distillation column is bounded

the models for maximum

and minimum column load. As

basis for further discussion the

following three linear models


Model N Model I

are

introduced:

column at nominal load column at maximum feed flow rate and


minimum feed

composition (increased load)

Model R

column at minimum feed flow rate and


maximum feed

composition (reduced load)


are

The feed data of the different models

listed in Table 5.1.

Table 5.1:

Operating

conditions for

design
Feed

purposes

Operating point
OP-N
OP-I

Feed flow rate

composition
0.8 0.7 0.9

(mol/min)
33
46

(mol/mol)

OP-R

20

The

simplest

way to

represent the column nonlinearity due to varying


a

operating points would be by


assume

multiplicative output uncertainty.


following form

If we

that the

uncertainty

for each model output is

independent of the
for the

actual value of the other two model outputs, the

output uncertainty holds

(Figure

5.2):

112

5 A Structured

Uncertainty

Model

Figure

5.2:

Multiplicative uncertainty

at

output

8
M

(jco)
0 0

0
0

y(jco)

"

8y2 (jco)
OS
3

Wy(jco)
.

y(jco)

(5.8)

(jco)

with

Nl--1
and

y(jco)

GN(jco)

d(jco)

u(jco)_

The transfer matrix

Wy

is
,

diagonal matrix with the uncertainty


^2
,

bounds for each output (w,r

wv

^1

bound for these uncertainties

can

be obtained

-^3

on

the main

diagonal.
a

An upper

by

calculation of the

standardized
d-

errors

y- of the

AGj (jco) models Gj (jco)


=

and and

AGR (jco) for each channel GR (jco) respectively:


,

ui

-^

yj

or

AGj(jco)

[GI(jco)-GN(jco)]G^1(Jco)

(5.9)

5.4 Model

uncertainty

113

AGR(jco)

[GR(j(o)-GNa<))]GN1afl
uncertainty weights
.

(5.10)

The upper bound for the standardized


errors

is the maximum of all


j

for the output y

In earlier papers it has

already been recognized that column nonlin by the

earity

is not well

represented by simple multiplicative uncertainty


two

bounds at model output (McDonald [5.2]). This fact is confirmed

uncertainty bounds for the


and

numerically evaluated

linear models

Gj

Gr

which include the flow

dynamics (Figure 5.3). The multiplicative

output uncertainty exceeds 80% (for


It is

Gu

) in the low-frequency range.

significantly smaller in the medium


for

frequency range,

but increases

sharply

frequencies above

0.1

rad/min, where the flow dynamics

influence the

dynamic

behavior. An uncertainty
in the

description with such


range is

high multiplicative uncertainty


tive for any control

low-frequency

prohibi

design.
are

Fortunately the
steady-state
decrease of the

errors

highly

correlated: The variation of the


a

operating points

causes

simultaneous

increase

or

singular

values of the transfer functions from the control


y

signals
trated

(L and V) to the model outputs

(T10, T44, T24).


Uj

This is illus
-

by the Nyquist

plots
or

for the individual channels

y.

(Figure
a

5.4). It

clearly shows

that the variation of the column load

causes

simultaneous increase

decrease of the

open-loop gains

in the low-

frequency domain. Thus


This
can

we can assume

that the

dynamic behavior of

the distillation column must lie "somewhere between OP-I and OP-R."

be

represented by
and

linear combination of the two column

models

Gz (jco)

GR (jco) (Figure
+ GR

5.5)

G(jco)

GTlU(jco)
=

(jco)
+

5G(jco)

GTlU(jco)

GR

(jco)

(5.11)

with

|6g|L-i

8Ge

or

8Ge

114

5 A Structured

Uncertainty

Model

Standardized

error

of T10,

Gj(s)

Standardized

error

of T44,

Gj(s)

10

10

10

10

10

10

10

10

10

10

10'

10"

Frequency (rad/mm)
Standardized
error

Frequency (rad/min)

of T24,

Gj(s)

10

10

10

10

10

10

Frequency (rad/min)
Standardized
error

of T10,

Gg(s)

Standardized

error

of T44,
1

Gr(s)
1

0.5

"""%, \
/
/

/
.

1
1
.

\ \

s^.-

-4l
i1
\
'

111

i-

10

10

10

10

10

10

10

10

10

10

10

Frequency (rad/min)
Standardized
error

Frequency (rad/min)

of

T24, Gr(s)

Legend
xF

Frequency (rad/mm)

Figure

5.3: Standardized model

errors

at

operating points OP-I and OP-R

5.4 Model

uncertainty

115

GL-T10(Jffl)

GV^T (J>)

GL_T44(jco)

GL^ T CJ0
6
1

1
1

V
\\

/
/

<z>
/
/

\
\
y

/
/

-250

500

Figure

5.4:

Nyquist plots

for different column loads

solid lines: Model N;


dash-dotted lines: Model I;
dashed lines: Model
x: 0)
=

R;

lxlO"3 rad/min;

o: co

lxlO"4 rad/min

116

5 A Structured

Uncertainty

Model

r-H.Gl J~^-^'
T+
r-

5G*3
1/2
u

Figure

5.5:

Uncertainty model due

to

nonlinearity

in the

low-frequency

range

The uncertainty parameter

may be either
a

complex

or

real. If

we

define it to be
and

complex we
means

allow

phase shift for all models between


are

GR
a

GT, that
line.

the
a

points of all models


fixed

in the defined set in the


not

Nyquist plots

and for

frequency

required

to be

on

straight

In this way

we

generate

plant which

covers

the

properties

of the distil

lation column at low and at high feed compositions, and at different feed flow rates without introducing additional conservatism. It is
to model such
a

impossible

behavior with

an

unstructured uncertainty

description!

Effect of transients
While
an

appropriate uncertainty model for different operating points

requires
sitions

highly

structured uncertainty

description, the effect of tran rejection, the


well
as

sients is rather unstructured:


on

During

disturbance
as

compo

tray 10, tray 44, and tray 24


from their

the

product
a

compositions will deviate


movement of the

steady-state values, caused by

nonlinear

composition profile toward one vapor/liquid equilibrium, the singular values


may

column end. Due to the


of the transfer

functions

Gu

change
.

in different

directions, e.g., towards


of

higher
.

singular values of G
Due to the

and lower

singular values
in the

high controller performance

low-frequency

range, tran
cause

sients do not affect the

low-frequency

range.

However, they

5.4 Model uncertainty

117

nonlinearity

in the middle and


a

higher frequency
w

range, which can be


as

described with

multiplicative uncertainty description


are

in

equation

(5.8). The uncertainty weights


values in the

chosen to have large

singular

higher-frequency
It is not

range and low singular values in the lowto calculate these


cause a

frequency range.
exactly.
point

possible

uncertainty bounds operating

input point, but the magnitude of the deviation from


cannot be

Each disturbance

will

variation of the
a

steady-state operating

predicted. The

selection of

appropriate transfer func

tions is discussed in

Chapter

6.

5.4.2

Unmodelled

dynamics Chapter
are

It has been shown in

2 that flow

dynamics affect the highdesign,


an

frequency behavior of distillation columns. If linear models which

neglect the flow dynamics


uncertainty model

used for control

appropriate

is necessary.

Most authors treat the effect of flow

dynamics in the
< x <

same

way

as

the

effect of an input time

delay

with 0

1 minute

([5.1], [5.3], [5.4],


a

[5.5]). The corresponding input uncertainty is often modelled with

multiplicative uncertainty, using


the

first order Pade

approximation

for

uncertainty bound ([5.4], [5.5]):


1

_I5
(5.12)

e-B

This

uncertainty

can

be combined with the other

input uncertainties
some

(Chapter
sented

5.3). Lundstrom et al. [5.1]

point

to the fact that

combi

nations of

delay uncertainty are not repre using simple uncertainty weights. They developed new and more
time
cover

gain uncertainty and

complicated uncertainty bounds, which


combined gain
control results

the whole domain of the

uncertainty

and time

delay uncertainty. However,


(Chapter 6) show
for the input
very

design

studies in this research

good

using simple

first-order

weights

uncertainty

description.

118

5 A Structured

Uncertainty

Model

Model

uncertainty due

to flow

dynamics could be represented by


as

multi

plicative output uncertainty,


tage that the

well. This
can no

approach has the disadvan


time

uncertainty bounds

longer be approximated by

delays.

5.5 Measurement
An additional
source

uncertainty
are

of uncertainty
a

the temperature measurements.


sensor

The
a

dynamic

behavior of

temperature

is well

approximated by

first-order

lag:

GT(8)
The time constant

~L-

(5.13)

TT

of this transfer function

depends

on

the tempera

ture measurement

position. While the time


sensor

clearly below
to

1 minute if the

is

constant will usually be placed in the liquid phase, we have

expect time constants


In the
case

up to 10 minutes if it is

placed

in the vapor

phase.
tion,
a

of the industrial distillation column under investiga

position
on

in the

liquid phase

cannot be
we

guaranteed because of the

small head

the

plates.

Therefore

have to consider time constants

up to 10 minutes for the control

design.

The gain

KT

of the

sensor

model

GT depends

on

the

sensor

calibration

and

on

the heat loss to the environment. The

sensor can

easily be cali

brated with

high accuracy. However, the dynamic effects of the heat loss


are

due to variations of the environment temperature


mate.

difficult to esti
cause a

They

concern

mainly the low-frequency

range and

slow

bias variation of the temperature measurements. This effect is compa

rable to variations of the setpoints for the control system. The of the control system is not affected if large bias variations
are

stability
avoided.

They

would lead to

product compositions
A

which

are

very distinct from

those at the

design operating points.


sensors

good thermal isolation of the

temperature
Because the

is thus recommendable.

u-analysis and u-synthesis guarantees the performance for proposed


to include the model for the

the worst case, it is

temperature

5.6

Specification

of the controller

performance

119

sensors

with

gain

KT

1 and

time constant

TT

10 min into the

column model. will be

Thereby further uncertainty blocks closed-loop system's stability due


w

can

be avoided. It

easily recognized later, that shorter


the

time constants
to the

TT

will not

endanger frequency

large output
in the
upper

uncertainties

specified

for the controller

design

range.

5.6
The

Specification

of the controller

performance

uncertainty model discussed above is structured. A controller


or a

design

robust

performance analysis requires the framework

of the

structured

singular value u, which expects the disturbance

inputs d by the

(feed composition, flow rate), the reference inputs


error

r, and the control

to be in

H^-norm

bounded set. This is illustrated


5.6.

frequency shaped plant in Figure

i!
We(8)
/ I

Wd(s)

+o
-L

*\J 1

[*

K(s)

GA(s)

T24
T10>T44
Figure
5.6: Performance

specification for the

uncertain

plant

The uncertain plant

GA(jco)

describes the nonlinear behavior of the


is defined
as

distillation column. The

performance objective
to be in the set

making the

weighted

control

error e

120

5 A Structured

Uncertainty

Model

sup

[jWe (j co)
2

(j co)

J2

<

Vco

R+

LrJ

The

following H^ bound is equivalent


T
d
r

to this

specification:

(jco)
^P

We
e.

(s) is

(usually diagonal)
in
a

matrix of transfer functions which

shapes
T

the maximum allowed


If

amplitude

of the transfer function from [d, r]


range,

to

We

is

large

certain

frequency

only

small control

error

is allowed there.

The matrix

Wd (s) shapes the frequency content of the disturbances and


case

setpoint changes. In the

of our distillation

column,

variations of the

feed composition and feed flow rate will affect the medium and lower

frequency
control

range. First order

lags shape
as

the

frequency

content of these

two disturbances

quite well. Because measurement noise enters the


same

loop
are

at the

position

the reference

inputs, the

corre

sponding weights
chosen

are

chosen to be constant. The

weighting

functions

discussed in the following chapter.

5.7

Summary

The complete uncertainty model is shown in Figure 5.7. It consists of the

input uncertainty (5.1), the model uncertainties (5.8) and (5.11), and the

performance specifications. Simple dynamic


sensors
are

models of the temperature

included in the column models. This relatively

complex

uncertainty model has the advantage that the entire operating range of
the distillation column is covered. The

large

conservatism of an unstruc

tured uncertainty description is avoided. Therefore, with


dures based
on

design

proce

the

u-synthesis

or

u-optimization, operating

we

can

expect high

controller

performance for

the entire

range.

5.7

Summary

121

6 Eh

rV)^ ** +
Us
0)

I
iL
iL

a t
u

a s

a
o

o
IO

Q.

122

5 A Structured

Uncertainty

Model

The input uncertainty bounds wu


tions
are
a

are

easily shaped. Only

few reflec

necessary about the


error

where

100%

is to be

steady-state error and the frequency expected. However, the output uncertainties

wy.

are more

difficult to

it is often necessary to
tions

shape. During the controller design procedure adjust them iteratively until nonlinear simula
This

show

satisfactory closed-loop dynamics.


Chapter
6.

problem

is

discussed further in

5.8 References
[5.1]

Lundstrom, P., S. Skogestad, and Z.-Q. Wang: "Uncertainty

Weight
30th

Selection for
on

H-Infinity

and Mu-Control

Methods," Proc.
U. K. (1991)

Conference

Decision and Control,

Brighton,

[5.2]

McDonald, K. A.: "Characterization of Distillation Nonlinearity


for Control
trol

System Design
(1987)

and

Analysis," The Shell Process Con


Morari, Butterworth, Bos

Workshop,

ed. D. M. Prett and M.

ton, 279-290

[5.3]

Postlethwaite, I., J.-L. Lin and D.-W. Gu: "Robust Control of

High Purity Distillation Column Using u-K Iteration," Proc. 30th

Conference
[5.4]

on

Decision and
M.

Control, Brighton, U.

(1991)

Skogestad, S.,
matic

Morari, and J. C. Doyle: "Robust Control of 111-

Conditioned Plants:

High-Purity Distillation," IEEE

Trans. Auto

Control, 33,12, 1092-1105 (1988)


and P. Lundstrom:

[5.5]

Skogestad, S.,
Distillation

"Mu-Optimal LV-Control of Eng., 14, 4/5,


401-413 (1990)

Columns," Comp.

Chem.

6.1 Introduction

123

Chapter

ja-Optimal Controller Design

6.1 Introduction
While the

synthesis and analysis of controllers using the structured


a

singular

value

(SSV) has attracted considerable attention among

aerospace and electrical

engineers (e.g., [6.8], [6.9]), it has been less


process control

commonly considered by

engineers. One

reason

for that

might be the lack of adequate structured uncertainty models for chem


ical processes. The uncertainty model discussed in the forms
a

previous chapter
uncer

suitable basis for


covers

u-optimal controller design. Since this

tainty model
antee

the

dynamic behavior of the industrial distillation

column for the entire operating range, the

resulting controllers

guar

stability

and

performance for all operating points.


the results of a

This

chapter presents

u-optimal controller design for the


a

LV control structure of the distillation column. After


most useful

summary of the

aspects of the SSV, the

design

of

state-space controllers by

u-synthesis

is demonstrated. Because the


a

implementation
a

of state-

space controllers in

distributed control system is

troublesome

project, the design of controllers with fixed and easy-to-implement


structures (PID control

structures) is considered in

special

section. A

comparison of the controller's performances in the time-domain termi


nates this

chapter.

124

|i-Optimal

Controller

Design

6.2 The structured


The uncertainty model

singular

value

approximating the nonlinear dynamic behavior

of the industrial distillation column (see simultaneous uncertainty blocks

Chapter 5) includes several


control

(8i; A-), thus representing a structured


design
based
on

uncertainty
methods

model.

Most of the well-known robust


are

(e.g., H^, LQG/LTR)

unstructured uncertainty
on

descriptions. The application of these methods

such uncertainty

models often introduces unnecessary conservatism in controller

design,

because these methods combine all the uncertainties in

one

large, fully
use

occupied uncertainty block. Thus


ties is

the

special
can
was

structure of the uncertain

neglected. This conservatism

be avoided

by the

of the
J. C.

structured

singular value

|i, which

introduced in 1982

by

Doyle ([6.5], [6.6]).


The structured

singular value (t

so

far has seldom been discussed in


are

textbooks. Therefore the most important facts about |i


within the

summarized

following complex uncertainty blocks.


can

three sections. The discussion is restricted to Results for mixed

real/complex

uncertain

ties

be found in [6.15]. The references [6.4],

[6.12], and [6.14]

contain additional informations.

6.2.1

Representation

of structured uncertainties

The definition of the structured

tainty

model for
6.1. The

singular value presumes that the uncer plant is rearranged into a special form, as shown in
process models and the

Figure

plant P consists of the

weighting

functions. It has three sets of inputs and outputs: The first set of

inputs and outputs is highly important. Within the


set

uncertainty model, this


uncertainty blocks. In
the

represent the output and input signals of the


the

our case

inputs

to the

uncertainty blocks
are

are

signals po,,0

and no. The

corresponding outputs

the

signals

p.^i.andTii.
The second set of inputs consists of all external reference lated

signals (disturbances d,
manipu
p,

inputs
u.

r), while the third set of inputs consists of all

inputs

The

corresponding

set of

outputs contains the outputs

6.2 The structured

singular value

125

A
=

Uncertainties

P
Plant

K
Controller

Figure 6.1: Standard representation of an


The definition of the vectors
z

uncertain

plant.
5.7

and

is related to

Figure

subject

to any

performance

measure

(e.g., the weighted control error),

and the measured

plant outputs

y,

respectively.
more

If the uncertainty model is structured (i.e., it contains

than

one

uncertainty block) the matrix A is uncertainty blocks


on

the main

diagonal diagonal. In case of the uncertainty


a

block

matrix with all

model for the distillation column considered in this block structure holds:

thesis, the following

diag (A

u,

6GI3,5y], 5^, 6^|Au


or,

C2 * 2,

8G e C, 8y.
R, 8

C )

alternatively,
C2 x 2,8G
e
e

diag (A

u,

8GI3,8 y,' 8y2',8 y3 IAu e


with

^i

C )

(6.1)

M-^ftl-*1
model,
the matrix A is
a

For

an

unstructured uncertainty

fully occupied

matrix without

predefined

structure.

126

(i-Optimal

Controller

Design

The rearrangement of

an

uncertainty model

into the standard form is

always possible.
[6.1]
as

The

MATLAB p-Analysis

and

Synthesis

Toolbox

well

as

the Robust Control Toolbox [6.3]

provide

efficient tools

for that purpose.

6.2.2

Definition of the structured

singular value
a

Let X be the set of all A matrices with


structure:

given, fixed block-diagonal

X-

{diag

8]Ir,...,8sIr,A1,...,Af 1^
singular

C,

A.

CmJxmJ}

(6.2)

The structured
m
=

value [6.7] of the Matrix Me Cmxm

with

Yfj

Ym. (Fig. 6.2) is defined by


1

min

HA(M)

'

Ae X

{omax(A)|(det(I

MA)=0)}
(6.3)

ifno(AeX)

solves

det

(I

MA) =0

Hence 1/u

(M) is the size of the smallest matrix

A which

moves a

pole
case

of the system shown in


of
a

Figure

6.2 onto the

imaginary axis. In the

nominally stable system M, 1/li(M)


matrix A. In
case

is the size of the smallest

destabilizing

of

nominally unstable system M,

Figure

6.2: M-A feedback connection

6.2 The structured

singular value

127

however,

|i(M)

is

not

defined,

and

the

numerical

results

are

misleading.
Some

important properties of\l [6.7] scaling matrices:

Let D be the set of diagonal


D=

{diagfD!

,...,Ds,ds

1Imi,...,ds

FImF]
(6.4)

|DiCrixrsDi=D*>0,ds+jR,ds+j>0}
and let U be the set of block-diagonal unitary matrices

{diag(Ulf U2, ...,Un)

|Ui6

C^U^U^}

(6.5)

With these definitions the

following properties of \i hold:


(6.6)

p(M)<u(M)<omax(M)
H(DMD->)
max
=

u(M)
inf

(6.7)
(6.8)

UeU

p(UM) <p(M)

<

ofDMD-1)
mM

DSD

Property (6.6) reflects the advantage of the structured singular value u: In the presence of structured uncertainty, usually the inequality holds.
Therefore,
u

is smaller than the maximum

singular value. by property (6.7),

The invariance of \i to

diagonal scaling
as

is indicated

which is essential for the

singular

values

as

well

for the DK-iteration for


u

approximate calculation of the structured u-synthesis. exactly.


All

No direct way has been found yet to calculate for the numerical

algorithms

computation calculate upper and lower bounds

according

to

property (6.8).
an

Both bounds represent

optimization problem.

The

optimization

problem for the


2S
+

upper bound is convex. For

simple

block structures with


to

<

3 the upper bound is

guaranteed

to be

equal

nA (M).

128

fi-Optimal Controller Design

The optimization problem for the lower bound is not

convex

and its

calculation may converge to local maxima. Nevertheless numerical

experience indicates that usually the difference between


lower bounds is within 5%, and almost

upper and

always

within 15%

([6.12],

[6.14]).

6.2.3

Robustness of stability and performance


we

Before

start to discuss robust

stability

and
we

performance within the


have to join the

framework of the structured

singular value,

plant

and the known controller K of the standard close the control loop

configuration
a

in order to

(Fig. 6.3). This is easily done by

linear fractional

transformation [6.13]:

M(P,K)
The

J,(P,K)

Pn+P]2K(I-P22K)-iP 12
outputs:

(6.9)

resulting plant M has

two sets of inputs and

M M12 M21

M22| [d] |_rj


A
^

(6.10)

d
r

M(P,K)

Figure

6.3:

Representation of uncertain control system with controller K and plant


P combined into the system M

6.2 The structured

singular value

129

The

input

sets

are

(1) the outputs from the uncertainty block A, and (2)

the disturbance and reference

inputs. The outputs, in turn,


A and the set of

are

the

inputs
P-

to the

uncertainty block

performance

measures

Theorem 6.1: Robust

stability

Let BX be the set of all block


ture and with

diagonal

matrices with

particular struc

infinity-norm-bounded submatrices:

BX

{diagr81Iri,...,8sIvAI,...,Afl
V

|6i6c,Ajec,.m8i|^i.hL:si}
The system shown in

(6.11)

Figure

6.3 remains stable for all A

BX if and

only

if
sup
co

uA(Mn)<l

(6.12)

Proof:

see

[6.6]

Theorem 6.1 allows the


tured uncertainties. If for

stability analysis of control systems with plot the


in the
u

struc

we

upper and lower bounds of |A

(Mu)

enough frequency points

frequency

range of interest and find

that the maximum value of

is smaller than one, the control system is

stable for the uncertainties

specified with the assumption


for any

|oJ

<.

1,

||Aj|
not

<

1. If p,

(M) exceeds
uA)
a

one

guaranteed
<l/(sup
m

to be stable.

frequency, However, for smaller uncertainties with


the control system is
\iA)
a

II5J
ii

iii~

and |A.||
ii

Jiioo

<l/(sup
w

stability is guaranteed.

Theorem 6.2: Robust performance


The

performance

of the control system is


with
A

robustly achieved if and only if

supp-(M)<l
Proof:

diag[A,Ap]

(6.13)

see

[6.6]

130

u-Optimal Controller Design

For the

application of theorem
uncertainty

6.2

we

have to add

one

uncertainty block

Ap

to the

structure A

(Fig. 6.4). Imagine that the perform

ance

specification

of the control system is met for all allowed distur


case

bance matrices A in the set BX. In this

the output p is bounded


<

by

HpIL
loop

<

1 for all

inputs [dT, rT]T

with

|| [dT, rT] T|L


Ap
with

1. If

we

close the
<

from p to

[dT, rT]T by introduction of the block Ap

with
<

|Ap|[

1,

the system will be stable. But if any block the

||Ap|

1 destabilizes

loop

[dT, rT]T,

the

specified performance
the

cannot be achieved in the frame


a

for all

possible plants within

specified

set.

Therefore,

work of the

SSV, the robust performance problem is handled like


robust

sta

bility problem. A test for


robust

performance will be similar

to

test for

stability.
it

Because the test for robust will be sufficient.

performance includes robust

stability,

usually

A 0

AP

<)

M(P,K)
performance setup for the SSV-framework

Figure

6.4: The robust

6.3 The
The

design
the

model
for the industrial distillation column

design of u-optimal controllers


on

is based
5.7
on

uncertainty

model

developed

in

Chapter

5 (see

Figure

page 121). For this model, the weighting functions for the

6.3 The

design model

131

input uncertainties,
output uncertainties,
reference

inputs, disturbance inputs, and

controller

performance

are

to be selected.

All

weighting

functions

are

chosen

as

diagonal

matrices:

Wd(s)
Wu(s)

diag[wXF(s),wp(s),wrio(s),wr44(s)]

(6.14)

=diag[wUL(s),wUv(s)]
=

(6.15)

Wy (s)
We(s)

diag[wyifl (s), wy^(s), wy^(a)]


diag|"we (s),we (s)l
weighting functions
is

(6.16)

(6.17)

The selection of the

primarily done

on

the basis

of physical considerations: Feed disturbances: The variations of the feed


flow rate will affect the lower

composition

and the feed


contents

frequency by
are

range. The

frequency

of these disturbances

are

modelled

first-order

lags. Typical weights

chosen here for the control

design

wv

Xp

(s)

XF1

-iwithK
+ 1
S

0.1

XF

mol/mol,

T
XF

180 min (6.18)

wF (s)

KF 1

*
+

with

lps

Kp

mol/min, Tp

120 min

(6.19)

Reference inputs: The reference inputs changes


0.2 C:
well

r can

be used to model
are

setpoint
constant

as as measurement noise. They weights, representing setpoint changes and

chosen

as

measurement noise of

132

u-Optimal Controller Design

wr

(s)

wr

(s)

0.2

(6.20)

Input uncertainties: An uncertainty of


ulated variables within the uncertainty is
a

10% is assumed for both

manip

wide

frequency

expected

to be much

range. For higher frequencies higher. An uncertainty of more

than 100% is assumed for

co >

0.5 rad/minute:

wUl(s),wUv(s)=0.1i^
Output
is

(6.21)

uncertainties: The resolution of the temperature measurements The


maximum

limited.

deviation of the

pressure-compensated
to be reasonable. In the

temperatures from their setpoints is usually significantly smaller than


1 C. A measurement
error

for AT of 10%

seems

higher frequency
mismatch. An

range the

output uncertainty
a

is affected

by model
cross

assumption

of

100%

error

for

co

1 /16 rad/minute has

shown
over

good
one

results in controller

design. Adjusting

of this 100%

is

of the

high-frequency possibilities resulting controller. Typical uncertainty weights are


of the
to influence the

behavior

wy

yiO

(s)

y44

(s)

ylA

(s)

0.liif
1 + 1.0/S

(6.22)

Performance weights: The performance weights "punish" the control


error

in the

frequency domain. These weights


with
a

have been selected

as

first-order

lags

large steady-state gain, which forces nearly inte

grating behavior of the controller. The cut-off frequency of these


is
a

weights

matter of

optimization:

If the

frequency

is too
a

high, robust perfor

mance

cannot be achieved. On the other hand

cut-off frequency speci

fication

may lead to

significantly lower than the maximum attainable frequency an unsatisfactory controller design. This holds especially for the uK-iteration which is discussed later. A typical performance specifi
cation, which allows
a

0.01 C

steady-state offset,

is

given by
(6-23)

^-^-^T^Sooi

6.4 Controller

design with u-synthesis

133

All

weights above

are

illustrated

by Figure 6.5

Uncertainty weights

Input

and

performance weights

10

10

10

10

10"

10

10

10"

Frequency (rad/min)

Frequency (rad/min) Weights


for

Figure

6.5:

u-synthesis

6.4 Controller
The

design

with

{i-synthesis
a

objective of the (i-synthesis


on

is the calculation of

stabilizing

controller K without restriction

the controller order and its structure,

which minimizes the SSV for all frequencies:


K=
arg
K

inf stabilizing

|Ui-(M (P, K)) I


"

(6.24)

>

As it is not
is

possible

to calculate the SSV

exactly,
u

the

design task (6.24)

usually approximated by K=arg


inf

the upper bound for


inf

(6.8) (6.25)

Kstabilizing"

De D

amax(DM(P,K)D-1)

The aim of the


of |i- for the

u-synthesis

is

perfectly reached, if the


is below

maximum value

closed-loop system (Figure 6.4)

one.

134

(i-Optimal

Controller

Design

6.4.1

Synthesis algorithms
not
a

The

u-synthesis is

trivial task. Yet

no

algorithms have been devel

oped which allow


The known

a one-step solution of the u-synthesis problem (6.24). algorithms require the repeated calculation of an H^ a

problem, alternating with


cannot

scaling of the plant. These algorithms

guarantee convergence of the iteration.

DK-Iteration
The

synthesis problem (6.25) is


no

simultaneous
matrices

optimization problem of
an

the

frequency-dependent scaling
If we

D and the controller K.

Because

direct solutions exist, Doyle [6.7] proposes

iterative

approach:

keep

the

diagonal scaling matrices


(DMD-1) ||

D constant, the mini H

mization of 1

||

inf
DgD

max^

Ml-

forms the

convex

problem
(6.26)

arg

infllDTUP.KJD-1!
Kll
1

II-

If we fix the controller K,


tion
are

equation (6.26) represents a convex optimiza problem for the diagonal scaling matrices D. These scaling matrices
a

optimized by
uA

u-analysis of the closed-loop system:


=

[?i (P, K) ]

inf o
DeD

(DMD-1)

(6.27)

The

frequency-dependent scaling

matrices D
.

are

approximated with
u-

stable, rational transfer functions D (s)

Alternating the HM controller


procedure
1.

synthesis and the optimal scaling,


synthesis problems
illustrated in

convergence is achieved for most

after several iterations. The iteration

is

Figure

6.6. The DK-iteration is finished either if the solu

tion does not show any further


convergence cannot be

improvement
Both of the

or

if

u <

However,

guaranteed:

single optimization prob

lems

are

convex, but not the overall

optimization problem (6.25). The

optimized scaling
mization
zation

matrices D

are an

optimal solution for the local opti


not

problem (6.27),

but

they

are

optimal for the global optimi

problem (6.25). Therefore,

the DK-iteration may converge to local

minima.

6.4 Controller

design

with

u-synthesis

135

K0

arg

infl^P.K)^

ZZZ3ZZI
^[^(P.Ko)]

=D>nfDamax(DMD-i)

Fit D(s) with

stable,

mm.

phase

transfer functions D (s)

T
K,
=arg

irfJDfsj^fP.KjD-'fB)!.
|

^[JjCP.K,)]

=D.nfD0-max(DMD-i)

Figure

6.6: DK-Iteration

ui<L-/terafJon A
new

algorithm for u-synthesis has been proposed by

Lin et al. [6.11].

Instead of fitting the


of the

scaling matrices D, this algorithm is based on a fit frequency-dependent SSV with a stable rational transfer func
plant is premultiplied with a diagonal u-approximating transfer function. Thus the peaks of u
of interest
are

tion. At each iteration step, the matrix of the

within the

frequency range

punished, and the algorithm


algorithm
is not

tries to flatten the

u-curve.

The convergence of this

proved.

The authors present "a reasoned

the sequence will

argument for believing that converge" [6.11]. A scheme of the uK-algorithm is

136

|i-Optimal Controller Design

K^arg inff?, (P,K)|m

u(jco)

^[^(P.Kq)]

Fit

|10 (jco)

.,

'..

ji0(s)

stable, miminum phase


"

z.iz'vi"
K,
=

arg

inf|jio(s)7'1(P>K)||oo

;::_
u(jco)

:.l."._::..:
=

u-fJiCP.K,)]
' '

:::::: ::\:
*.=[?&

i
K=

arg
'

infNji. (s)(l0(s)?1(P,
"
'

K)||

J...
6.7: uK-Iteration

Figure

usually
gence

converges

more

slowly than
u

the

DK-iteration, and the


on

conver

properties

are

strongly dependent

the fit of the

u-curve.

Even

if convergence is

achieved,

is not minimized for all

frequencies.

6.4.2 The

Applying the DK-iteration apphcation


of the DK-iteration to
our

design problem

was

not

successful because convergence of the

algorithm

is not attainable. Most

6.4 Controller

design

with

u-synthesis

137

likely, the

main
or

problem

is the fit of the

fully occupied

2x2

or

3x3

(including
the

excluding the

measurement of

T24, respectively)

block of

D-scaling
and

matrices. This block results from the

repeated scalar
a

uncertainty block

8GI.

If we fit each position of this block with

scalar,

stable,

minimum-phase

behavior for the

minimum-phase resulting MIMO system is guaranteed. However,


a

transfer function,

unnecessary conservatism is introduced

thereby,

since

minimum
not for the

phase behavior
single
6.4.3

is

only required for the MIMO system, but


problem

scalar transfer functions. This

remains to be solved.

Applying the uK-Iteration


of the uK-Iteration does

The

apphcation

yield

convergence.

However,

it

is necessary to

the

slightly modify algorithm. premultiplication of u-curve-fitting transfer function jlj (s) increases the order of the
model at each iteration step. This

the

The

design
more

easily leads

to models with

than 200 states.

Unacceptable calculation times and numerical problem is avoided by


a

problems
tion

result therefrom. This

an

order reduc

step after the augmentation of the plant and before the HM design.
balanced truncated realization

The order reduction method utilized is

[6.1].

Experiences
A

typical

course

of the iteration is shown

by Figure
u-

6.8. In the
<

frequency
achieved,

range where the

performance specification

(ja>)

1 is not

the upper bound of the SSV is forced down at each iteration step. After
six

steps, the solution is reached for which


at the

no

further improvement is

possible. If we look
min,
we

frequency

range between

IO-2 and 10_1 rad/


the

discover that the first controller exhibits much better robust


than the final

performance
task, but
not

design. This results from


one.

"flattening"
design

behavior of the uK-Iteration, which leads to any solution of the

necessarily to the optimal

This behavior of the iteration scheme may lead to strange results. Even

if the robust

performance criterion
an

is

achieved, the simulation of the


oscillation. It

closed-loop

behavior may exhibit

insufficiently damped

138

|i-Optimal Controller Design

10

10

10

10

10

10

10

Frequency (rad/min) Figure


6.8:

Convergence

of the uK-Iteration

has to be

emphasized here that such


This
in the

an

oscillation is consistent with the

performance specification.

performance of the closed-loop system is


rather than in the time-domain!
or

specified
Slightly

frequency-domain
removes

increasing the

performance requirement
this

the uncertainty

specifications usually
Another

problem.
area:

problem of the

uK-iteration is the small convergence

The

performance specification We(s) maximum achievable performance, otherwise the


converge. As
a

(see (6.17)) has to be close to the


iteration does not times have to be 2h
on a

last drawback the


a

long computation

mentioned. The CPU time for

design usually exceeds

SUN

SPARC 2 workstation!

6.4 Controller

design with u-synthesis

139

Analytically
In

or

numerically linearized models? developed.


case

chapter

4 two different types of linear models have been

The main differences between these model

types consist of the low


In the

frequency gains and the representation of flow dynamics.


the

of

analytically linearized models, the relative uncertainty


range due to variation of the

in the low

frequency

steady-state operating points

is

essentially smaller (see Chapter 5).


For both

types of models, state-space controllers have been designed. In

order to achieve

an acceptable controller design (oscillation free) with analytically linearized models, the low-frequency gains of the distur

bance

weights (6.18) and (6.19)


to the not

must be

approximately doubled. With


of the output

respect

higher frequency

range the fact of unmodelled flow

dynamics does

dominate the

shape

uncertainty

weights

for the tray temperature

T10

and

T44.

Both weights may be

kept equal for both types of linear models. However, the uncertainty

weight for the temperature

measurement in the middle of the column

T24

has to be increased for the

analytically linearized models due


adaptations
controllers

to the of the

unmodelled flow

dynamics.
the

In accordance with these

weighting functions,
identical

resulting state-space yield nearly closed-loop behavior with a small advantage from using the

numerically linearized models.


Complex
chosen
or

real

uncertainty block

8G
8G
may be
a

Within the structured


as a

uncertainty models, the uncertainty


real

complex or real uncertainty. The choice as

uncertainty

reduces the uncertainty for the entire


tions. To avoid this

frequency

range.

However, the

resulting closed-loop behavior exhibits insufficiently damped oscilla

problem, the performance weights


are

and the distur

bance

weights

must be increased. The

state-space controller designed


to the

with the modified


a

weights

not

superior

design obtained

with

complex uncertainty block 8r.

140

(i-Optimal

Controller Design

\\K-Iteration
The

results

for three temperature

measurements

input vector of the controller may consist either of the pressurecompensated tray temperatures T10 and T44, or of all three tempera
tures. The

temperature T24 is close

to the feed

tray and its

response to

feed flow disturbances is faster than that of the other two temperatures.

Therefore,

an

improved

controller

design should result from this addi

tional temperature measurement.

Several

synthesis attempts have shown that it is possible to increase the


circa 30% up to
1 1+ 20580s

performance specification by
w

100

(6.28)

After the convergence of the

uK-Iteration, the final controllers


a

were

reduced to

an

order 20

using

balanced truncated realization of the

control system [6.1]. The u-plot for the reduced-order controller (Figure

6.9) using the uncertainty, input, and performance weights (6.14)-(6.23)


demonstrate the excellent robustness properties of this controller.

Figure
and

6.9:

Robust

performance

stability

for the

\i-optimal

state-space controller (controller

inputs: T10, T44, T24)

10

10

10

Frequency (rad/min)

An

analysis of the nominal closed-loop system (with plant model G^)


Figure
6.11. The

is

shown in Figure 6.10 and

singular values of the


outputs
for the

transfer function from the reference

signals

to the controlled

(Fig.

6.10 a) y indicate

of interest. The

a good set-point tracking frequency range singular values of the individual transfer functions from

6.4 Controller

design

with

u-synthesis

141

10J

10J

10'

10

Frequency (rad/min)

Frequency (rad/min)

a)
Figure
6.10:

b)
with the

Singular values for the nominal closed-loop system state-space controller (controller inputs: Tig, T44, T24)
b) Transfer functions from disturbance signals
Dash-dotted line:

u-optimal

a) Transfer function from reference signals to controlled output signals


to controlled

output signals

TF
T

solid line:

Tx

10

HUM

IIJJIL

cr-T-rrr

J 1 nil

1I

Mini.

10

10'

10'

Frequency (rad/min) Figure Singular values of the sensitivity function at u for the nominal closedloop system with the \i-optimal state-space controller (controller inputs: Tig,
6.11:

T44, T24)

142

u-Optimal Controller Design

the two disturbance inputs to the controlled outputs


a

(Fig.

6.10 b) show

maximum

of the

sensitivity

to

these disturbances in the midrange the

frequency

range. While in the

high-frequency

sensitivity

is

smaller due to the

low-pass characteristics of the plant, the large


effective compensation in the
at
u

controller gains
range. The

cause an

low-frequency

plot of the sensitivity

(Fig. 6.11)
(6.29)

Su(s)
confirms the

[I

K(s)G(s)]-1
in the
common

good robustness properties

unstructured
a

uncertainty representation. The maximum value (=1.6) guarantees

stability phase

margin of at least 35 [6.4].

The simulation of step responses using the rigorous described in Chapter 3 demonstrates the time-domain. Two disturbances
are

dynamic model
behavior in the

closed-loop

simulated: An increase of the feed


an

composition from
3.6 mol/min.
as

0.8 to 0.9

mol/mol, and

increase in the feed flow


as

by

Figure

6.12 shows the

top and product impurities

well

the control

errors

for these disturbances and for maximum and

minimum feed flow rates. To estimate the

sensitivity

to

errors

in the

manipulated variables,
for the
same

10%

error

of the controller outputs AL and AV


are

test bench has been simulated. The results

represented

by the thin lines

in

Figure 6.12. The steady-state offsets of the product

compositions
tures
on

are

caused

by controlling pressure-compensated tempera

trays 10 and

44 instead of the

product compositions.
espe

The simulation results confirm the

good robustness properties,

cially the low sensitivity

to

errors

in the

manipulated variables.
error

At both

operating points, the overshoot of the control


For the

is small. in feed

an increase compensation of the first disturbance reflux and boilup must be reduced. The second composition

distur

bance is
an

an

increase in feed flow

rate, which has


at

to be

compensated by
various
reasons.

increase in reflux and

boilup (see Figure 6.13). The large difference


even

between reflux and

boilup

steady-state has
a

First, the reflux and feed

are

subcooled and

partial condensation of the


corresponding

vapor flow thus increases the

liquid flow

rates below the

6.4 Controller

design with u-synthesis

143

Ft=0=20
0.020

mol/min
0.020

Ft=0=46

mol/min

I
g
a
o

0.015

0.010

S
o

Top composition
Bottom

Top composition
Bottom composition

composition

0.005
o

0.005
40 o 10 20 Time (h)

10

20
Time (h)

30

30

40

Ft=0=20
0.4

mol/min

Ft=0=46

mol/min

1: 0.2

'

!!
/K W'""}i
^

0.0

-8
a.

0.2

B
0.4

!! ii
*

I
error error

Control Control

T-10

T-44

06

10

20

30

40

10

20

30

40

Time (h)

Time (h)
state space controller

Figure

6.12: Simulation results with for


an

\i-optimal

(controller

inputs: Tjn, T44, T24)


at t=0 h and
an

increase in feed

composition (0.8

->

0.9 mol/mol)

increase of feed flow rate (+ 3.6 mol/min) at t=20 h

Upper plots: Product composition Lower plots: Control error


^

L, V equal

to controller

output
error

AL with +10% error, AV with -10%

144

|i-Optimal

Controller

Design

Ft=0=20 mol/min
70
1

Ft=0=46
140

mol/min

60

\
V
.

/'

/'

^
s

50

Boilup
Reflux

40

\
30 20

V-J
V

10

20

30

40

10

20

30

Time (h)

Time (h)

Figure

6.13: Simulation results with

\i-optimal state-space

controller (controller in

puts:

Tiq, T44, T24)


an

for

an

increase in feed

composition (0.8

0.9 mol/mol) at

t=0 h and
^^-

increase of feed flow rate (+ 3.6 mol/min) at t=20 h


to controller

L, V equal

output
error

AL with +10% error, AV with -10%

trays. Second, the major part of the feed leaves the column

as

top

product.
If we compare the
we

plots for

the minimum and maximum feed flow rate,

recognize

an

essentially slower rejection of the feed composition


at maximum feed flow rate is not

disturbance at the maximum feed flow rate. A distinct improvement of


the

performance

linear time-invariant feedback controller.

possible using a Higher controller gains would


at this

improve the disturbance compensation simultaneously


destabilize the control
rate. A closer look at these

operating point, but


that

loop figures demonstrates


more

at the minimum feed flow

especially
for

at

high
in

feed flow rates the controller response is feed

sluggish

changes

composition but not

for disturbances in the feed flow rate. This fact

6.4 Controller

design

with

u-synthesis

145

is

explained by the

course

of the

manipulated

variables in

Figure

6.13.

An increase in feed

composition

at minimum feed flow rate forces the

controller to reduce the reflux and the

boilup by

=11

mol/min, while

at

maximum feed flow both flow rates have to be reduced

by

=30 mol/min!

Since in practice

step change of feed composition is improbable, the

rejection of feed flow variations has much higher significance. [iK-Iteration results for
When
are

two

temperature measurements
on

only the pressure-compensated temperature


as

trays 10 and 44
to

used

the controller

inputs, it becomes extraordinarily difficult


an

achieve convergence of the uK-Iteration and

oscillation-free closedas

loop dynamics.

For

design

purpose, the

same

weights

in the

previous

design

for all three temperature measurements have been used.


was

The final controller

reduced to order 20

by

balanced truncated of the reduced order

realization of the control system [6.1]. The controller

u-plots

(Figure 6.14) shows

worse

robustness

properties of this
to the controller

controller in the

higher frequency
criteria.

range
as

(compared

with three measured temperatures

input), only just matching the

robustness and

stability

Figure
and

6.14: Robust

performance \i-optimal

stability

for the

state-space controller (controller

inputs: T10, T44)

10

10

10

Frequency (rad/min)

An

analysis of the sensitivity functions (Figure 6.15) exhibits


sensitivity
at the control
error

small

maximum

e, but

an

evidently

reduced

146

u-Optimal

Controller

Design

Sensitivity

at

I
CO

10

10"

10'

Frequency (rad/min)
Sensitivity
at
u

10

10'

10

10

10

10'

Frequency (rad/min)

Figure

6.15:

Singular values

of the

sensitivity functions

at

{upper plot)

and at

{lower plot) for the nominal closed-loop system with the \i-optimal state-space
controller (controller inputs:

Tig, T^

6.4 Controller

design with u-synthesis

147

stability margin
the SSV and the

at

u.

This illustrates the direct

relationship between
measures.

common

unstructured robustness

Nevertheless the simulation results in Figure 6.16 demonstrate


controller

high

performance, paired

with

larger sensitivity

to

input

errors.

If we compare the controllers with 3 temperatures and 2 temperatures


as

input,

we

must state that the "control

qualities"

in the time-domain

are

very similar. As mentioned


a

before, the intuition of a control engineer


measurements due to
a

is to expect

better

performance for more


obviously

faster
to

state-estimation. This is

not the case! It will be


course

possible

give

an

explanation for this result in the further

of this

chapter.

Ft=0=20
0.020

mol/min
0.020

Ft=0=46

mol/min

I
o

0.015
o

0.015

B
a
o

el
o

03

ID

S
o

0.010

a.

S
o

0.010

Top composition
Bottom

Top composition
Bottom

composition

composition

0.005 o
10 20

0.005 30
40

10

20

30

40

Time (h)

Time (h)

Figure

6.16: Simulation results with

\i-optimal state-space controller (controller composition (0.8


->

inputs: Ti0, T^
t=0 h and
-"

for

an

increase in feed

0.9 mol/mol) at

an

increase of feed flow rate (+ 3.6 mol/min) at t=20 h V

L,

equal

to controller

output
error

AL with +10% error, AV with -10%

148

u-Optimal Controller Design

6.5

Design of controllers with fixed


industry PID
or

structure
are

In the process
common.

advanced PID control structures


a

very

Therefore, the implementation of


are

controller

design and
on

its

acceptance

substantially improved
control elements (such

if the design is based


is the

PID

control structures. The

corresponding design objective


as

u-optimal
first-order

tuning

of

simple
a

PID

controllers,

lags)

within

fixed control structure:

arg
K

inf stabilizing

|MM)|| II A II~

(6<,o.ou;

30)

K with fixed structure

The solution of this

design objective is extremely difficult. Because no synthesis methods exist, (6.30) must be solved by a parameter optimiza
tion approach. During this optimization the SSV has to be calculated repeatedly for a number of frequency points. However, the maximum of the SSV may be very sensitive to the number of frequency

points calcu

lated. In order to
can

simplify the numerical treatment, the design objective


a

be

approximated by

summation of the cube of the SSV for all k

frequency points:
k

arg

inf
e

u| {Jr[P, K(0) ] }
i
=

(6.31)

Summing the cube, large values of the SSV have the design objective becomes closer to (6.30).
The calculation of the SSV presumes

much

more

weight and

nominally stable control loops.

Within u-synthesis, the controllers

are calculated by solving an H^ problem, which always guarantees nominal stability. However, during a

parameter optimization, nominally unstable control loops may be gener


ated. Therefore, the the

design objective (6.31)

must be

supplemented with

boundary

condition for nominal

stability:
<0

Re

Kn *,*!. K}

(6.32)

6.5

Design

of controllers with fixed structure

149

A second

boundary condition

is the robust

stability criterion,

which

should be fulfilled for the final parametrized controller:

uA{^[P,K(8)]}
This constrained parameter tial

<1

(6.33)

optimization problem is solved by

sequen

quadratic programming [6.10].


u-synthesis methods, this approach has shown price
of
even a

In contrast to the

high

reliability,

at the

higher computation

times.

However, the

excellent results

justify the effort.


structures

6.5.1

Diagonal PI(D) control

The

diagonal PI(D) control


frequently
used

most

structure (Figure 6.17) is the simplest and composition control structure for distillation

columns. Due to the

high interaction between the sluggish.

two control

loops,

this
or

structure is difficult to tune, and the response to

setpoint changes

disturbances is known to be very

r10

:^j
f>!
+

PID!

Distillation

Tio

Column with

r44

J
i
_

PID2

inventory
control

T44

Figure

6.17:

Diagonal

PID control structure

The

design

model
same uncertainty model as u-synthesis, excluding the temperature measurement

The

design model for this optimization is the

that used for the

150

(i-Optimal Controller Design

on

tray 24. The weighting functions

are

the

same

transfer functions

as

discussed in section 6.3.


Results

for PI control

The matrix transfer function of the

diagonal

PI control structure is

given by

KR

l+TI,s
1

L(s)

TI,s
KR
2

e10(s)

(6.34)

V(s).

l+TI2s e44(s)
TI2s
optimization for the
as

Table 6.1 summarizes the results of the parameter

analytically

and

numerically linearized

column

models,

well

as

for

complex and mixed real/complex u-analysis.


A comparison of the different optimization results shows quite similar

parameters for

complex u-analysis and

mixed

real/complex u-anal

ysis. However,
and the

a numerically analytically linearized models: Using the analytically linearized

significant difference exists between the

models, the time constants

TIj

are

much smaller and the


reason

corresponding
are

low-frequency gains
smaller

are

much

higher. The

for that

the

low-frequency gains designs show


a

of these linear models. An underestimation

of low-frequency uncertainty results therefrom. Simulations with these controller

faster, but insufficiently damped closed-loop


an

behavior. Of course, with

increase in the output

uncertainty of the

Table 6.1: Results for the Model linear ization Numerical

diagonal

PI control structure

U-analysis
Complex
Mixed R/C

KRt
(mol/min/C)
-14.09 -11.27 -11.52 -14.58

TIi
(min)
137 141 49 60

KR2
(mol/minTC)
2.49 3.15

TI2
(min)
34 52

Analytical

Complex
Mixed R/C

6.92
6.36

56
41

6.5

Design

of controllers with fixed structure

151

design model, the design


experiences corresponds

can

be

improved.

This leads to results compa

rable to those obtained with the

numerically linearized models. These

to those of the

u-synthesis.
is not

Consequently, the design with


further discussed. Due to the
mixed

the

analytical linearized models


times

extremely large computation


the

using

real/complex u-analysis
and

and the very similar optimized tuning,


on

the further discussion will focus

optimization with complex

u-

analysis

numerically

linearized models.

The upper bounds for robust

stability and performance (numerically


the

linearized models,
structure
are

complex u-analysis) using

diagonal

PI control

shown in

Figure 6.18. While stability is guaranteed for the frequency


range, the range.

specified
mance

uncertainties and for the entire

perfor

specification performance

is not met in the lower

frequency

However,

robust

is achieved within the upper

frequency range.

1.5

a
1

/T\
RP

\/\
RS^

Figure
and

6.18: Robust

stability

for

performance diagonal PI

control

0.5

\^
10" 10'

CO

10

10"

Frequency (rad/min)

The transfer functions from the reference and disturbance inputs to the

temperature measurements for the nominal closed-loop systems (Figure


6.19) shows
a

high condition number for the tracking behavior within

the most important


the

frequency

range. This

means a

high sensitivity of
inputs.

tracking behavior

to the direction of the reference

152

|i-Optimal Controller Design

io1

r->y

8 io
3

_____^"\

1
S
io

^\
10

,n-2 10

\
10
10

10

10

10

Frequency (rad/min)

Frequency (rad/min)

a)
Figure
6.19:

b)
values for the nominal

Singular

closed-loop

with

diagonal

Pi-controller

a) Transfer function from reference to output signals

b) Transfer functions from disturbance to output signals


Dash-dotted line: TPj, solid line:
T
_.

These conclusions tions

are

confirmed

by the results of the nonlinear simula


as

(Figure 6.20). They demonstrate the sluggish disturbance rejec optimally tuned diagonal PI control. However, performance plot, the
maximum control

tion of the

expected
is suffi

from the robust

error

ciently

small. Another positive result is the small sensitivity to input

uncertainty.

Results The
use

for PID control


of real PID control instead of PI control

gives additional degrees

of freedom for the controller


not

design. Since true differential behavior is realizable, the parameters for PID controllers with a first order lag
are

in series

optimized (real

PID controllers). The

following

transfer

function for the controllers holds:

L(s)

GK1
0

o" e,0(s)
GK2 644(8)

(6.35)

V(8)_

6.5

Design of controllers with fixed

structure

153

Ft=0=20
0.020

mol/min
0.020

Ft =o=46

mol/min

h
0.015

0.015

I
o

6
a
o

^
I V--'

H i v.--

CO

0.010

o.

0.010

6
o

Top composition

\r
0

^
Bottom composition

Bottom composition

0.005
0 10

0,005 20 Time (h) 30 40 10 20 Time (h) 30 40

Ft==o=20 mol/min
0.4 k

Ft=0=46
0.4

mol/min

0.2

i \
I

0.2

f
''

--

0.0

^T^=

I
U CD

0.0

! r>

0.2

!!
-

P.

-0.2

a
-0.4
error error

0.4

ControlI
Control, .It

T-10
T-44

Control Control

error error

T-10
T-44

0,6 0

-0.6

10

20 Time (h)

30

40

10

20

30

40

Time (h)

Figure

6.20: Simulation results with


-

diagonal

PI control for
an

an

increase in feed

composition (0.8

0.9 mol/mol) at t=0 h and

increase of feed flow rate

(+ 3.6 mol/min) at t=20 h

Upper plots: Product composition


Lower plots: Control

error

L, V equal

to controller

output
error

AL with +10% error, AV with -10%

154

u-Optimal Controller Design

with 1+TIs

TI.TD.s2
+

GKi^=KRi

TI^l

sTL,)

(6.36)

The optimal tuning results show the

unacceptably large controller gains

in

high-frequency
can

range. A

high amplification of the

measurement

noise

be avoided

by

various methods:

"Punishment" of high additional

frequency controller output by

weighting functions frequency gains by additional

Limitation of high

boundary conditions
In order to

keep the uncertainty model invariant, the differential


a

behavior of the controller was limited by

minimum bound of 2 min for

the time constants TL of the first-order


constants
are

lags. The resulting tuning

given

in Table 6.2.

Table 6.2: Results for the


Controller
KR

diagonal real
TI

PID control structure


TD

TL

mol/min/C
PID1 -15.97 4.40

(min)
101

(min)
7.41 15.2

(min)
2.00 7.16

PID 2

39.0

Results achieved with

numerically

linearized model and

complex u-analysis

The

u-plots (Figure 6.21) for the diagonal PID control

structure show

an

improvement of the robust performance. However, the design objective


of robust

performance

u-{Jr[P,K(0)]}
is

<1

(6.37)

by far not reached.


a

The simulation results in


as was

Figure
time.

6.22 illustrate the

same

sluggish behavior slight

obtained with the PI controllers. The im

provement is

reduction of the

settling

6.5

Design

of controllers with fixed structure

155

Figure
and

6.21: Robust

performance
real

stability

for

diagonal

PID control

10"J

10"

10

10

Frequency (rad/min)

Ft=0=20
0.020

mol/min
0.020

Ft_o=46

mol/min

A
o

I
S
o

0.015

l
\

!
**"
.'s*-

'_

CO

ft

6
o

0.010

jr

Top composition
Bottom

Top composition
Bottom

composition

composition

0.005
o 10 20 30
40

0.005
0

i,

10

20

30

40

Time (h)

Time (h)

Figure

6.22: Simulation results with


-

diagonal

PID control for


an

an

increase in feed

composition (0.8
-

0.9 mol/mol) at t=0 h and

increase of feed flow rate

(+ 3.6 mol/min) at t=20 h

L,

equal

to controller

output
error

AL with +10% error, AV with -10%

156

u-Optimal Controller Design

An

analysis of the controller's singular values shows large higha

frequency gains despite


TL

limitation of the minimum filter time constant


a

(Figure 6.23). This makes


,2

first-order filter for the reference inputs

Figure

6.23:

Singular

values of the

diagonal

PID controller

'"lO"5

103

10'

101

Frequency (rad/min)

high-frequency controller gains would be possible. However, decreasing high-frequency limits annihilate the
necessary. A reduction of these

improvements achieved

over

the

diagonal PI control

structure.
we

Summarizing the results for the diagonal PI(D) control structure,


can

conclude that this control structure is

absolutely

not suited for

high performance.
6.5.2

PI(D) control structures with two-way decoupling

The major

disadvantage of the diagonal PI(D) control


two control

structures is the interac

neglect of the interactions between the


tions
can

loops. These

be

partially

cancelled

by

use

of

decoupling techniques.
is shown in

simple

controller structure with


can

decoupling

Figure
or

6.24.

The

decoupling elements

be static (static

decoupling)

dynamic

(dynamic decoupling).
The

tuning

of the

decoupling control decouplers


.

structure for
on an

distillation column

is difficult. Often

are

based

inversion of the

plant's
is very

transfer function G(s) sensitive to

The resulting

closed-loop behavior
errors.

input uncertainty and decoupler Skogestad


in

Summarizing the

research results,

[6.17] concludes that (two-way) decou-

6.5

Design

of controllers with fixed structure

157

r10

*"
PIDi
+

10

Distillation Column
with

inventory
PID,

control

l44

Figure

6.24: PID control structure with static

decoupling

piers should

never

be used for

high-purity distillation columns with the decoupling


seems

LV-configuration. On
less sensitive to

the other hand one-way

to be

input uncertainty and should be preferred [6.18].

Results for static two-way


The

decoupling
structure is static
are

simplest decoupling

decoupling.

Here the two

decoupling elements Cj and C2


this structure the
same
are

constant factors. The results for

obtained with the


as

same

weighting functions and with diagonal PI(D) control struc

uncertainty model

used for the

ture.

Table 6.3 summarizes the control with static controller


are

u-optimal parameters for PI


The

and real PID of the PID

decoupling.

high-frequency gains

small

enough that no boundary conditions concerning this

criterion

were

necessary.

The results for the

decouplers

are

somewhat

surprising. They indicate


one-way

that the
us

optimal decoupling

is very close to

decoupling!

Let

examine this control structure in detail:

158

u-Optimal Controller Design

Table 6.3:

u-optimal parameters
static

for PI(D) control with

decoupling
TD
TL

Controller

or

KR

TI

decoupler
1 2

No.

(mol/min/C)
-5.21 3.71 -13.1 4.56

(min)
22.8
46.8 51.6 62.1

(min)
-

(min)
-

(-)
-0.0240
1.11

1
2

7.83
5.11

8.43 3.07

-0.217 1.03

Results achieved with

numerically linearized model and complex u-analysis

The

decoupler parameter C2 is close to one. Therefore the output of the top composition controller causes
or

any variation of
a

simultaneous

increase

decrease of reflux and

boilup

by almost the

Thus this controller shapes the composition


an

same magnitude. profile within the column by

adaptation of the separation. decoupler parameter


moves

The other

C}

is small.
a

Consequently the output


on

of

the bottom controller


In

composition controller has


the composition

small effect

the reflux. This

profile within the column.


an

light of this interpretation, the limited advantage of


no

additional

temperature

measurement in the middle of the distillation column is

easily explained. Since


measurement,
an

setpoint is available for such

temperature

improved feedback
nor

may be calculated neither for the

composition profile
This

for the

composition profile's position.

special behavior of the control system has its significant advan


the

tages for the closed-loop behavior. The u-plots in Figure 6.25 demon
strate

superior

robust

performance

of the

decoupling control
a

structures. While the

optimal tuning

for PI control shows

peak of the

robust

performance plot
values.

within the medium

frequency

range, the PID

control structure shows

nearly flat and significant smaller structured


a

singular

Using

decoupling

control structure, the additional


a

degrees of freedom
controller

in the controller

design allow
in

significantly better

performance, especially

the

important mid-frequency

6.5

Design

of controllers with fixed structure

159

-a
>

,RPa

/
0.5
-

//\
/

-RS-

03

tn

0
10 10

10

10

10

10

10

10'

Frequency (rad/min)
Figure
6.25: Robust

Frequency (rad/min)
PI control

performance and stability for

{left)

and

real PID control

{right) with

static

decoupling

range. Because of the better

performance using PID controllers, the

further discussion focuses


The

on

that control structure.

singular value plots of the loop transfers from the reference and

disturbance

signals

to the

output signals (Figure 6.26) illustrate the

io1

r-y

oitude
C8

10

_^\

"^\\

IO"'

\
10

,n-2
10 10 10 10 10

10

Frequency (rad/min)

Frequency (rad/min) b)
closed-loop system for
PID control

a)
Figure
6.26:

Singular

values for the nominal

with static

decoupling
to
T

a) Transfer function from reference to output signals

b) Transfer functions from disturbance


Dash-dotted line: Tp^,,, solid line:
r
*

output signals
^
~*

^f

160

U-Optimal Controller Design

better controller
are

performance

as

well. The condition numbers of

Tr
and

much smaller than those of the

diagonal PI(D) control structure,

the The

tracking

behavior is

significantly improved. (Figure 6.27)


confirm

simulation

results

improved
the

controller

fundamentally performance. The sluggish behavior has vanished,


the
errors are

and the maximum control

comparable

to those obtained with

diagonal PI(D) control


has

structures. While the

sensitivity

to

input

uncertainty

increased,

it is still small.

Ft=0=20
0.020

mol/min
0.020

Ft_0=46

mol/min

I
o

0.015

g
a
o
XD

6
o

0.010

Top composition
Bottom composition

Top composition
Bottom composition

0.005
o 10

0.005 20
30

40

10

20

30

40

Time (h)

Time (h)

Figure

6.27: Simulation results for PID control with static


in feed

decoupling
an

for

an

in

crease

composition (0.8

>

0.9 mol/mol) at t=0 h and

increase of feed

flow rate (+ 3.6 mol/min) at t=20 h


^^^^^

L,

equal

controller output
error

AL with +10% error, AV with -10%

6.5

Design

of controllers with fixed structure

161

PID Control with

dynamic decoupling

Using lead-lag transfer functions for the decoupler elements

Cx

and

C2

Ci(B)
a

Kci|g
listed in Table 6.4.
control with

(6.38)

dynamic decoupling
a

structure is realized. The additional

degrees

of

freedom allow

further improvement of the control


constants
are

design. The

resulting optimal tuning


Table 6.4:

u-optimal parameters for PID dynamic decoupling


KR
KC TI

Controller

TD

TL

(mol/min/C)
PID1 -22.2
5.68
-

(-)
-

(min)
80.2 59.4
-

(min)
19.6
12.6

(min)
44.8 24.7 7.42

PID 2
CI C2

-0.138 1.07

117
53.0

71.43

Results achieved with

numerically

linearized model and

complex u-analysis

The simulation results exhibits


worse

performance which
a

is

insignificantly

than that of the


more

u-optimal
a

state-space controllers (Figure 6.28).


control structure with
a

However, the

difficult initialization of

dynamic decoupling in

distributed control system is

disadvantage.

6.5.3

PID control structures with one-way

decoupling

The results for two-way

decoupling
are

have shown

optimal

results for
In this

decoupling
section the

structures which

close to one-way

decoupling.
are

optimal

tuning results for one-way decoupling simple


to initialize. In order to

discussed.

This control structure is


control system and

particularly easy to implement in a


keep
the

distributed

decoupler as

simple pling.

as

possible, the discussion is limited

to static one-way decou

162

jl-Optimal Controller Design

Ft=0=20
0.020

mol/min
0.020

Ft=0=46

mol/min

fi

a
0.015
O

!i
i\
-'

'\
{\/~.

'-'

TV
CO

S
o

0.010

Top composition
Bottom composition
Bottom

composition

0.005 o 10 20
30 40

0.005 10 20 Time (h) 30 40

Time (h)

Figure

6.28: Simulation results for PID control with


in feed

dynamic decoupling for


an

an

in

crease

composition (0.8

->

0.9 mol/mol) at t=0 h and

increase of feed

flow rate (+ 3.6 mol/min) at t=20 h


^~~

L, V equal controller output


AL with +10% error, AV with -10%
error

Two different

decoupler

structures

are

possible
a

if we set either
zero.

Cj

or

C2

of the control structure shown in for the two-way

Figure
to

6.24 to

While the results for the


=

decoupling lead
0),
no

us

expect

good performance
case

firstcase(C,
In fact, the

inference is

possible
a

for the second

(C2

0).

optimization results show insufficient performance for the

second

case

(C2

0). Therefore
of the

reversal of the

decoupling control
compo

structure with

shaping

composition profile by the bottom


not lead to results

sition controller and

moving

the

composition profiles position by the top

composition controller does


obtained with the other

comparable

to those

decoupling

structure.
0

The tuning parameters for the controller with C j Table 6.5. The

can

be found in
us

corresponding

u-curves

(Figure 6.29) let

expect

6.5

Design

of controllers with fixed structure

163

Table 6.5:

u-optimal parameters for PID static one-way decoupling


KR

control with

Controller

or

TI

TD

TL

decoupler
1

No.

(mol/min/C)
-10.5 5.35

(min)
45.7 67.4

(min)
2.18

(min)
5.01

(-)
0 1.05

13.4

13.9

Results achieved with

numerically

linearized model and

complex u-analysis

CD

73
>

a
.3
T3
CD

0.5

Figure 6.29: Robust performance and stability for real PID-control with
one-way

decoupling

Frequency (rad/min)

performance somewhere between that of the PI control with static


way

two-

decoupling and that

of the real PID control with static

decoupling.

The simulation results in

Figure

6.30 support this


a

interpretation.

Therefore this controller represents

structure which is

simple and
a

easily implemented

in

distributed control system, distinguished by

sufficiently high

controller

performance.

164

u-Optimal

Controller

Design

Ft=0=20
0.020 r"1

mol/min
'
'

Ft=0=46
0.020

mol/min

0.015

0.015

0.010

0.010

Top composition
Bottom

Top composition
Bottom

composition

composition

0.005 o 10 20
Time (h)

0.005 30
40 10

20 Time (h)

30

40

Figure
an

6.30: Simulation results for PID control with static one-way increase in feed

decoupling
an

for

composition (0.8

->

0.9 mol/mol) at t=0 h and

increase

of feed flow rate (+ 3.6 mol/min) at t=20 h


^^

L, V equal controller output


AL with +10% error, AV with -10%
error

6.6

Summary
PID control
structures

The comparison of the state-space controllers obtained with

obtained

by u-synthesis by u-optimization leads to

surprising results. The frequently heard opinion that state-space controllers are much superior to PID control structures apparently is
not true for this distillation column. The PID control structures with

decoupling
structures

exhibit

nearly the
a

same

performance

as

that achieved with

state-space controller of
are

higher order, provided

that the PID control


u-curves

optimally

tuned. The visual results of the

and

6.6

Summary

165

simulation

plots shall be supported by numerical integral

measures.

For

purposes of comparison, the


t=40h

square of the control

errors

ISE

J
o

[e20(t)+e|4(t)]dt,

(6.39)

and the integral of the

time-multiplied

absolute control

errors

t=40h

ITAE

| C|eio<t>| |e44(t)|]
+ 0

* dt

(6-40>

have been calculated and summed up for both

operating points and all


errors,

controllers. While ISE


ITAE

punishes especially large control


has
a

the

performance

measure

higher importance for the


to

process

industry

because it punishes any

undesirably sluggish disturbance


the result for the state-space
can

rejection. Both criteria, relative

controller using 3 temperature measurements,


value of the SSV (RP) and the value of the
k

be found in Table

6.6. The last two columns in this table state the maximum absolute

optimization criterion

f(0)

u|{^[P,K(0)]}
i= 1

(6.41)

relative to the value for the state-space controller


measurements. The

using

temperature

high correlation

of the ITAE and the is the

optimization by the

criterion

are

obvious. The

single exception

state-space controller

using

temperature measurements, which

may be caused

convergence

problems

mentioned before.

This table

effectively illustrates the high performance achieved with


structures. The

simple and easily realized PID-control

approach

has

proved

to be

an

efficient tool for the

u-optimization optimal design of

controllers with fixed structure.

166

u-Optimal Controller Design

Table 6.6:

Comparison of controllers
Relative

in time-domain

Control structure

Relative
ITAE

Max.

u-

Relative

ISE

2>f (j<)
0.85 1.00

State-space controller,
3

temp,

measurements

1.00

1.00

State-space controller,
2

temp,

measurements

0.82

1.05

1.04

1.54 9.03 3.98

Diagonal PI Diagonal

control

3.13

2.89

2.14

PID control

2.08
2.42

1.87
1.74

1.53
1.13

PI control with static twoway

decoupling

1.55

PID control with static

two-way decoupling
PID control with

1.44

1.23

0.91

1.19

dynamic

two-way decoupling
PID control with static one-way

1.18

1.12

0.88

1.12

decoup. (Ci=0)

1.99

1.51

0.97

1.34

6.7 References
[6.1]

Balas, G. J., J. C. Doyle, K. Glover, A. Packard, and R. Smith:

u-

Analysis
and The

and

Synthesis Toolbox, MUSYN Inc., Minneapolis MN,

MathWorks, Inc., Natick, MA (1991)

[6.2]

Balas, G. J., A. K. Packard, and J. T. Harduvel: "Application of u-

Synthesis Techniques

to Momentum

Management

and Attitude

Control of the Space Station," Proc. 1991 AIAA Guidance, Navi

gation and Control Conference,


[6.3]

New

Orleans,

LA (1991)

Chiang, R. Y.,

M. G. Safonov: Robust Control Toolbox User's

Guide, The Mathworks Inc., Natick, MA (1992)

6.7 References

167

[6.4]

Dailey, R.

L.: "Lecture Notes for the

Workshop
on

on

H and

Methods for Robust Control," IEEE

Conference

Decision and

Control, Brighton (1991)


[6.5]

Doyle, J. C: "Analysis of Feedback Systems with Structured Uncertainties,"IEEProc., 129, Pt. D., No. 6,
242-250

(1982)

[6.6]

Doyle, J. C: "Performance and Robustness Analysis for Struc


tured

Uncertainty,"

Proc.

of the 21st Conference

on

Decision and

Control, (1982)
[6.7]

Doyle, J. C: "Structured Uncertainty in Control System Design,"


Proc.

of the 24th Conference


FL (1985) K.

on

Decision and

Control, Ft. Lauder

dale,
[6.8]

Doyle, J.,

Lenz, and A. Packard: "Design Examples Using

u-

Synthesis: Space Shuttle Lateral Axis FCS During Reentry,"


NATO ASI Series F:

Computer and Systems Science, 34,128-154

(1987)

[6.9]

Enns, D. F.: "Rocket Stabilization


Value Synthesis

as

Structured

Singular
67-73

Design Example,"

Control

Systems, 11, 4,

(1991) [6.10] Grace, A.: Optimization Toolbox


User's Guide, The Math-

Works, Inc., Natick, MA (1990)


[6.11] Lin, J.-L., I. Postlethwaite, and D.-W. Gu: "u-K Iteration: A New

Algorithm for u-synthesis," Automatica, 29,

219-224 (1993)

[6.12] Maciejowski, J. M.: Multivariable Feedback Design, Addison-

Wesley Publishing Company, Wokingham, England (1989)


[6.13] McFarlane, D. C, and K. Glover: "Robust Controller Design

Using Normalized Coprime Factor Plant Descriptions," Lecture

168

u-Optimal Controller Design

Notes in Control and


Berlin (1990)

Informations Science, 138, Springer-Verlag,

[6.14] Packard, A., J. Doyle, and G. Balas: "Linear Multivariable


Robust Control With
426-438 (1993)
a u

Perspective,"

Trans.

oftheASME, 115,

[6.15] Packard, A., and J. Doyle: "The Complex Structured Singular

Value," Automatica,

29

1, 71-109 (1993)

[6.16] Shinskey, F. G., ''Distillation control for Productivity and Energy

Conservation," 2nd ed., McGraw-Hill, New York, 194-203 (1984)


[6.17] Skogestad, S.: "Dynamics and Control of Distillation Columns
A Critical
-

Survey," Preprints of the

3rd IFAC

Symposium

on

Dynamics and Control of Chemical Reactors, Distillation Col


umn

and Batch Processes,

April 26-29, 1992, College Park, MD,

1-25

(1992)

[6.18] Skogestad, S., and M. Morari: "Implications of Large RGA Ele


ments
on

Control Performance," Ind. Eng. Chem. Res., 26, 2323-

2330 (1987)

[6.19] Skogestad, S., and P. Lundstrom: "MU-Optimal LV-Control of


Distillation Column,"

Comp. Chem. Eng., 14, 4/5, 401-413 (1990)

7.1 Introduction

169

Chapter

Controller

Design for Uncertainty

Unstructured
A

Comparison

7.1

Introduction design for the


entire
a

A controller

operating

range of the distillation

column (see

Chapter 6) requires

structured uncertainty model incor

porating two linear models, and


the
ties
can

a huge computational effort. Naturally, question arises what controller performance and robustness proper

be achieved if we

use

simpler design methods, based

on

just one

plant model

for the nominal


or

operating point (Model Gn)

and classical

design methods
A few of these

simple unstructured uncertainty bounds.


are

simpler methods

discussed in this

applied
mance.

in

straightforward
to

manner, and the

chapter. They design results are

are

not

guaranteed
inherent

represent the optimum achievable controller perfor


an

However, the results give problems


of the

impression of the limits and


on

application of design methods based

simpler uncertainty concepts, and they allow optimal


results

presented

in the previous

a comparison with the uchapter. The weighting func-

170

7 Controller

Design for Unstructured Uncertainty

Comparison

tions of the structured


same as

uncertainty model used for the u-analysis previous chapter.

are

the

those used in the

7.2
A

Diagonal

Pi-control
seems

diagonal
are

Pi-control scheme

to be most

frequently

used in

conventionally
lers
tion
a

controlled distillation columns.

Usually

these PI control

tuned on-line. Due to the


we

large

time constants of the

composi
to lead to to
use

dynamics,

cannot

expect this on-line


to the

tuning approach

controller

performance close
as

optimum. The attempt

tuning rules such


results in
an

Ziegler-Nichols

unstable MIMO

loops often closed-loop system, because these tuning loops


into

for the individual SISO

rules do not take the interaction between the two control


account.

While the
use

following

two

simple and model based tuning methods make


a

of the classical

design methods, they try to

interactions. Both methods lead to

pay attention to the loop nominally stable controller design.


at all

However, sufficient stability margins for the closed-loop system

possible operating points

cannot be

guaranteed.

7.2.1

The BLT method

The

Biggest Log Modulus Tuning


is
a

was

proposed by Luyben

in 1986

([7.5], [7.6]). This method Nyquist stability


given by
y(s)
=

multivariable extension of the classical

criterion. The
=

closed-loop system
(s) and
a

(Tr_^y) with a square


PI control law K(s) is

nominal model G (s)

Gu

diagonal

[I

G(s)K(s)]-1G(s)K(s)r(s)

(7.1)

The characteristic

equation of the multivariable system

is the scalar

equation
det(I
+

G(s)K(s))

(7.2)

7.2

Diagonal Pl-control

171

If we

plot (7.2)

as a

function of frequency, the number of right half-plane

zeros of the

closed-loop characteristic equation are determined. In order


plot like the SISO scalar Nyquist plot,
a new

to make this multivariable

Luyben

introduces

function W(s):

W(s)

-l+det(I

G(s)K(s))

(7.3)

The closer this function the closer the MIMO

approaches
is to

the

system

(-1,0) point in the Nyquist plot, closed-loop instability. The design

objective is defined
L.
=

as

20

log

W(J()
1+W(ja

<

2p Voe R+

(7.4)

where p is the number of

procedure
settings

starts

with

inputs/outputs of G(s). The proposed tuning independent Ziegler-Nichols settings for PIloops.
In
a

controllers of the individual control


are

second step these

detuned

by

factor F

K,

Ki

ZN;
=

Tli

TIm

(7.5)

in order to achieve the

design objective (7.4).

Results
The

of the BUT tuning

tuning results for the nominal model


are

GN (s)

of the distillation
was

process

listed in Table 7.1. A detuning factor F of 3.82

necessary

to achieve the

design objective
too

(7.4). The

proportional gain KR of the top

composition controller is

large for satisfactory setpoint tracking and

Table 7.1:

Tuning

constants with BLT-method

Controller

KR

TI

(mol/min/C)
PI1 -47.1 6.74

(min)
95.2 171.8

PI 2

172

7 Controller

Design for Unstructured Uncertainty

Comparison

measurement

noise attenuation. A
same

plot of the structured singular


as

values (with the the

uncertainty and performance weights


illustrates the insufficient robust

used in and

previous chapter)
any

stability

robust

performance of this composition control design (Fig. 7.1).


further

However,

detuning would reduce the low and highcomposition controller


to
an

frequency gains of
insufficient level.

the bottom

absolutely

Figure

7.1:

u-plots

for

diagonal

Pi-control law tuned with BLTmethod

10

10

10

Frequency (rad/min)

7.2.2

Sequential loop closing sequential loop closing


SISO controller is
was

The idea of the

introduced
one

by Mayne ([7.8],
corre

[7.9]). First,

designed for
has been

pair of input and

output variables. When this

design

completed, the

sponding control loop

is closed and the next

pair of input and output

variables is chosen. Thus the interaction between the control taken into account. This
It is

loops
7.2.

is

design procedure

is illustrated in

Figure
be

an

advantage of this method that each single loop


one or

can

designed
draw
may

using classical methods. However, this method has


backs:

some severe

First, the selection of the first


deleterious effect
on

two

input/output pairs

have

the behavior of the remaining


sequence

loops [7.7].

There exists little

help for this

problem. Second, this method

cannot guarantee robustness for the entire

if the

plant G(s)

is not

operating range. Especially diagonal dominant, that means the condition

7.2

Diagonal Pi-control

173

G3(s)

G2(s)

G^s)

Figure

7.2:

Sequential loop closing

iG^aco)] >|Gij(jco)|
is not

VcoeR+

(7.6)

satisfied,

we

have to

expect robustness problems.

Design Results
The

sequential loop closing idea has been applied to the composition For each control problem represented by the nominal model GN(s)
.

SISO
a

loop,

phase margin of at least

60

degrees

and for both controllers

maximum

high frequency gain

of 18 mol/min/C has been


a

required.

The results of both

possible

design sequences and with


to the linear

minimal inte

gral

absolute

error

(LAE) for the rejection of feed composition and feed

flow disturbances (with Table 7.2.


An

respect

model)

are

summarized in

analysis

of the robustness to unstructured

peturbations
and

shows
for the for the

maximum values for the sensitivities of

Se=2.6, Se=2.4

Su=2.1 Su=1.9

Top

Bottom

design

sequence,

and of

and

174

7 Controller

Design

for Unstructured

Uncertainty

Comparison

Table 7.2: Results of the

sequential loop closing


Til
KR2

Design

sequence

KR1

TI2

(mol/min/C)

(min)
101.9 52.6

(mol/min/C)
10.09 8.78

(min)
55.8
214.5

Top

->

Bottom
-

-18.0
-18.0

Bottom

Top

sequence Bottom

->

Top. These stability margins

are

insufficient. The
are

results of the trated

analysis using the structured uncertainty model


7.3. Both controller designs

illus

by the u-plots in Figure guarantee robust performance nor robust stability.

can

neither

a
a

CO

10

10

10

10

to

10"'

10

10"'

10'

Frequency (rad/min)

Frequency (rad/min) b)

a) Figure
7.3:

|l-plots
a) Top

for the
->

sequential loop closing designs


design
sequence
sequence

Bottom
>

b) Bottom

Top design

7.2.3 The

Optimized

robust

diagonal Pi-control

objective of this controller design is a maximization of the distur bance rejection capabilities with the boundary conditions of sufficient stability margins. As a measure the IAE as defined by
LEnd

of the disturbance

rejection capabilities

IAE

[]e10(t)|

|e44(t)|]dt

(7.7)

7.2

Diagonal Pi-control

175

is

suitable

measure.

It is calculated for step responses to feed compo

sition and feed flow rate of the

closed-loop system. If we

tune both PI-

controllers in order to minimize the


erties of the

IAE-criterion, the robustness

prop

closed-loop system form boundary conditions for the Stabihty bounds in terms of the sensitivity at
are

minimum achievable IAE.

the plant input and output

well established. If

we

require
the

phase

margin of at least 35

degrees sensitivity bounds hold

(which is

relatively small),

following

Se(jco) Su(j<o)
The

[I

G(jco)K(jco)]-1

<1.7

VcogR+

(7.8)

[I

KCJoojGGffl)]-1

<1.7

VcoeR+

(7.9)

optimal parameters which


by trial and
error or

minimize the IAE criterion constrained parameter

are

found

either

by

optimization.

Results
The results for this

design approach
previous

are

given in Table

7.3. The

corre

sponding u-plots (Fig. 7.4) illustrate the improved robust stabihty prop
erties

compared

to the

two methods. While

robust

stability, the

robust
a

performance

is

design guarantees substantially worse than the

u-optimal design of

diagonal Pi-controller design (see Figure 6.18,

page 151). An analysis of the controller behavior in the time domain

(Figure 7.5)

shows

extremely sluggish disturbance rejections.

Table 7.3:

Tuning

constants with

optimizing method
TI

Controller

KR

(mol/min/C)
PI1 PI 2
-5.10

(min)
600.0 86.2

4.92

176

7 Controller

Design

for Unstructured

Uncertainty

Comparison

Figure

7.4:

u-plots
method

for

diagonal
op

Pi-controller

designed by

timizing

10"

10

10

Frequency (rad/min)

Ft=0=20
0.020

mol/min
0.020

Ft=0=46 mol/min
"1

,MUMIIIM"MMI

"

h
0.015
o

11
o

!
V

i \
5

g
a
o
TO

i \
\.
Bottom

I
o

0.015

I1,
\
fr

g
a
o

Top composition
Bottom

composition

composition

0.010
o

s
o

0.010

^^^
0.005 0 10
20 30

l"
0.005
,

yj
10 20 30 40

40

Time (h)

Time (h)

Figure

7.5: Simulation results with


an

diagonal

PI controller

(designed by optimiz
0.9 mol/mol) at t=0 h

ing method) for


and
an

increase in feed

composition (0.8

increase of feed flow rate (+ 3.6 mol/min) at t=20 h

Upper plots:
Lower

Product composition Control


V
error

plots:
L,

-^^^^^

equal

to controller

output
error

AL with +10% error, AV with -10%

7.3 Pi-control with

decoupling

177

7.3 Pi-control with

decoupling
a

The basic idea of decoupling is


increase the
more as

reduction of the

loop interactions. design

If we
on

diagonal

dominance of the system, the


a

task takes

the characteristics of

multiloop SISO design problem. However,


previous chapter,
a

emphasized already
sensitivity
a

in the

reduction of the

loop
an

interactions does not increased


mance

automatically imply better control. Due input


errors, the maximum

to

to model and

perfor
may be

of

controller
even

exhibiting sufficient stability margins


compared
to that of
a

strongly reduced
In the

diagonal

PI controller.

simplest

case,
or

as

discussed in this section, the


constant

plant behavior
or

is

"compensating" prepling" matrix. Different approaches for the selection of these interaction
a

altered

by

postmultiplied

"decou

reducing matrices

are

proposed:
a

Davison [7.3] recommends


the

steady-state decoupling of the

process. For

"decoupled" process G* (s) holds

G*u_y(s) =Gu^y(s)G-'u^y(0)
or

(7.10)

G*u_y(s) =G-'u_y(0)Gu^y(s)
With
a

state space

representation of the
to

process, the

decoupling matrix

is calculated

according

G"1u_>y(0)
The choice of tion
a

(CA^B)-1
or

(7.11)

premultiplication

postmultiplication

of this interac

reducing

matrix is another

degree

of freedom for the controller

design.
Mayne [7.9] proposes a reduction of the high-frequency interactions of the plant. The corresponding decoupling matrix is calculated by

G-u_y(j~)

(CB)-1

(7.12)

178

7 Controller

Design

for Unstructured

Uncertainty

Comparison

As before, the choice of

pre-

or

postcompensation has

to be decided

during the controller design. Ryskamp [7.11] suggests


idea of
a a

decoupling

scheme which is based

on

the

composition profile

control: The difference in the

temperature
of the two

deviations should be used to set the reflux ratio, and the

sum

temperature deviations should be used


This scheme is called Another interesting

to set the reboiler heat

duty.

"implicit decoupling."
based
on a

approach,
at

singular value decomposition


Brambilla et al.

(SVD) of the process

steady

state, is

presented by

[7.1]. Let the SVD of the steady-state transfer matrix of the process

Gu_y(0)

be

Gu^y(0)
where U and V
are

UIVT,
X

(7.13)

unitary matrices and

is

diagonal matrix

containing the singular values


compensator
D (at

diag(Oj,a2).

plant-inverting

plant input) according


D
=

to this SVD is the matrix

VZ~'UT

(7.14)

In order to avoid

high sensitivity

to

input

errors

due to the
a

perfect

decoupling

at

steady state, Brambilla


F
-

et al.

[7.1] introduce

matrix F

al+ (l-a)E

(7.15)

and define

a new

compensation
D
=

matrix D

as

VFI^UT
0...1

(7.16)
allows
=

The

single parameter
a

with

continuous shift
a

between

plant-inverting compensator (a
remove

1) and

compensator

which does not

the effect of the


a

directionality
on

of the process

(a

0). The tuning parameter

has to be chosen
in the

the basis of (1) the

magnitude of the assumed


process to the model

errors

model, (2) the sensitivity of the

errors

(Relative Gain Analysis of D), and (3) the

7.3 Pi-control with

decoupling

179

required performance

in terms of reduction of interactions and direc

tionality (Relative Gain Analysis of Design results


The four

G"1^

(0) D).

proposed compensation matrices

are

summarized in Table 7.4.

In order to calculate

described in section

"optimal" controllers, the optimization approach 7.2.3 has been applied to the different compensated
not

plants. However,
lers

it

was

possible

to achieve any

acceptable control

using

the

proposed compensation matrices, except for the SVD-

based compensator. This SVD-based compensator is


almost the
same

distinguished by
obtained
as
a

one-way

decoupling

structure

as

we

result of the

u-optimal decoupling (see Chapter 6).


Table 7.4:

Compensator matrices
Position of com

Type

of compensator

Compensator
matrix

pensator

Decoupling at
co
=

Plant input

or

-0.636 0.168 -0.728 0.195

plant output
Plant input
or

Decoupling at
0)
= oo

0.380 -0.295 0.875

plant output

-0.193_

Implicit decoupling

Plant output

-1 -1 -1 1

SVD-based compen
sation (a
=

0.8)

Plant input

0.901 0.082 0.955 0.391

The parameters of the additional

IAE-optimal PI controllers (with respect


for the

to

an

boundary condition

proportional gains IKRJ

<

18 mol/

min/C)

are

given in Table 7.5. The n-plots for this controller design

(Figure 7.6) demonstrate good


damped
error

However, the simulation results


oscillation at
rate. The

performance and robust stability. (Figure 7.7) show an insufficiently higher frequencies for the minimum feed flow
is
a

robust

damping of these oscillations


change of the reflux
L and

significantly better for a +10%


-10%
error

in the

in the

change of

180

7 Controller

Design

for Unstructured

Uncertainty

Comparison

Table 7.5:

Optimal

PI

tuning constants

for

plant

with

SVD-based Controller

compensation
KE TI

(mol/min/C)
PI1 -18.0 18.0

(min)
47.4 116.0

PI 2

1
S-c

RP1

03

1
0.5

/
-

<7\
-

RS-

**

Figure

7.6: u-plots for SVD-based compensation with optimally

tuned

diagonal

PI control.

s
CO

10

10

10

10'

Frequency (rad/min)

boilup V. These unwanted specification


controllers' controller
in the

oscillations

are

allowed

by the performance
a

frequency

domain!

They require
on

detuning

of the

proportional gains which

the other hand, reduces the

performance.

7.3 Pi-control with

decoupling

181

Ft=0=20
0.020
'"1

mol/min
0.020

Ft=0=46

mol/min

|
o

0.015

\
\
"

g "a
o
en

.J v..
/

I
a
a
a

0.015

/x
i
/

X*

i
:y
\^

p.

s
o

0.010

e
o

0.010

'

Top composition
Bottom

composition

Bottom composition

0.005
0 10
_

0.005 20 Time
30 40 0

10
_

20

30

40

(h)

Time (h)

Figure

7.7: Simulation results with SVD based compensator and


an

diagonal

PI

con

trol for

increase in feed

composition (0.8

0.9 mol/mol) at t=0 h and

an

increase of feed flow rate (+ 3.6 mol/min) at t=20 h

Upper plots: Product composition


Lower plots: Control

error

L,

equal

to controller

output
error

AL with +10% error, AV with -10%

182

7 Controller

Design for Unstructured Uncertainty

Comparison

7.4

H^ optimal design
minimizing design ([7.2], [7.4], [7.7], [7.10]) of multivari

The H^-norm

able controllers have

proved

to be

powerful method for robust, model-

based controllers.

HM Design specification
The

closed-loop system with


with the

the

plant G(s) and the controller K(s),

augmented
W

weighting

functions

Wd (s) We (s)
,

Wu (s)

and

(s) is outlined in Figure 7.8. This scheme is often called S/KS/T-

weighting scheme. The matrix Wd (s) is a diagonal matrix of transfer functions and represents the frequency content of the feed composition,
feed flow rate, and reference

input signals. The selection of these input


same

weights is discussed applied here.

in section 6.3. The

weighting functions

are

zfi(s)

zu(s)

Zy(s)

d(s)rOO-

Figure

7.8:

Augmented closed-loop system with weighting functions


for the H^

design

H^ optimal design

183

All other

weighting functions

are

chosen

as

diagonal frequency-depen
are

dent weights because the performance and robustness properties equal for all channels:

We (s)

diag

[we (s), we (s) ]

(7.17)

Wu (s)

diag [wu (s), wu (s) ]

(7.18)

Wy(s)
The

diag[wy(s),wy(s)]
is

(7.19)

performance
a

of the

closed-loop system

specified

in terms of the

sensitivity function by the weighting function


with
static

We (s).

A first-order
a

lag

gain of 100 has been

specified

to achieve

nearly

inte

grating behavior.
The bandwidth of the

closed-loop system punishes

is limited

by the weighting

function W

(s)

which

the transfer function

T[dT)rT]T_>y
A first-

from the disturbance and reference

signals

to the

plant outputs.

order
A

lead-lag

transfer function is suitable for this task.


a frequency-dependent limitation helps to achieve sufficient stability margins for
u.

weighting of the plant inputs allows


sensitivity function
at

of the control energy and


the

As done with W

(s),

first-order

lead-lag

transfer function has been selected. The

poles and

zeros

of the
e

weighting functions
u

were

adjusted until the

sensitivity functions at

and at

of the
as
a

closed-loop system had attained


the

approximately the
(with 2

same

peak values

u-optimal controller design


and

temperature measurements),

high performance,
1

(jco)

(7.20)

were

achieved. The best weighting functions

are

given by
(7.21)

We(s>

100i+^20i

184

7 Controller

Design

for Unstructured

Uncertainty

Comparison

cl+ 520s Wu(s)=-5T7T3T


,

(7.22)

,1 +1500s wy(s)=0JT+T5T
,

(7.23)

Design results Despite the fact that the singular values of sensitivity functions for the H^- design (Figure 7.10) and for the u-synthesis (Figure 6.15) are nearly
identical, the u-analysis shows significant differences. The p>plots of the
H^ design (Figure 7.9) show much higher peak values in the low and

mid-frequency sensitivity
was

ranges. The simulation results

(Figure 7.11) allow


in the

conclusion with respect to the of the

larger structured singular values: The


to
errors

closed-loop performance
A reduction of this
common

manipulated
errors

variables is
not

large.

sensitivity

to

plant input
scheme.

possible

using the

S/KS/T

weighting

Figure

7.9:

u-plots

for

H^

op

timal controller

10"

10"

10

10

Frequency (rad/min)

H optimal design

185

Sensitivity at

cs

10"

10

10

10

10"

Frequency (rad/min) Sensitivity


r.

at
ji

u
i
i

10'

1iiii

mi

1ji' i

'

j 'il

1ii

11

nj

11 in

1ir-rrrm

1r-i

11 ra

/_>=-

itude
a
2
io

/ /~
__X
10
2

/
10 10 10

"

Frequency (rad/min) Figure


7.10:

Singular

values of the sensitivity functions at

{upper plot) and

at

{lower plot) for the nominal closed-loop

system with the H^ controller

186

7 Controller

Design

for Unstructured

Uncertainty

Comparison

Ft=0 =20
0.020

mol/min
0.020 r

Ft=0=46

mol/min

(mol/ )
osit on
Comp

h.
0.015
v..
..

J\

0.010

If

Af-

\
0.005

Top composition
Bottom

Top composition
Bottom

composition

composition

1,,

0.005 20 Time (h) 30

10

40

10

20

30

40

Time (h)

Figure 7.11:

Simulation results with the


-

HM-controller for
an

an

increase in feed

composition (0.8
(+ 3.6 mol/min)

0.9 mol/mol) at t=0 h and

increase of feed flow rate

at t=20 h

Upper plots:
^^^^^

Product

composition
error

Lower plots: Control

L,

equal

to controller

output
error

AL with +10% error, AV with -10%

7.5

Summary

187

7.5
The

Summary
application of design methods for unstructured uncertainty
to obtain

to the

composition (or temperature) control problem shows that it is extraordi narily difficult
the

performances which
was

are

comparable
a

to those of

u-optimal controllers. Despite the high effort for


not

robust

tuning of

the Pi-control structures, it

possible

to achieve any

satisfactory

result.
Better results
were

obtained using the

H^ -minimization approach.
stability for the

The

resulting
ating
are

state-space controller guarantees


identical to those of the
to

entire oper

range and the

singular values of the sensitivity functions (Se, Su)

nearly

u-optimal state-space controller.


input uncertainty demonstrates
a

Nevertheless, the high sensitivity


the limits of controller

simple

unstructured uncertainty bounds. Even


on an

robust

design based

be very sensitive to

unstructured uncertainty model tends to uncertainty at operating points different from input
a

the design point. The advantages of presented in Chapter 6 are obvious.

|x-optimal controller design

as

7.6 References
[7.1]

Brambilla, A., and L. D'Elia: "Multivariable Controller for Distil


lation Column in the Presence of Strong

Directionality and Mod

el Errors," Ind. Eng. Chem. Res., 31, 536-543 (1992)

[7.2]

Dailey, R.
ods

L.: "Lecture Notes

for the Workshop

on

H and ]i Meth
on

for Robust Control,"

1991 IEEE Conference

Decision and

Control, Brighton,
[7.3]

December 9-10 (1991)

Davison, E. J.: "Multivariable tuning regulators: The feedfor


ward and robust control of
IEEE Trans. Aut.

general servomechanism problems,"


35-47

Control, AC-21,

(1976)
to

[7.4]

Glover, K., and


mal

J. C.

Doyle: "A State Space Approach

HM Opti

Control,"

Lecture Notes in Control and

Information Sciences,

135, 179-218, Springer-Verlag, Berlin (1989)

188

7 Controller

Design for Unstructured Uncertainty

Comparison

[7.5]

Luyben, W. L.: "Simple Method for Tuning SISO Controllers


Multivariable

in

Systems," Ind. Eng. Chem. Process

Des.

Dev., 25,

654-660 (1986)

[7.6]

Luyben,

W. L.: Process

Modeling, Simulation, and Control for


McGraw-Hill, New York (1990) Design, Addison-

Chemical

Engineers,

2nd ed.,

[7.7]

Maciejowski,

J. M.: Multivariable Feedback

Wesley Publishing Company, Wokingham (1989)


[7.8]

Mayne,

D.

Q.: "The design

of linear multivariable

systems," Au-

tomatica, 9, 201-207 (1973)


[7.9]

Mayne, D. Q.: "Sequential design of linear multivariable


tems," Proc. IEE., 126,
6, 568-572 (1979)

sys

[7.10] Raisch, J., L. Lang, und E.-D. Gilles: "H^-Reglerentwurf fur


Zwei-

und

Dreistoffdestillationsprozesse", at, 41, 6, 215-224

(1993)

[7.11] Ryskamp, C. J.: "Explicit


Chemical

vs.

implicit decoupling

in distillation

control," Chemical Process Control II, American Institute of

Engineers,

New

York, 361-375 (1982)

8.1 Introduction

189

Chapter

Feedforward Controller

Design

8.1 Introduction
It is
a a

drawback of feedback control that

corrective action necessitates

deviation of the controlled variables from their setpoints. This disad


can

vantage

be

overcome

by the

use

of feedforward control. A major and

probably the most frequent disturbance of a distillation column is a change in the feed flow rate. Because the feed flow rate is always
measured, it
can

be used

as

controller

input. An appropriately

designed
atures

feedforward controller takes most of the necessary corrective

action before the

product compositions and the controlled tray temper change. However, because of model errors and other unmeasured
a

disturbances

feedforward controller alone will


so

never

be able to

yield

perfect control
Within this

that feedback control will still be needed.

chapter, the design of linear time-invariant feedforward


our

controllers for

distillation column is discussed. The

proposed design

methods take into account the wide operating range of the distillation
column and the unmeasured feed composition.

190

8 Feedforward Controller

Design

8.2 The

design problem
design objective
is
a

8.2.1

The

The

objective of feedforward control

reduction of the control

error

in

presence

of feed flow rate disturbances. The main

problem

is the

nonlinear behavior of distillation columns. The


a

perfect
on

control action for


the actual and

rejection of

feed flow disturbance

depends
may be

measured feed flow rate and the unmeasured feed composition. A controller
others.

design

for

one

operating point

unsatisfactory
a

at any

Consequently,
column.

it is

impossible
the

to

design

perfect
is
a

linear timerange of
a

invariant feedforward controller for the entire

operating

distillation

Hence

controller which improves the for the


worse.

design objective compensation of feed flow disturbances


range, but
never

feedforward

largest possible part of the operating

makes it

perfect solution of this design objective would be


the ideas discussed in
we

an enormous

task.

Fortunately,
If

good Chapter design the feedforward controller simultaneously for the models
results:
minimum feed flow rate and maximum feed

5 lead to very

GR(s) (representing
minimum feed

composition) and GT(s)

(representing maximum feed flow


we

rate and

composition),

obtain

design which improves the


range.

compensation of feed flow disturbances for the entire operating


8.2.2 The

One-step

or

two-step design?
in
a

design of feedforward controllers is feasible either

one-step

design, simultaneously with the feedback controller,


for the

or as a

second step

closed-loop system (Fig. 8.1) [8.3]. The design

of

feedforward
because the affects the

controller for the

feedback controller

open-loop system is not recommended shifts the poles and, consequently,

dynamics of the system.


A
in

u-optimal one-step design using


5 is

the

uncertainty
to the

structure

presented
the

chapter slightly modified by the additional input

tempting. For that

purpose, the

uncertainty

structure is

controller, i.e.,

8.2 The

design problem

191

a) One-step design

xF

F-f-

wF(s)

KF(s)
K(s)

b) Two-step design
A
*

Step

1: Feedback

design
1
A *-

P
1

K(s)

Step

2: Feedforward

design

F-f
r

-*P

wF(s)

Kp(8)
K(s)

Figure 8.1: Design of feedforward controllers


a) Simultaneous design with feedback controller
b) Design
as a

second step for the closed

loop system. The weighted plant P*

may be

simpler uncertainty

structure than the

plant

P.

192

8 Feedforward Controller

Design

weighted feed flow signal. However, this approach has


backs:

certain draw

Convergence

is unattainable

using the uK-Iteration in

our case

Using the u-optimization approach, the one-step design needs significantly more computation time than the two-step design.
(The computing
with k>2.)
cost is

proportional

to {number of

parameters)11
the feed flow
compensa

For

acceptable results,

the

weighting function for


cause a

signals

must be modified: Small

improvements in the
to the

tion of feed flow disturbances

dominance of the reference

and the feed composition

inputs with regard


gains

performance
part result

specification. Very
therefrom.

small

in the feedforward

Consequently, the discussion


controllers for the

is focused

on

the

closed-loop system, i.e.,


is

as a

design of feedforward second design step. Since


by
means

feedforward control does not affect any stability properties of the closed-

loop system,
two

the

design

relatively simple.

It is discussed

of

examples.

8.3

H^-minimization
a

The H^-minimization [8.4] is well suited for

feedforward controller

design.

Before

we use
a

the numerical tools available


a

(e.g., [8.1], [8.2]),

we

have to build up

closed-loop plant with


an

controller K(s). As

previously designed example, the u-optimal state-space controller

feedback

using all
we

temperature

measurements is selected (see section 6.4.3). If

wish to improve the compensation of feed flow disturbances for the plant models GR (s) as well as for GT (s) we have to close the feedback
,

loops
The The

for both models

separately, define Figure


is
a

the desired

performance, and

limit the

high-frequency output
is outlined in

of the feedforward controller 8.2.

KF (s)

design plant

performance weight We(s)

diagonal

matrix of the transfer

functions

we(s)

H^-minimization

193

Wu(8)

Gr(b)
u

K3
KF(S)
uF
Jl

K(s)

6
We(s)

GT(s)
u

K3
Figure
8.2: The

K(s)

tO
a

augmented plant

for

design of the feedforward

controller

KF(s) by H^-minimization

Wp(s)
It demands the

diag[w(s),wp(s),wp(s),w(s)]

(8.1)

same

performance for both column models and both

controlled temperatures. The transfer function

we(s)

is chosen

as

first-order lag with


until

|TF _J

high pole of we (s) is adjusted 1 is achieved. The final transfer function becomes
a

static

gain. The

We(S>
If do not

100TT2380i

(8.2)

we

specify
we

any

high-frequency limits of the feedforward


a

controller output,

obtain

controller with

large high-frequency

gains. This is undesirable because


feed flow fluctuations
a cause

measurement noise and short-time

unnecessarily, large control actions. Using

diagonal transfer function matrix

Wu(s)

for

the

feedforward

controller output uF

according to

194

8 Feedforward Controller

Design

Wu(s)
with the

diag[wUF(s),wUF(s)]

(8.3)

lead-lag transfer function


(s)
0.5
1
+

104s

l+2.5s

(8.4)

controller behavior similar to

first order lag is obtained. The

singular

values of the controller and the transfer functions from the disturbance

inputs
8.3. If

to the control
we

error

(for the nominal model)


8.3 b with

are

shown
we

by Figure

compare

Figure

Figure

6.10

b,

recognize the

significant improvement of the feed flow disturbance compensation


(dash-dotted lines).

10J

10J

10"

10

icr

io

io

io

Frequency (rad/min)

Frequency (rad/min)

a) Figure
8.3: a)

b)

Singular

values of the feedforward controller


to the

b) Singular values of the transfer function from the disturbance inputs d


controlled output

signals

y for the nominal model


Xr,
_,.
>

G^

with feedforward and


r
_.

>

feedback control. Solid line: T

dash-dotted line: Tw

Nonlinear simulations confirm these


interest of

expectations (Figure 8.4).


are

In the
in all in the

consistency, the

same

disturbances

simulated

as

previous chapters. Of
feed

course, the response to the step


one

changes

composition

remains identical to the


error

shown in

Figure

6.12.

However,

the maximum control

during

the compensation of the

H^-minimization

195

Ft=0=20 mol/min
0.020

Ft=0=46
0.020

mol/min

I
o

0.015

I
o

0.015

g
a
o

g
a
o

0.010

P.

0.010

S
o

B
o

O
-

Top composition
Bottom composition

Top composition
Bottom composition

0.005
0
10 20

0.005
30 40
i

10

20

30

40

Time (h)

Time (h)

Ft=0=20
0.4
1
.

mol/min
0.4

Ft=0=46

mol/min

}'
.

J.
-

0.2

(K)

i!
i

0.2

Tempratue

0.0

0.0

-0.2

i
Control

tfl

-0.2

-0.4
error

-0.4
T-10 T-44

Control
Control

error error

T-10
T-44

Control

error

-0.6 0 10 20

-0.6 30
40

10

20

30

40

Time (h)

Time (h)

Figure

8.4: Simulation results with

\i-optimal
an

state space controller


an

(controller in

puts:

Tig, T44, T24)


-

and feedforward controller for

increase in feed

composi

tion (0.8

0.9 mol/mol) at t=0 h and

increase of feed flow rate (+ 3.6 mol/

min)

at t=20 h

Upper plots: Product composition


Lower plots: Control

error

L, V equal

to controller

output
error

AL with +10% error, AV with -10%

196

8 Feedforward Controller

Design

feed flow disturbance is

approximately halved, and for the maximum


even more.

feed flow rate it is reduced

8.4
The

Optimization approach
implementation of state-space controllers
are

in

distributed control
as

system is difficult. Of
well. Most desirable

course, this holds for feedforward controllers

feedforward controllers with

simple

and

easily implementable
The

structure.

singular values of the H, norm-minimizing state-space controller


a

suggest

feedforward controller structure with


to the

first-order

lag and

different

gains for the outputs

reflux L and the boilup V:

KF(s)

KRj
KR,

1 1+Ts

(8.5)

The parameters of this simple control structure


constraint

are

computed by
norm

parameter optimization [8.5]. The objective may be of

different kind: One

possibility is the minimization of the H


_^

of the

transfer function

TF

for the

plant shown

in

Figure

8.2. This

design

objective

has certain

disadvantages, however: degrees of freedom resulting


structure,
it

Due to the few

from
to

using this
obtain
a a

simple

controller

is

not

possible

controller which is close to

design specifications for

wide

frequency

range.

The H^-norm

minimizing parameters strongly depend


=

on

the

allowed maximum for y, with y controller

|TF_^J

If

we

allow y>5

designs
a

with

obtained. But for

large enough gains (KRL, KRy) are performance specification allowing y 1, we


~

attain small controller

gains and the improvement of the distur

bance compensation is insufficient Most of the feed flow disturbances entering this distillation column
are

step changes. Consequently,

we are

able to define

an

appropriate design

8.4

Optimization approach

197

objective
for
a

in the time domain. It is the minimum absolute control

error

step change

in the feed flow rate. The

design objective becomes


E

[T,

KRL, KRV]

arg [T,

inf

(8.6)

KRL, KRV]

with

f {|e10 (tOl
|
1UR

I ^E

le^ (t)|

|e10 (t)\
I
1UI

\eu (t)|}dt.
| ^*I
I

(8.7)

The performance measure E is calculated for a step response plant input F, employing the plant illustrated by Figure 8.5.

to the

'io.

'44t

'10,
"44,

Figure

8.5: Plant structure for the

optimization

of

feedforward controller parameters

If we select the

u-optimal PID-controller with one-way decoupling as the

feedback controller K (see section 6.5.3), and limit the time constant T

by a lower bound of 5 minutes, the following simple optimal feedforward


controller results:

198

8 Feedforward Controller

Design

KF(s)
The
8.6

1.5
=

1
+

2.6 l

5.0s

(8.8)

singular
a.

values of the feedforward controller


8.6 b
we

are

shown in Figure

In

Figure

find the

singular values

of the transfer func

tions

Td

for nominal closed

loop system with this feedforward

controller. It demonstrates the low sensitivity of the feedback and feed


forward controlled distillation column to variations of the feed flow rate.

T,

10'

d -*y

<D

'V
*>

=1

10

0
'

\\
\
\

6
s
1U

/ /

/
/
/

\v
\
\

/
,n2 /
10
10
i

\
10
10

10

10

10

10

Frequency (rad/min)

Frequency (rad/min)

a) Figure
8.6:

b)
structure

a) Singular values of the feedforward controller with fixed

b) Singular values of the transfer functions for the nominal closed loop system
from the disturbances

inputs d

to the controlled
,

output signals

(Feedback
F-*y

and feedforward control). Solid line: T

dash-dotted line: T

The simulation results


deviation of the

(Figure 8.7) demonstrate that the


a

maximum

product compositions for


comparison

step change of the feed flow

rate is very small. A


same

with the simulation results for the

feedback controller but without feedforward control in

confirms the substantial improvements controller.

by

this

Figure 6.30 simple feedforward

8.5

Summary

199

Ft=0=20
0.0201
'

mol/min
'

Ft=0=46
0.020

mol/min

I
o

0.015

s
a
o

a.

B
o

0.010

Top composition
-

Top composition
Bottom composition

Bottom

composition

0.005
o 10

0.005

20

30

40

10

20 Time

30

40

Time (h)

(h)

Figure 8.7: Simulation results with ^-optimal PID controller with one-way decou pling and a simple feedforward controller for an increase in feed composition
(0.8
->

0.9 mol/mol) at t=0 h and

an

increase of feed flow rate (+ 3.6 mol/min)

at t=20 h
~-"-

L,

equal

to controller

output
error

AL with +10% error, AV with -10%

8.5
The

Summary
compensation of feed flow disturbances
in the time domain
can

be

improved by using

feedforward controllers. H,-norm minimization and the minimization


of the control
errors

(for feedforward controllers with


as

fixed structure)
as

are

efficient

design methods. Frequency domain

well

time-domain results demonstrate the obtained


even

pleasing improvements which

are

by a feedforward controller of order one. A comparison

of the ISE and ITAE criteria (see section 6.6) in Table 8.1 demonstrates

200

8 Feedforward Controller

Design

improvements
tured

up to 50%! As mentioned

previously, the

maximum struc

singular

value |i is not

good performance

measure

if we include

the feedforward control in the structured

uncertainty model.

Table 8.1:

Comparison

of controllers in time-domain Relative ISE

Control structure

Relative ITAE
1.0

Max n-

State-space controller,
3

temp,

measurements

1.0

0.85

State-space controller,
3 temp, measurements 0.63 0.51 0.86

and feed forward control PID control with static one-way


1.99

decoupling (C 1=0)
PID control with static one-way

1.51

0.97

decoupling (C 1=0) and simple


feedforward control

1.13

0.87

1.05

8.6 References
[8.1]

Balas, G. J., J. C. Doyle, K. Glover, A. Packard, and R. Smith:

u-

Analysis

and

Synthesis Toolbox,

MUSYN Inc.,

Minneapolis MN,

and The MathWorks, Inc., Natick, MA (1991)

[8.2]

Chiang, R. Y., and M. G. Safonov: Robust Control Toolbox


er's

Us

Guide, The MathWorks, Inc., Cochituate Place, Natick, MA

(1992) [8.3]

Christen, U., M. F. Weilenmann, and H. P. Geering: "Design of

H2 and H Controllers

with Two

Degrees of Freedom," Proc. of

the 1994 American Control

Conference, Baltimore, MA (1994)

8.6 References

201

[8.4]

Glover, K, and J. C. Doyle: "A State Space Approach


mal Control," Lecture Notes in Control and

to

H, Opti

Information Science,

135,179-218, Springer-Verlag, Berlin (1989)


[8.5] Grace, A.: Optimization Toolbox

User's Guide, The Math-

works, Inc., Natick, MA (1990)

202

8 Feedforward Controller

Design

9.1 Introduction

203

Chapter

Practical

Experiences

9.1 Introduction
In simulations the
ronment.

performance of controllers
measurement

is tested in

sterile envi

Lacking

noise, operator actions, and varying


a

environmental conditions, the results of these simulations represent well established basis for

a comparison of different controller designs. However, only the implementation of a controller in the real plant

proves its

performance. While in the literature

great number of design

methods has been tested

proposed and the resulting controllers have been


very few results of
an

by simulations, only

implementation

at

real industrial distillation column have been

reported.

This

chapter complements the simulation results presented in previous chapters with the results of a controller implementation in the distrib
uted control

system (DCS) which is coupled with this distillation

column. The first section describes the

implementation including
use

the

handling

of constraints. Further sections discuss the

of pressure and

compensated temperatures, the controller performance observed,


economic aspects. A short summary concludes the

chapter.

204

9 Practical

Experiences

9.2 Controller

implementation
objective of
any control

In the research field the

design
in
an

is

high
are

controller

performance.

A control

design implemented

industrial

environment must consider many additional aspects. A few of them

listed below.

Simple implementation: As mentioned previously, state-space control


lers
are

difficult to
on

implement

in

DCS. Therefore the control scheme


on

should be based

fixed low order structures, e.g.

PID control

or on

advanced PID control structures.


Robustness: The control

design
drifts,

must

guarantee stability for the entire


time variations due to
corro

operating

range of the

column, including
etc.

sion of trays, transmitter

Easy

to initialize: The switch from manual to automatic control must be easy


to

simple and

understand.

Operators often

are

semiskilled

workmen who cannot and should not be

expected

to have

an

engineering
intensive

background.

complex

initialization

procedure

of

control scheme
an

unnecessarily

increases the risk of

errors

and requires

operator training.

Handling of
column from

constraints: Constraints

are

necessary to

prevent the

flooding, weeping,

overpressure,

overtemperature, etc.

Often it is sufficient to limit reflux and reboiler heat

duty.

Performance: Despite

the

requirements listed above, the performance of

the control scheme should still be

high. proposed within this thesis,


compromise
to

Comparing

the different control schemes

the PID control structure with one-way feedforward controller evolves


as

decoupling including the simple


among all these

the best
is

requirements. The control scheme


1. Initialization

simple

initialize1,
Lactuai.

robust to

of the control scheme (see Figure 9.1): First, the output of the top adjusted to achieve r^
=

composition
bottom

controller in manual mode is

Then the

top composition

controller is switched to automatic mode.

Second, the output of the


=

composition

controller in manual mode is adjusted to achieve rq

Qactuai-

After that the controller is switched to automatic mode.

9.2 Controller

implementation

205

plant uncertainty,
exhibits
a

it allows

simple handling

of constraints, and it

high performance

in simulations.

This control scheme has been

implemented in the DCS installed at the


Eckardt PLS 80E). The controller

plant
are

considered here (i.e.,

an

inputs

estimated tray

to be easier to understand than pressure

compositions Xj, which for the operators have proved compensated temperatures.
are

The

proportional gains of the controllers

easily converted for these

controller inputs.
A scheme of the

implementation

is shown in

Fig.

9.1. The

handling of

constraints is realized

by using the anti-windup facility of the standard

PID controller blocks within the DCS. The

following ideas have been

realized:

If the

setpoint for the reflux controller rR becomes smaller than

its minimum limit

Rmjn,

the top

composition

is allowed to rise

above the setpoint (=> top composition purer than the top composition controller must be

required),
from

and

prevented

windup.

setpoint for the reflux controller rR exceeds its maximum limit Rmax, the top composition is allowed to decrease (=> top
If the

composition less
controller all

pure than must be


case

required), and the top composition


if at

again

prevented from windup. However,

possible, this

should be avoided.
hold for the

Equivalent
controller.

constraints

bottom

composition

This

policy establishes individual loop


as

constraints for the top

composition
we

control

well

as

for the bottom

composition
are

control

loop. Since

have to include the feedforward control and the one-way

decoupling, the
following four

outputs of the

composition

controllers

limited

by

the

signals entering the anti-windup facility of the PID controllers:

RFBl,max
RFBl,min

Rmax_RFF
Rmin
~

(9<1)

RFF

(9-2)

L10

ei +

92<T +

TCorr)
11Q.

PIDl
+
^

"FBI

e3p
PIDR
-

rR
%

Valve

ii
n

R
FBI,
max

Rmax RFF
"

Po
vFBl,min

PlO

Po

^(P51-Po)
Rmin_RFF
Q

Composition
FBI

Kq/R

estimation

P51
Feedback controllers with constraints

Composition

Po

^44

estimation

5o(P51-P0)

Decoupling

0]+02(T +

TCml)
Q

^*
PID2
ii

FB2

i9_
PIDQ

"Valv

0,p +

04p2
n, Q

FB2,
max

Qmax^FF^FBl
Q

_Q

FB2,
mm

Qmin-^FF-^FBl
Figure
9.1: Controller

LAG

VFF

implementation
Q

Feedforward controller

Kqf/rf

FF

9.3

Composition

estimators

207

-FB2,max Q-FB2,min

"

Qmax- FF- -FBl Qmin" ^FF- -FBl

(9-3)

(9-4)

These individual constraints make unnecessary the

configuration of

variable

structure

control

in

the
a

DCS.

However, the maximum


more

constraint of the reflux may lead to

top product quality significantly


much

below the

product specification, which is

undesirable than
simulation

deterioration of the bottom


well
as

product quality. Fortunately, practical experiences have shown that the reboiler heat duty
as

exceeds its maximum limit

Qm,v first.

In this case, the behavior of the

control scheme is identical to that of a with reboiler heat

single composition control scheme

duty

set at maximum

Qmax
duty

and top

composition

controlled

by reflux flow
as

rate.

If the reflux

well

as

the reboiler heat

reach their minimum

constraints, both

products

become purer than desired.

9.3

Composition
problems, the
was

estimators
any partic parametrization of the composition estima
not
cause

While the ular


tors

implementation of the controllers did


correct very troublesome.

In

first step the parameters of the

estimators were calculated by regression of {Tpx} data (see Chapter 2). However, the correlation of the estimated compositions on tray 10 and
44 with the

product compositions analyzed

once

day proved

to be

unsatisfactory.
Hence

operating data
was

were

recorded for two weeks. Since the feed


was

composition
simulations.

almost constant, it

possible

to compare these

measurement data with

tray compositions calculated by steady-state


the
errors a

Minimizing
was

between the estimated and the

calculated tray
sures

compositions,
are

correction of the estimated tray pres

by

20%

necessary to correct the estimates. Since pressure not

sensors on

tray 10 and 44

installed, the

pressures

on

these trays

are

calculated

by

linear

interpolation between top

pressure and

208

9 Practical

Experiences

bottom pressure (see

Fig. 9.1). The

error

in the pressure

compensation
could

might

have been caused

by

this

interpolation.
or

Other

error sources

have been incorrect

{Tpx} data

pressure measurements. Once the

parameters of the estimators had been


tors worked pressure variations' influence

adjusted,

these

simple

estima

fairly satisfactorily. Nevertheless, the compensation


on

of the

the tray temperatures is the limiting

factor for the overall

performance of the control scheme. This will be

shown in

more

detail in the

following

section.

Of course, the effort for the parametrization of the estimators is

high and
could be

the

performance of the
in

control scheme is limited


gas

fairly by them. In

view of these two

points, the installation of on-line


our case a

chromatographs
polymerization
inputs

preferable. However,
chromatograph
or

the

light component polymer


use

izes at temperatures

exceeding
in
a

certain level. Since

plugs
is

gas

short time, the

of pressure compen
as

sated temperatures

estimated tray

compositions

controller

indispensable.

9.4 Controller
The controller

performance
results the recorded deviations of the esti
in the presence of several

performance observed matched the simulation


9.2 and 9.3

quite well.

Figures

mated tray

depict compositions from their setpoints

feed flow disturbances and at two different feed flow rates.

The

large

measurement noise of the estimated

tray compositions is

caused

by the noisy low-pass

pressure

readings
in

in the column bottom.

Using

first-order

element

series

with

the

bottom

pressure

measurement, the noise could be significantly reduced. Unfortunately,


at the time of the installation of the control

scheme, the capacity of the


is

DCS there

was

exhausted. Even for the


space left. As

configuration of this simple element,


capacity available, the bottom

was no

soon as new

pressure measurement will be filtered.

During the recording of these operating data, the setpoints


constant. In

were

kept

Figure

9.2 the feed flow rate


rate at t=0 h
was

was

increased

steps. The feed flow

260 1/h (49

by mol/min), while the

401/h in four

9.4 Controller

performance

209

Feed

Tray

10

40

60 Time (h)

100

Tray

44

L5

>

;,

H-osS 8
.

Ml
53S
,*
*

ik'3.' MLi >S Jss ieS ills


165 SflftSf WBt3B%jff"WB5f t HB'Sff

t??^

?'

"

jj'

j;

0.5

o
i-t

a.

+a

a
0)

CO

"g
20 40 Time (h) 60

80

100

Figure

9.2: Recorded

one-way

operating data with installed PID control decoupling and feedforward control.
Deviations of estimated tray

scheme

including

Top:
Middle:

Deviation of feed flow rate from 260 1/h (49 mol/min)

composition and
on

of pressure of pressure

compensated temperature
Bottom:

from setpoint

tray 10

Deviations of estimated tray

composition and
tray 44

compensated temperature

on

210

9 Practical

Experiences

Feed

P 3

30

40

Time (h)

Tray

10

-a

-0.5

g
a
<a

a,

0.5 V 30 40 70
P

Time (h)

Tray

44

13

-0.5

l^^diAilw
a
-

05

>
a)

10

20

30

40

50

60

70

Time (h)

Figure

9.3: Recorded

operating data with installed PID control scheme including


and feedforward control.

one-way

decoupling

Top:
Middle:

Deviation of feed flow rate from 170 1/h (32 mol/min)

Deviations of estimated tray

composition
setpoint

and of pressure

compensated temperature
Bottom:

from

on

tray 10

Deviations of estimated tray

composition
tray 44

and of pressure

compensated temperature

on

9.4 Controller

performance

211

feed

composition
was

was

approximately

0.85 mol/mol.

Although boilup

the feed

flow rate

out of the

design
by

range, the reflux and

remained

within the range covered


flow rate at t=0 h
once
was

the controller

design.

In

Figure 9.3, the feed


was

170 1/h (32 mol/min) and it

increased

only

by 101/h.
errors

The control

in presence of these feed flow disturbances remain


to

extraordinary small. In fact, it is almost impossible


control
error

separate the
of this

from the measurement noise and the effect of all other

unknown disturbances. This proves the advanced PID control scheme.


The

high performance

simple

advantages of the controller implementation are demonstrated best


comparison of the product compositions analyzed
once a

by

day before
shown

and after the installation. At the


bottom

beginning

of this

project, top and


are on

composition were controlled manually. The results

the left-hand sides of Figure 9.4 and 9.5.

Obviously,

the average

product

compositions
the The

are

found far from their


are

setpoints,

and the variations of

product compositions

very

large.
analysis results

right-hand sides of Figure

9.4 and 9.5 show the

beginning after the adjustment of the composition


the variations of the
average

estimators.

Clearly,

product compositions
are

are

much smaller and the

product compositions

close to the desired results. However,


as

despite the high performance of the control scheme Figure


can

illustrated

by

9.2 and

9.3, significant variations of the product compositions

still be detected. Please remember that pressure measurements of


are

tray 10 and 44

lacking. Therefore the influence of the large pressure


adjustment of the controller setpoints, the results will
extensive

variations (bottom pressure: 120-190 mbar) to the tray temperatures


cannot be

perfectly compensated and


on

an

setpoints depending presented improve


are

the feed flow rate is necessary. Since the results

achieved with almost constant

even

further

as

the operators

gain

more

experience

with the setpoints.

212

9 Practical

Experiences

Manual operation
0.3

Controlled

xxx
~

0.25

Xx
x

0.2
X

Xx

0.15

xx

x x X

XX

0.1
x
X

x*-xx

"

0.05

X X x

XX
*

>**
35

^
70

Days

Manual

operation
0.25
r

Controlled

0.25

0.15

0.05

Figure

9.4:

Analysis data

of top and bottom

product

Top: Top composition 1-xjj


Bottom: Bottom

composition

xg

Dashed line: Average

composition

9.4 Controller

performance

213

Manual

operation

Controlled

0.1

0.2

0.3

0.1

0.2

0.3

Top composition (mol/mol)

Top composition (mol/mol)

Manual

operation

Controlled

0.1

0.2

0.3

0.1

0.2

0.3

Bottom composition (mol/mol)

Bottom composition (mol/mol)

Figure

9.5:

Histograms

of analyzed

product qualities

Top: Top composition 1-xjj


Bottom: Bottom

composition xg

214

9 Practical

Experiences

9.5 Economic aspects


The management decision for
or

against the installation of

control
some

system depends primarily

on

the economic
In

feasibility

and to

degree

on

ecological improvements.

our case

the installation of the

control scheme

yields

the

following

most

important improvements:

More uniform
=>

product qualities
necessary

Less

overpurification

=>

Energy savings (which


mean

is

an

ecological advantage, too)

Reduced
=>

of light component in bottom


a

More top

product with

market value of

>

250000

$/a

Increased maximum column load


=>

The installation of
are

an

additional column

can

be avoided
a

These pay-offs

complemented by side effects, for example

deeper
as a

understanding
plant.

of column
a

dynamics by the operating staff,

which

consequence achieved

better

operation of other columns in the

same

These financial benefits must be

weighed against the investment

costs.

Hardware and software expenses

exclusively

for this

project total
a man

approximately 50000 $. It is

not unreasonable to estimate the necessary

engineering effort for

similar

project

to be less than half


very

year.

Therefore the economic benefits

are on a

positive

side.

9.6

Summary
implementation of the PID control
on

The results of the


way

structure with one

decoupling

and feedforward control


of this

the real

plant confirm the

high performance
tions. The main

simple control scheme indicated by simula

problem of the implementation was, except for over coming high psychological resistances, the correct parametrization of
the

composition estimators.

A solution of this

been

possible without

an

extensive
use

problem never would have comparison of simulation and oper


of pressure

ating data. Nevertheless, the

compensated tempera-

9.6

Summary

215

tures

or

estimated tray compositions remains the achieved

limiting factor of the


economic

overall

performance of the implemented control scheme. The

advantages
effort

by

this

simple

control scheme exceed the financial

by far.

216

9 Practical

Experiences

10.1

Introduction

217

Chapter

10

Conclusions and
Recommendations

10.1 Introduction
This thesis treats all the necessary steps for

composition control

design
a

for

an

industrial

binary

distillation column. Each of these steps

produced
sion, they

new

insights

into various

aspects of the control design. Since


a

chronological discussion of these steps would lead to


are

thematic confu

summarized in the four sections

Controller

synthesis
or

State-space

PID control?

How many temperature measurements? Column models


a

This thesis does not presume to present


control

final solution to all distillation


up

problems.

The ideas

presented

come

against

many gaps in

research, limits of distributed control systems, and


ation between

problems of cooper
section the most

industry and university. In the last topics


are

important

aspects of these

discussed.

218

10 Conclusions and Recommendations

10.2 Controller

synthesis
design of robust controllers for the dual
compo

This thesis discusses the


sition control

problem
are

of an industrial

binary distillation column.


over
a

Distil

lation

columns

usually operated
rates.

wide

range

of feed

compositions and feed flow


antee

Consequently,
not

controller must guar


at
a

stability and

high performance
operating

only

single operating

point, but for the


The
common

entire

range of the distillation column.

tured

design methods are based on unstruc uncertainty models, for example a multiplicative uncertainty at
estimate of the
are

robust controller

plant output. An

corresponding uncertainty bounds has large


to allow any controller

shown that these bounds Nevertheless


a a

too

design.
on

solution of the

design problem
a

is

possible.

It is based

structured uncertainty model which to


an

large

extent avoids the

unnec

essary conservatism of

unstructured

uncertainty description.
as

This

model treats the nonlinear column behavior

several simultaneous

uncertainties and quite well describes the column

dynamics for all

oper

ating points within the predefined operating

range.

Utilizing this uncertainty model, priate design methods


space
are a

feedback

controller

synthesis
state-

requires the framework of the structured singular value


the uK-Iteration for the

\i. The appro

synthesis of

controllers and

constraint parameter

optimization for the

synthesis of controllers
mance

with fixed structure. These methods lead to

feedback controllers which


and

are distinguished by a high controller perfor stability within the entire operating range, guaranteed

paired with

low

sensitivity

to

errors

in the

manipulated variables.

A drawback of this

design approach is the high effort for uncertainty

modelling

and

computation of the controllers. In principle, comparable

results could be obtained and the

computational effort could be signifi


on

cantly
these

reduced by using

design methods based


are

arbitrary small

unstructured uncertainty bounds. However, it has been shown that


common
as

design methods

not well suited for ill-conditioned

plants such

high-purity distillation columns.

10.3

State-space

or

PID control?

219

The ideas of the feedback controller feedforward control

synthesis

can

be extended to the

design.

A simultaneous controller

design for the


using HMoptimization

closed-loop

models at maximum and minimum column load


or

minimization (for state-space feedforward controllers)

in the time-domain (for feedforward controllers with fixed structure)

yield controllers, which greatly improve the compensation of feed flow


disturbances.
The theoretical and simulation results the
are confirmed by the results of simple PID control structure with

practical implementation of

one-way

decoupling and
controller

satisfactory

a simple feedforward control scheme. The very performance achieved without any expensive on

line composition

analyzers leads

to

high

which justify the effort of the control

economic and ecologic benefits design and implementation.

10.3
A

State-space

or

PID control?

comparison of the different state-space controllers with optimally


an

tuned advanced PID control structures has demonstrated

unex

pected result:

The performance of u-optimally tuned advanced PID control


structures is

only insignificantly

worse

than the

performance

of

high-order state-space controllers


This statement is of great

significance for industrial practice. It holds

for the feedback

as

well

as

for the feedforward control


a

design. The imple

mentation of advanced PID control structures in

distributed control

system requires
be

much less effort than that of state-space controllers and

increases the acceptance of the control

design by the operators. It


settings.

must

emphasized, however,

that the

high performance

of the PID control

structure is achieved with unconventional controller

The

optimal tuning of PID control


an

structures with

decoupling for this


in

distillation column caused

additional insight. The optimal controller


an

performance
essence

is achieved with

implicit decoupling scheme where

220

10 Conclusions and Recommendations

the bottom

composition

is controlled

by moving the composition

profile,

and

the top composition is controlled

by intensifying

or

weakening

the S-form of the composition profile.

Since the position and

shape

of the

composition profile
a

at

steady-state
it

depends essentially

on

the actual and unmeasured feed

composition,
variables.

is difficult to make any inference from

composition

or

temperature

measurement in the column middle to the

manipulated

Very similar considerations hold for the relative performance


state-space controllers. For the
same

of the

reason,

the estimation of the


no

composition profile by
better
controller

the inherent observer has

advantages. The
of freedom in the
in the low- and

performance results only from the higher degree


design,
which allows
a

higher performance

mid-frequency range without destabilizing the closed loop system in the


high-frequency range.

10.4 How many temperature measurements?


A

comparison of a control design including


a

temperature measurement
this measurement

in the middle of the column with

design excluding

leads to the following statement

Additional temperature

or

composition

measurements in the

middle of the distillation column have


on

no

significant

influence

the maximum controller

performance.

The

reason

for the very limited

advantage of additional temperature


is their unknown

measurements for the control

design

setpoint, which
two pressure-

depends
mance

on

the

actual, unmeasured feed composition. The high perfor


can
or

of the control design

be achieved with
two

just
tray

compensated

temperatures
control

estimated

compositions.

Dispensing
costs

with additional measurements reduces

the installation

of the

system and increases its economic viability.


are

However, if regression models

used to estimate the

product

compo-

10.5

Column models

221

sitions

based

on

temperature and flow measurements, additional


are

temperature

measurements

of great

advantage.

10.5 Column models


All results of this thesis distillation column.

are

based

directly or indirectly on models of the


of the
are on

Especially the model-based adjustment


However, the control design
or

composition estimators clearly proved that such

process models may be based

absolutely

necessary.

linear models that include


Within the structured

exclude flow

dynamics.
a

uncertainty model,

multiplicative uncertainty

is included for each measured tray temperature, whose

bounds exceed 100% for

uncertainty frequencies above 1/16 rad/min. Since the flow


range, the

dynamics affect
justified:

the

high-frequency

following

statement is

Including or excluding flow dynamics


is

in the linear models

insignificant for the controller design.


an

This has

impact
at

on

the

design effort.

If

controller

design

can

be

based
tion
not

on an

analytical linearization of a simple model for the composi

dynamics

particular steady states, steady

rigorous dynamic model is


a

absolutely

necessary. The

states of

column may be calcu


as

lated with

common

flowsheeting

programs such

ASPEN

PLUS

or

PROCESS

and the controllers

designed

can

be tested

using a simpli

fied nonlinear model without flow

dynamics.

10.6 Recommendations

10.6.1 Academic research

Multicomponent distillation: The results of this thesis


these results to other

are

based

on

the

example of a single binary distillation column. While the adaptation of binary columns is expected to be straightforward,

222

10 Conclusions and Recommendations

the

uncertainty modelling of multicomponent distillation columns

needs additional research.

[i-synthesis:
singular
for

The robustness

analysis of controllers using the structured


a

value ^i has shown to be


convergence

reliable and

outstanding
insufficient.

tool.

However, the
robust

properties of the corresponding algorithms


are

u-synthesis (DK-iteration, uK-iteration)

More

algorithms

are

absolutely

necessary.

Decentralized control:

Generally, the design of robust controllers with


an

simple

structures is at

early stage of development.

In the

case

of this

distillation column, it

was

relatively

easy to propose

potential control
many
more

structures and to solve the


eter

design objective optimization (u-optimization). However, dealing with


a

with

constrained param

control

loops simultaneously,

solved. For

problem example, the high performance


the

of the

loop pairing

is still not

of the controllers in this

thesis has been obtained


common

using the LV control configuration. Since (single loop pairing of


to minimize the inter

methods for control structure selection

controlled variables and

manipulated inputs) try

actions between the individual control

favor other control

loops, certainly these methods configurations. Therefore methods for the selection
are

of control structures

necessary, which include

simple multivariable
are

control schemes. Similar arguments hold for the controller tuning. The
current methods for the

tuning of multiloop SISO control schemes


or

known to be either very conservative methods would be very desirable.


10.6.2 Decentralized control

else to lack robustness. Better

systems

Today

control engineer in the research field is familiar with modern


as

and flexible software tools such


contact with
one,
arouses
a

MATLAB or MATRLXX.
even

His first
modern

decentralized control system (DCS),


of

with

feelings

working

in the
a

analog computing
seems

era.

The

replacement
idea for the

of the old consoles with

computer

to be the

only
a

development

of the DCS. The inherent

possibilities for
are

faster,

more

flexible, and simplified controller implementation

not

exhausted yet.

10.6

Recommendations

223

10.6.3

Cooperation industryuniversity
complains of the inadequate cooperation between
a

Often the industry

university and industry. Some typical problems of such


the

cooperation

were encountered during the course ofthis project. The main problem is

divergence between the interest of the partners in the project.


are

University researchers
tion of the actual
are

interested in

deeper insights into basic prob

lems and their solution, while the process

industry wants

rapid solu

industry problem. Additionally chronically overworked with everyday problems, thus unable to spend enough time to concern themselves with such a project. This leads
the contact persons in
an

to

insufficient flow of communication. different

Consequently,

both partners

speak

stand the industrial

languages: needs,

the university researcher does not under


while the industrial

counterpart does

not

understand the mathematical methods. Therefore it is of


tance that

high impor

the aims and

responsibilities of both partners in the project


as

are

spelled

as

clearly

possible

at least

one

control engineer of the industrial partner actively

follows the progress of the project

points could be kept in mind, industry and university could be avoided.


If these two

many

problems between

224

10 Conclusions and Recommendations

Curriculum vitae

Name Date of birth Place of birth

Hans-Eugen Musch
June 19,1965

Freiburg
German

im

Breisgau, Germany

Nationality

1971-1975 1975-1984

Primary school
Humanistic
at

gymnasium Kolleg St. Sebastian

Stegen

near

Freiburg

1984

Abitur

1984-1985

Military service

1985-1989 1989

Chemical
Masters

engineering studies
in Chemical

at the ETH Zurich

degree

Engineering CDiplom")

Since 1990

Research assistant at the Measurement and Control

Laboratory, ETH Zurich

You might also like