You are on page 1of 21

International Journal of Control

ISSN: 0020-7179 (Print) 1366-5820 (Online) Journal homepage: http://www.tandfonline.com/loi/tcon20

A unified derivation and critical review of modal


approaches to model reduction

D. BONVIN & D. A. MELLICHAMP

To cite this article: D. BONVIN & D. A. MELLICHAMP (1982) A unified derivation and critical
review of modal approaches to model reduction, International Journal of Control, 35:5,
829-848, DOI: 10.1080/00207178208922657

To link to this article: http://dx.doi.org/10.1080/00207178208922657

Published online: 12 Mar 2007.

Submit your article to this journal

Article views: 35

View related articles

Citing articles: 36 View citing articles

Full Terms & Conditions of access and use can be found at


http://www.tandfonline.com/action/journalInformation?journalCode=tcon20

Download by: [RMIT University Library] Date: 19 June 2016, At: 22:49
INT. J. CONTROL, 1982, VOL. 35, No.5, 829-848

A unified derivation and critical review of modal approaches


to model reduction

D. BONVINH and D. A. MELLICHAMPt

Six of the most commonly-used modal redu'ction techniques are represented by


means of signal flow diagrams connecting inputs and state variables. In this repre-
sentation, the derivation and the cross-comparison of the various methods are greatly
facilitated. Each method is then analysed to determine key characteristics, i.e.
[i} steady-state agreement, (ii) initial-value agreement, (iii) independence of the
reduced model on the choice of retained state variables and inputs. and (iv) eigen,
vector orientation. Realistic, large-scale systems are used to evaluate the static and
dynamic characteristics of each model reduction method numerically. These studies
Downloaded by [RMIT University Library] at 22:49 19 June 2016

reveal Borne important, and often unrecognized, weaknesses of several traditional


techniques and confirm the st.rengt.h of the newly-developed technique of Lite.

Symbols used
A (n X n) plant matrix
B (n x r) input matrix
C" C 2 first, second effect glycol concentration
diagonal gain matrices in Davison's second-modified technique
D l • i , D 2•i
E
matrix relating %1 to Z2 in the Litz technique
(n x r) canonical input matrix
G
H, H-l (n x n) right, left eigenvector matrices
Hl
first effect enthalpy
I identity matrix
m order of the reduced model
M matrix defined by (46)
n order of the large-scale model
N matrix defined by (47)
P, Q matrices defined by (57)
Q weighting matrix
r dimension of the input vector
R matrix defined in the Table
S matrix defined by (50)
T matrix defined in the Table
r-dimensional input vector
matrix defined by ~51)
(n x n) left eigenvector matrix (V = H-l)
first, second effect hold-up
n-dimensional state-variable vector
n-dimensional canonical-variable vector
(n x n) eigenvalue matrix

Received 30 October 1981.


t Department of Chemical and Nuclear Engineering, University of California,
Santa Barbara, California 93106, U.s.A.
t Present address: Abteilung fiir "Prozessautomation, Sandoz AG, Basel,
Switzerland.
002l}-7179/82/S,0, 0829 $04-00 <0 1982 Taylor & Francis Ltd
830

Superscripts
indicates a time derivative (x)
denotes an approximate variable (xd
A denotes an intermediate approximate value (~l,i)
T indicates a transposed matrix (G?)
indicates a complex conjugate (G l )

Subscripts
I refers to a dominant subset, Xl
2 refers to a non-dominant subset, X2
R refers to a reduced model, A R
S refers to a reduced model, Bs
i refers to the ith input (gl, i)
Hold lower case letters indicate vectors (u)
Downloaded by [RMIT University Library] at 22:49 19 June 2016

Bold capitals indicate matrices (A)

I. Introduction
Because of its importance in systems analysis and in the design of controllers,
model reduction methods have received considerable attention over the past
20 years. The review paper of Genesio and Milanese (1976) indicates how
many techniques have been proposed and discussed in the literature.
Among the various classes of reduction methods, modal techniques have
continued to be of interest since they permit explicit (or closed-form) formula-
tion, i.e. the reduced model is derived directly from the linear large-scale
system through algebraic relationships. The need for analysis of input/output
data and for the use of optimization procedures, such as is the case with
least-squares techniques, is simply avoided. However, several key decisions
need to be made before any modal method can be used efficiently; in parti-
cular, the user must predetermine:
(i) an appropriate order for the reduced model,
(ii) the dominant modes of the original system, i.e. the modes to be
retained in the reduced model, and
(iii) the most sensitive state variables, i.e. the state variables to be retained
in the reduced model.
A satisfactory resolution of these three points can be obtained quickly
using the analytical methods developed by Bonvin and Mellichamp (1982)
which are based on a generalization of the structural dominance method of
Litz (1979 a; b, 1980). A remaining unresolved point, then, is the choice of
a specific model reduction method.
Davison (1966) proposed one of the first modal reduction techniques. As
pointed out subsequently by Chidambara (Chidambara and Davison 1967 a)
the method does not provide for steady-state agreement between the dominant
state variables of the original and reduced models. Further arguments of
Chidambara and Davison (1967 b, c) led to several variations of Davison's
original approach: Chidambara's first method, Chidambara's second method
and Davison's' first-modified' method. A year later, Davison (1968) pro-
posed his' second-modified' method.
1Hodal approaches to model reduciion. 831

Simultaneously, Marshall (I!166) developed a technique which preserves


the steady state of the original system by exciting the modes in the reduced
model differently from those in the original system.
Chidambara (1969) then presented two modal techniques which are optima]
in the sense that they minimize deviations between the original and reduced
models. However the methods do not result in explicit formulations for the
reduced models and, therefore, are not considered here. Later, Fossard (1970)
proposed a modification to Davison's original method which ensures both
initial and final (steady-state) agreement between the original and reduced
models.
In a review and comparison of the above techniques made by Wilson
(1974) Chidambara's first method was shown to be equivalent to Davison's
first-modified method, and Chidambara's second method equivalent to
Marshall's approach. Because of the quite different ways in which the
Downloaded by [RMIT University Library] at 22:49 19 June 2016

equivalent methods had been derived and because of the presence of algebraic
identities in the final expressions which hindered comparisons, some of these
equivalences were originally overlooked. Several additional techniques have
been developed and described in the literature (Genesio and Milanese 1976),
but since they mainly represent minor extensions to one or another of the
procedures mentioned above, they are not discussed in this study.
Just recently, Litz (1979, 1980) proposed a new modal technique which is
optimal in the sense that the integral of the square of the errors between the
dominant state variables in the original and approximate models is minimized.
Unlike Chidambara's technique (1969), the new approach provides an explicit
formulation for the reduced model.
In the present study we develop a graphical representation of each model
reduction technique, i.e. of the signal flows from the inputs to the outputs
of the system. As a result, each of the techniques is derived from a unified
(common) viewpoint. The derivations may differ substantially from the
original ones, but corresponding reduced models are equivalent. The use of
a signal flow representation facilitates comparison of the techniques. Equally
important, it furnishes a clear visualization of the assumptions involved in
developing each technique and, as well, provides insight into the advantages
and disadvantages of each approach.

2. Signal flow representation and derivation of the techniques


A diagram is used to represent each of the dynamic models in modal
form; the various methods can then be derived easily by inspection and
interpretation of the signal flows.

2. I. Original model
Consider the deterministic, continuous, linear system

x=Ax+ Bu (1 )

where xe R», UER' and A and B are matrices of appropriate dimensions. A.


zero initial condition is assumed, in accordance with the usual model reduction
approach
x(O) =0 (2)
.832 I.!.. Bonvin and D. A. Mellicharnp

On transformation to Jordan canonical form, using the transformation


X= Hz (3)
the system becomes
z=Az+Gu (4)
with
z(O) = ° (5)

The essence of all modal approaches to model reduction consists of neglect-


ing the dynamics associated with fast modes, i.e. those which die out quickly
when perturbed. Hence, it is representation ally useful to partition the above
relationships in terms of dominant and non-dominant modes, as well as
important and less important state variables. A proper partitioning is
accomplished routinely with the help of the generalized structural dominance
Downloaded by [RMIT University Library] at 22:49 19 June 2016

analysis proposed by Bonvin and Mellichamp (1982) so as to determine, among


other things, the dominant modes, the most sensitive state variables, and an
appropriate order for the reduced model. Hence, the dynamic system given
by (3)-(5) can be partitioned as follows

Xl]
[X z
= [H ll H1z][Zl] (3 a)
H Z1 H zz Zz

1 1
[ ~1]
Zz
= [A °][Zl] + [G ] U
° Az Zz Gz
(4 a)

and
Zl(O)]
[zz(O)
= [0] (5 a)
°
where subscripts 1 and 2 refer to '.dominant ' and' non-dominant' quantities
of dimension m. and (n-rn), respectively. The dynamic system expressed by
(3 a )-( 5 a) is depicted in a form of signal flow 'diagram in Fig. I ..

~1 (0) =Q

+ 1\1
+

Figure 1. Original model.


Modal approaches to model reduction 833

The coupling between the inputs and the states of a linear system can be
altered so as to reduce the dimensionality of the dynamic system from n to
In, yielding an approximate model in which the (n-m) non-dominant modes
associated with A z are neglected. This procedure is standard for any modal
approach to model reduction; however, the way in which the reduced system
is rearranged to achieve certain required properties differs from one method
to the other.

2.2. Davison's technique


The key simplifying assumption consists of neglecting entirely the contri-
bution of the fast modes, i.e. their contribution to the model response is taken
to be zero (zz = 0) as shown in Fig. 2. (A tilde, C), is used to denote an
approximation to the original model.) As a result, Xl and Xz are approxi-
mated by Xl and xz ; the contribution of Zz to Xl and Xz is missing which
Downloaded by [RMIT University Library] at 22:49 19 June 2016

results, among other things, in steady-state offset. By inspection of Fig. 2,


the following relationships can be written immediately

(6)

(7)
and
(8)

The following reduced model is then obtained

~l = HllA l Hn-l Xl + HllGlu (9)


with
(10)
and the initial values
(II )
and
(12)

Equation (10) provides an algebraic relationship between the neglected and


retained state variables in the reduced model.

~1 (0) =Q
~

Figure 2. Davison's original method.


8:34 D. Bonvinand D. A. Mellichancp

2.3. Davison's first-modified and Possard'e techniques


Both of these methods possess the same underlying assumptions, the
difference resulting solely from the choice of the initial condition treatment.
Z2 is approximated by an algebraic equation so as to obtain steady-state
agreement between the large-scale and reduced models
Z2= -A 2- 1 G 2U=G 2U (13)

~1 (0)

+ ~1
+
Downloaded by [RMIT University Library] at 22:49 19 June 2016

~2 + ~2
~2 1------....::..., +

Figure 3. Davison's first-modified and Fossa-d's methods.

Figure a indicates the signal flow representation from which the following
relations can be written immediately
i, =A,z, + G,u (14)

Xl = Hnz 1 + H'2G2U ( 15)


and
X2 = H 21Z, + H 22G2U ( 16)

Equations (14)-( 16) combine to yield the reduced model


i 1 = H"A, H,,-l s, + [HnG, - HnA, H ll - ' H'2G2]U + H 12G 2U (17)
and
X2 = H 2, Hn-l x, + [H 22 - H 21.H,,-1 H 12]G 2U (18)
The first-modified approach of Davison assumes that
z,(O)=O (19)
and, therefore, from (15) and (18)
X,(O) = H I 2 G2U(0) (20)

x2(0 ) = H 22G2U(0) (21 )


1£ u(O) # 0, this leads to a jump in the system response at t = 0, which is not
realistic. Fossard (1970) corrected this ambiguous situation by modifying
the initial condition (19) to
(22)
which yields
(23)
Modal approaches to model reduction 835

However, the initial value of x2 is no longer zero


x 2(0) = [H 22- H 2l Hll-l H 12]G 2u(0) (24)

2.4. Marshall's technique


Marshall's technique was developed independently of Davison's original
method. However, it is instructive to note that it can also be derived in an
alternative way, i.e. Davison's original method can be modified in two steps
in order to achieve steady-state agreement between the original and reduced
models.

Step 1
Neglect completely the contribution of Z2 to the reduced model and then
modify the mode excitation G l to Gl so as to achieve steady-state agreement
Downloaded by [RMIT University Library] at 22:49 19 June 2016

for Xl' It follows then from Fig. 4 that


(25)
and
(26)

~1 (0) =Q

+ ~1

Figure 4. Marshall's method.

As a result, however
(27)
exhibits a steady-state offset.

Step 2
Consequently, G2 is determined so as to obtain zero steady-state offset
between X 2 and x2
G2=[H 22-l H 21Hll-l H 12-I]A2-l G 2 (28)
From Fig. 4, it can be seen that
il=Alzl+G1U (29)
and
(30)
83(i D. Bonvin and D. A. 111ellichamp

which, combined with (25), give the reduced model


.i 1 = H uA 1 H u - 1 Xl + H UG 1u (31 )

X2 = H 21H U - 1 Xl + H 22G2U (32)


with
(33)
and
(34)

At this point, it is worthwhile to notice that, for the case of constant


inputs, i.e. Ii = 0, Fossard's and Marshall's methods are equivalent in the
sense that the reduced models are identical. This feature was first pointed
out by Wilson (1974).
Downloaded by [RMIT University Library] at 22:49 19 June 2016

2.5. Davison's second-modified technique


In this approach compensation is introduced to eliminate steady-state
errors by considering a single input at a time. As in Davison's original
method, Z2 is neglected in evaluating Xl' For the ith input, the steady-state
errors of i 1, i and i 2 , i (where the brevet is used to denote the intermediate
x1,iand X 2,i components before gain compensation) are compensated by gain
elements represented, in Fig. 5, by the diagonal matrices 0l,i and 02,i'
respectively.

~1, i (0) =Q
~1

Figure 5. Davison's second-modified method.

Under steady-state conditions and considering only the effect of the ith
input, agreement between the original and reduced-order models is obtained
when the jth elements in the main diagonals of 0l,i and 02,i are taken to be

(d .) = ( H uA1-1 gl, i)j + ( H 12A2-1 g2,i)j (35)


1" i (H 11 A 1-1 g I, t.).1
and

(36)

respectively. It follows from Fig. 5 that

(37)
.l'! odal approaches to model reduction 837

and
(38)
which gives
(39)
For the ith input, the dynamic model for X, is similar to that of Davison's
original method
(40)

On consideration of all the r inputs, the approximations to X, and X2 become


r
x1 = L D1.iXi. i (41)
i=l
and
Downloaded by [RMIT University Library] at 22:49 19 June 2016

x2 = L 02,iH21HIl-1 x,,; (42)


i= 1

The initial conditions are matched for both x, and X 2, i.e.


x" ;(0) = 0 ,,; H Il %" ;(0) = 0 (43)

X2,i(0 ) = 0 2, ; H 21 H n - 1 01,;-1 X i(O) = O (44)


"
Figure 5 also shows that the eigenvector elements transforming %"i into x,,;
differ from the original ones, i.e. those for the transformation %" i into x,,;,

2.6. Litz' technique


The contribution of the fast modes to the reduced model is approximated
by an optimal linear combination of the slow modes
(45)
The approach is depicted in Fig. 6, from which can be obtained the expressions
(46)
and
(47)
which yields
(48)

h1 (0) =Q
~ -
h1 !:!11
[s! - ~r1 1;!21 I--- + 111
- +
-'-- -
<;2 1;!12 + 112
§, r---::-
!:!22 I - - - +
-

Figure 6. Litz' method'


838 D. Bonvin and D. A. Mellichamp

E is determined optimally through the use of Lagrange multipliers so as to


minimize a weighted integral of the square of the errors between 2: 2 and Z2'
Litz (1980) showed that the same conditions also minimize the integral of the
square of the errors between Xl and Xl' Despite these stringent requirements,
a closed-form expression for the reduced model is still obtained, in -oontrast
to Chidambara's methods (1969).
The matrix E is given by Litz (1979 a, 1980) as
E=A 2-I[S+(G 2-SU-'G I)(G IT U-IGI)-IGITjU-IA I (49)
with
I _
(5);;= - A",+;+ X; (G 2QG IT)i; (i= I, ... , n-m), (i = 1, ... , m) (50)

1 - T
Downloaded by [RMIT University Library] at 22:49 19 June 2016

(U);;=-~(GIQGI)ij (i,j=I, ... ,m) (51)


I\i + 1\;
Q is a matrix of weighting factors associated with the inputs in the optimiza-
tion procedure, and X; and G, are the complex conjugates of A; and G I ,
respectively. Again, from Fig. 6
zl=Alzl+Glu (52)

which can be combined with (46) to give the reduced model


jl = MAl M-IX I + MGlu (53)
The matrix E was derived in such a way as to ensure steady-state agreement
between the original and reduced models, i.e. between Xl and Xl as well as
between X 2 and x2 • Since ZI(O)=O, it follows from (46) and (47) that
(54)

Finally, the reduced model possesses different eigenvector elements than the
original model; the new eigenvectors span the reduced space in an ' optimal'
way.

3. Dependence or independence of the various reduced models on the choice of


retained state variahles
In an attempt to derive reduced-order models with increased accuracy,
much attention has been devoted, with good reason, to the best choice of
retained (dominant) modes. However, very little has been said concerning
the choice of retained state variables, Xl' Is it then implied that, with every
modal technique, such choice does not affect the deviation between the
original and reduced models, i.e. between Xl and Xl' and X2 and x2 ? Trilling
and Klein (1979) discussed the application of both Davison's and Marshall's
techniques to an industrial distillation column system. They found the
choice of retained states highly important, but ·their arguments were of a
qualitative nature and no conclusive results were provided.
The analysis here of this state dependence/independence property uses the
signal flow representations for all six techniques (Figs. 2-6). Retaining a
different set of state variables basically amounts to exchanging certain rows
in H (and H ll , H 12 , H 21 , H 22 ) and in B, and corresponding columns in H-I.
Modal approaches to model reduction 839

The operation, however, does not alter A 1 and A 2 • The mode excitation
term G = H-1B is not modified by exchanging rows in B if columns in H-1
are exchanged in a similar fashion; as a consequence, G 1 and G 2 are invariant
under this transformation.
Figure 2 shows that retaining a different set of state variables with Davison's
technique corresponds to an exchange of rows in H u and H21> but the reduced
model, given by Xl and X2 , remains unchanged. (Unchanged here does not
mean that the All and Bll matrices remain unaffected but rather that the
static and dynamic responses of the reduced-order model do not depend on
the choice of Xl and x 2 . )
With both Davison's first-modified and Fossard's techniques, G2 =
-A 2- 1 G 2 is independent of the choice of retained states and, consequently,
the same is true for the reduced model (Xl' x2 ) if the initial condition is
independent of that choice (see Fig. 3). This certainly is the case with
Downloaded by [RMIT University Library] at 22:49 19 June 2016

Davison's first-modified technique (Zl = 0). Fossard's method, however,


possesses the following canonical initial condition

(55)

Since H u- 1 has an entirely different column structure for two different sets
of retained states and H 12 would involve mainly an exchange of rows, Fossard's
method is dependent on the choice of retained state variables.
Ma.rshall's technique was shown earlier to be identical to Fossard's method
fur the case of constant inputs. Consequently, we conclude that Marshall's
red uced models also depend on the choice of retained states.
Davison's second-modified technique is depicted in Fig. 5; Xl,; and x2, i
(for the same reasons as with Davison's original technique) and 0l,i and
° 2, i (for reasons which are apparent from an inspection of (35) and (36)) are
independent of the choice of retained states. Consequently, the entire reduced-
order model is similarly independent of this choice.
Finally, Litz' technique, as shown in Fig. 6, will also be independent of the
choice of retained states if the matrix E is independent of that choice. This
is indeed the case since E is basically a function of G lO G 2 , Al and A 2 , all of
which are independent of the choice of retained state variables.
To summarize then, among the six techniques investigated in this study,
Fossarrl's and Marshall's methods produce reduced models which depend on
the choice of retained state variables while with the other four methods that
choice is basically irrelevant. In the light of those findings, one source of
confusion mentioned in the introduction can be eliminated:
(i) Models obtained with Davison's original technique are independent of
the choice of retained states, contrary to the early statements made by
Davison (1966) and Trilling and Klein (1979).
(ii) Numerical tests confirmed that the accuracy of Marshall's technique
is highly dependent on the choice of retained states. This property has been
almost totally ignored in the literature, the most recent occurrence, to the
authors' knowledge, being the study of Iwai and Kubo (1979) which dealt
with the determination of the optimal eigenvalues to be retained in a reduced-
order model. Their suggested approach, which used Marshall's method and
840 D. Bonoin. and D. A ..111ell£champ

an ISE criterion to expre~~ the degree of optimality of any choice of eigen-


values for the reduced model, failed to include the effect of choosing alternative
sets of state variables.
Davison (1966) and Trilling and Klein (1979) proposed combinatorial
approaches, based on the system eigenvectors, to determine the best choice
of retained states. The generalized structural dominance method recently
developed by Bonvin and Mellioharnp (1982) is, however, much more suited
to that task since it directly indicates the most sensitive' state' variables
associated with each of the modes. Furthermore, the need to investigate
many 01' all of the combinations of retained states among the original state
variables is obviated through use of this procedure.
Downloaded by [RMIT University Library] at 22:49 19 June 2016

4. Dependence or independence of the various reduced models on the choice of


retained inputs
One characteristic of the first five techniques (excepting Litz' method)
is that the choices of how many and which inputs to include in the model do
not affect A IV the reduced A matrix. Similarly, for these five methods, the
columns of Bit may be added or removed as the corresponding individual
inputs are added to or removed from the reduced model. Litz' method is .
completely different since the matrix M, which determines both A R and BR ,
is functionally dependent on which inputs are included in the reduced model
and how these are weighted. Litz' method is a global one in the sense that it
attempts to optimize information flow from the inputs to the states through
orientation of the reduced system eigenvectors. Hence, including unimportant
inputs (or weighting them excessively) is likely to reduce the model's accuracy
with respect to more important inputs. The computational implications
of this property are clear: with the five earlier techniques the designer does
not need to make an early decision concerning the inputs. to be retained in the
reduced-order model. With Litz' method. only those inputs which are
absolutely required should be included;. consequently, it is important that
some method, such as that proposed by Bonvin and Mellichamp ('1982), be
used to select the most effective system inputs of those available, in advance.
Only these should then be included in the reduced model.

S. Comparison of the techniques


The Table contains a summary of the details of implementation and key
properties which can be identified immediately from an analysis of the'
reduced-order models. All of the reduced models, except for Davison's
second modification. can be written in the form
i 1 = A lt x1 + BRu + Bsu (56)

with an algebraic relationship available to reconstruct x2 from Xl and u


x = PX
2 1 + Qu (57)

Also included in the Table are:


(i) the values of x1 (0) and x2 (0), in order to determine if a technique
conserves the zero initial conditions.of the original model,
Downloaded by [RMIT University Library] at 22:49 19 June 2016

First-modified Second-modified
Davison Davison Fossard Marshall Davison Litz

AH H"A,H n
-, H"A.H,,-' H"A,H"
-, H"A,H,,-' H"A,H n
-, MA,M-'
Bn HnG, Hng, H"G, H"G, H"g,,; MG,
+A RH'2T +ARH'2 T +AnH'2T
B~ 0 -H l2T - H'2T 0 0 0 '-'
0:::
c
-L NM-' R..
P H2[H" H 2, H,,-' H 2, H l l -, H 2, H,,-' x,= ~ O"i X,,; ~
i= 1
>@
~
Q 0 RT RT RT X2= ~ O 2 i H 21 0 -e
C
i-I
-:- X
'H 11-, X1,i
;:,..
'o"
x,(O) 0 - H'2 T u(O) 0 0 0 0 '"ce
x2(O) 0 - H 22Tu(O) RTu(O) RTu(O) 0 0
c
Steady-state No Yes
g
Yes Yes Yes Yes R..
agreement ~
...
I ndependence of Yes Yes No No Yes Yes '"
"'-
choice of x, ~
~.
'c"
Independence of A H Yes Yes Yes Yes Yes No
on choice of u '"
New eigenvector No No No No Yes YC"
clements

R=H 2,H,,-1 H'2-H22' T=A 2-'G2, O.,i=diag{(d'.;)j}, 02,;=diag{(d 2,i)j}' M=H n+H l2 E, N=H 2,+H 22E, (d',;)j' (d2,i)j and E
given hy (35), (3(\), (49)-(51).
Summary of t.hc six modal techniques. 00
.;:.
.....
842 D. Bonvin and D. A. 211 ellichamp

(ii) a statement concerning steady-state agreement between the original


and reduced models,
(iii) an indication of the independence (versus dependence) of the reduced
model on the choice of retained states and on the choice of retained
inputs, and
(iv) an indication of the eigenvector orientation relative to the eigen-
vector elements in the original model.

It is seen that Davison's original and first-modified techniques suffer from


potential steady-state-offset and initial-value problems, respectively. Fossard's
and Marshall's approaches also exhibit undesirable non-zero initial values,
which, however, are restricted to ;(2. Whether this latter defect is serious or
Downloaded by [RMIT University Library] at 22:49 19 June 2016

not is dependent on the particular application.


As to Davison's second-modified and Litz' methods, both appear to possess
important desirable features, i.e. they both retain the original initial conditions,
they maintain steady-state agreement, and they are independent of the choice
of state variables retained in the reduced-order model. Both methods also
result in a new orientation of the eigenvectors.
At this point it would be difficult for the designer to choose a particular
technique from among the six investigated in this work. One technique
might be quite appropriate for a particular problem but fail completely with
another problem. This peculiar feature caused much of the early confusion
and argument as researchers were able to formulate dynamic systems for
which their favourite technique happened to work efficiently while other
techniques failed. In order not to fall into the same trap, this study purposely
does not use any , constructed' example, but instead investigates the model
reduction of three systems which are thought to be representative of realistic
large-scale systems that have become well-known through previously published
studies.

(a) A homogeneous fixed-bed autothermal reactor model; the lumped


system was obtained by discretizing the PDE model using both the methods
of orthogonal collocation and finite differences; the system is unstable and
the eigenvalues are widely spread in the complex plane. The results for this
system are reported by Bonvin et al, (1982).

(b) A heterogencous catalytic fixed-bed reactor model; the lumped system


which was obtained by discretizing the PDE model using orthogonal colloca-
tion exhibits narrowly-spread, mostly complex eigenvalues.

(e) A double-effect evaporator model; an originally lumped system with


only real eigenvalues and two integrating modes.

Furthermore, both a qualitative comparison of the techniques and an


assessment of their strengths and weaknesses are possible since the six methods
have been derived from a unified viewpoint. Finally, a numerical cross-
comparison can establish some measure. of their accuracy in representing the
original systems and their relative power, i.e. how large a reduction ratio
can be used without significant losses in accuracy.
Modal approaches to model reduction 843

5.1. Catalytic fixed-bed reactor


This reaction system, which involves two catalytic fixed-beds in series,
was studied extensively by Foss and co-workers (Wallman 1977, Silva 1978).
They derived a 14th-order linear approximation to the reactor, with seven
catalyst particle and seven reactor wall temperatures forming the state vector.
The system model was reduced to seventh order for purposes of control system
implementation by the original investigators. As part of the present work,
the model reduction step was extended to include all of the modal techniques
discussed above.
Figure 7 shows the exit temperature response to a 4·5°C step increase in
the inlet temperature of the feed to the second bed. The curve 0 indicates
the temperature response of the original 14th-order linear model, while all the
other curves correspond to seventh-order approximations which involve only
Downloaded by [RMIT University Library] at 22:49 19 June 2016

the feed temperature as single input. The seven slowest modes and the seven
particle temperatures are retained in the reduced models, as suggested by the
original investigators (the best choice of retained modes and states is discussed
by Bonvin and Mellichamp 1982). It is seen that Davison's original and his
first-modified techniques give poor approximations in terms of steady-state
and initial-value matching, respectively. Davison's second-modified technique
failed completely when applied to this dynamic system. Marshall's and
Fossard's techniques are identical since a constant input (step change) is
applied; they provide a relatively good approximation to the 14th-order
system. However, the model obtained using Litz' approach is significantly
more accurate and constitutes an excellent approximation to the original,
linearized reactor model.

z
o
ti
s
w
o
w
0::
::::>
ti
0::
w
Q.
2
~ -16

-24';:-_ _-!;-_ _--:-_ _-:-_ _-!:-_ _~


o

Figure 7. Catalytic fixed-bed reactor. Exit temperature response to a 4·5°C


step change in temperature of the feed to the second bed. Model reduction
ratio = 14 : 7. 0: original 14th-order model; D, MD1, M, F, L: 7th-order
reduced models according to Davison, Davison first modification, Marshall,
Fossard and Litz, respectively.
844 J). Bonvin and u. A .31' ellichamp
Figurc 8 shows the exit temperature response to .a . 4'5°C, one-minute
rectangular pulse perturbation (i.e. a non-constant input) of the inlet feed
temperature to the second bed. As above, curve 0 indicates the temperature
response of the original 14th-order system. Curve 1\'[7 represents the tempera-
ture response of the seventh-order redueed model obtained using Marshall's
technique. Again, the seven slowest modes and the seven particle tempera-
turcs were retained, thus represcnting exactly the reduced model derived and
used by the original investigators for controller design and implementation.
Curve L7 identifies a seventh-order model obtained using Litz' technique and
retaining the same seven slowest modes. Litz' technique had not been
reported at the time of the original work, and it is interesting to note that this
reduced model represents a better approximation to the large-scale system
than could be obtained earlier. Even more important,. the third-order
Downloaded by [RMIT University Library] at 22:49 19 June 2016

approximation, L3, obtained using Litz' technique and retaining the three
slowest eigenvalues, still represents the original 14th-order system fairly
accurately. The significance of Fig. 8 is obvious: as more powerful and
more accurate reduction techniques are made available, models of significantly
lower order can be derived. Such advances facilitate the design of multi-
variable control systems and can eliminate the need for state reconstruction.

z
o
~
s
w
o
w
a:
::J
!;,i
a:
~
::;;;
W
f-

Figure 8. Catalytic fixed-bed reactor. Exit temperature. response to a 4'5°C,


one-minute pulse perturbation of the 'feed to the second bed. 0: original
14th·order model; M7, L7; 7th-order reduced models according to Marshall
and Litz ; L3; 3rd-order reduced model according to Litz.

5.2. Double-effect evaporator


A very thorough model reduction study starting with a l Oth-order model
of a double-effect evaporator has been presented by Wilson (1974). This
dynamic system has been discussed in the literature extensively; the reader
1),[odal approaches to model reduction 845

is referred to the case study of Fisher and Seborg (1976) for a physical
description of the evaporator.
A single model reduction example is discussed below: the lOth-order
evaporator model developed by the original investigators was reduced to
fourth-order using Marshall's (or Fossards) techniques. Figure 9 compares
the transient response of the original and reduced models for the case of a
step perturbation in feed flowrate. Wilson et al. (1974) included the following
four state variables in their reduced models: the hold-ups in the two effects,
11'1 and 11'2' the enthalpy in the first effect, H" and the concentration in the
second effect, 02' They concluded that the fifth-slowest mode had to be
retained instead of the fourth slowest one in order to guarantee good dynamic
response; this feature is easily seen in Fig. 9 by comparing curves 3 and 2,
which represent results obtained in the original work.
Downloaded by [RMIT University Library] at 22:49 19 June 2016

N
U
Z
o
i= -0.06
<[
0:
f--
Z
w
~ -0.12
o
u
(fl
(fl
w
...J -0.18
z
o
(fl
z
w
::;; -0.24
is

-0.30~0-----d:,--------;:';;---~----'!;;;------,~

Figure 9. Double-effect evaporator model. Reduction ratio = 10 : 4. 0: original


LOth-order model; I, 2, 3: all 4th-order reduced models according to
Marshall: 1=(AI-A4 ; WI' 11'2' C\, C2)· 2=(A1 - A4 : 1I'1,1I'2,H\,C2). 3=
(A1 - A3 , A.: WI' 11'2' H" C2 ) ·

However, the original investigators did not elaborate on the choice of


retained states nor on the importance of this choice. In fact, when the
generalized structural dominance analysis proposed by Bonvin and Mellicharnp
(1982) is used to determine the most sensitive state variables, W\, W2' 0\
(instead of H\) and O2 are found to be hest and, consequently, should be
retained in the reduced model for maximum accuracy if Marshall's and
Fossard's methods are to be used. Curve I shows the transient response of
the reduced model obtained when the four dominant modes and the best
set of state variables (11'1' 1V 2, 0\ and 02) are retained. This curve represents
an excellent approximation to the original lOth-order system and certainly a
846 D. Bonvin and D. A. Mellichamp

better one than either curve 2 or 3. Hence, with Marshall's and Fossard's
techniques, the choice of the state variables to be retained in the reduced
model is seen to be of the highest importance for deriving accurate approxi-
mations. If this choice is constrained, these two techniques should be used
with extreme caution; other techniques whose accuracy is independent of
the choice of retained state variables might be preferred. This fact has been
substantially overlooked in the literature.
No attempt will be made here to rank the six modal reduction techniques;
however, several important points can be made based on our results.
(i) Fossard's and Marshall's techniques usually work well when the retained
state variables are carefully chosen. Since the reduced relation for x2 exhibits
a non-zero initial condition, this feature may represent a disadvantage.
Downloaded by [RMIT University Library] at 22:49 19 June 2016

(ii) Davison's first-modified method exhibits non-zero initial conditions


for both Xl and x2 • Hence significant discrepancies between original and
reduced models will exist for short times.
(iii) Davison's second-modified method utilizes a set of gains chosen solely
to eliminate steady-state offset. The introduction of these gains changes
the original eigenvector orientation in the reduced model. Furthermore, the
new eigenvectors appear not to be oriented optimally, and our results have
shown extremely large errors in transient tests of certain systems.
The new technique of Litz was found to work extremely well with all of
the large-scale systems studied so far (in this study, and by Litz 1979, 1980).
The method permits larger "reductions in size for the same accuracy; further-
more, its implementation remains relatively easy and straightforward. The
power of Litz' technique appears to be related to the defect in Davison's
second-modified method, i.e. the orientation of the reduced-system eigen-
vectors in Litz' method is an optimal one, designed to minimize the differences
between original and reduced models. On the other hand, Litz' method
requires that the choice of inputs and their associated weightings be given
careful thought, since the final model will be a function of these characteristics.

6. Summary and conclusions


All the model reduction techniques have been developed here as if they
were extensions of Davison's original method, these modifications having
been made to provide, among other things, steady-state agreement between
the large-scale and reduced models.
Little attention was given to potential numerical difficulties which may
arise with some of the methods (e.g. zero eigenvalues or singularities of certain
matrices) or to special cases (e.g. an unstable system or repeated eigenvalues).
Such assessments are of relatively minor importance for the purposes of this
study which has been concerned mainly with putting the various approaches
on a common basis in order to compare them. Furthermore, if a specific
formulation for a reduced model fails, it is often the case that another, equi-
valent, expression can be used efficiently. Various formulations yielding
identical reduced-order models are presented in the Appendix for that purpose.
At odal approaches to model reduction 84i

This study points out the fact that the accuracy of two of the techniques
(Marshall's and Fossards) is critically dependent on the choice of the retained
state variables. Such choice is irrelevant to the accuracy of the other tech-
niques, unless numerical problems are encountered (e.g. near-singular matrices,
etc.).
Litz' technique is shown to be both highly accurate and powerful, thereby
permitting large reductions in model size. At this stage of evaluation, the
technique shows only a single potential weakness, viz. the dependence of the
reduced model on the choice of inputs. Litz' approach is something of a
hybrid method, linking modal to least-squares reduction techniques. It
possesses the advantages of both approaches, i.e. it gives an explicit formula-
tion for the reduced model and is optimal in the sense that it minimizes the
weighted integral of the squarc of the deviations between the original and
Downloaded by [RMIT University Library] at 22:49 19 June 2016

reduced models.

Appendix
Equivalent formulations of some reduced models
(a) For all but Litz techniques:

A R = HlIA I HlI-I = All + A l2 H 21 HlI-l = All - A l2V 22V-I 21


(b) For Davison's technique

BR= HIlG 1= HlImlH-IB]


where m[ . ] signifies the first 11/ rows of [ . ].
(c) For Davison's first-modified, Fossards and Marshall's techniques:

B R= HllG l+ HIlA1HlI-l H l2A2-l G 2= HIlGI+AnHI2A2-l G 2


=Au[HlIA1-l G 1 + H l2A2-1 G 2]
=AR[All-Al2A22-1 A21]-1[BI-A12A22-1 B2]
=A n m[A-IB]
= BI-[H12A2-ARHl2]A2-1 G 2
=Bl-A12V22-1 A 2-l G 2
= B1 - A 12[ H 22 - H 21 H Il - l H l2]A2-I G 2
p= H 2IH lI-l= -V 22-1 V 21
Q= [H 21 HlI-1 H 12- H 22 1A2- 1 G 2= - V 22-1 A 2-1 G 2
(d) For Davison's first-modified and Fossards techniques:
Bs= - H 12A2-1 G 2=A H - l BIt-mfA--'B]

REFERENm~s
HOKVIN, D., and :\'IELLICHAMF, D. A., 1982, lilt. J. Control, 35, 80i.
BONVIN, D., RINKER, R. G., and lVIELLICHAMF, D. A.. 1982, Ohern. Ellgng. Sci. (to
appear),
CHIDAMBARA, M. R .. and DAVISOK. K .J., 19ni a, I.E.E.E. Trans, autom. Control,
12, 119; 19ni b, Ibid., 12,213; 196i c, Ibid., 12, i99.
CHIDAMBARA, !If. R .. 19f19, Proc. Joint auiom . Control Con]: Denver.
848 .'ifodal approaches to model reduction

DAVISON, K J., 1966, I.E.E.E. Trans. autom. Control, 11,93; 1968, Ibid., 13, 214.
FISHER, D. G., and SEDORG, D. K, 1976, Multivariable Computer Control-A Case
Study (North Holland/American Elsevier).
FOSSARD, A., 1970, I.E.E.E. Trans. autom. Control, 15,261.
GENESIO, R, and MILANESE, M., 1976, I.E.E.E. Trans. autom.. Control, 21, 118.
IWAI, Z., and KUDO, Y., 1979, Int. J. Control, 30, 823.
LITz, L., 1979 a, Regelungstechnik, 27, 80; 1979 b, Ibid., 27, 273; 1980, 2nd IF AC
Symp. on Large-Scale System.'!, Toulouse.
MARSHALL, S. A., 1966, Control, 10, 642.
SILVA, J. M., 1978, Ph.D. Dissertation (Berkeley: University of California).
TRILLING, U., and KLEIN, H. J., 1979, Regelungstechnik, 27, 37.
WALLMAN, P. H., 1977, Ph.D. Dissertation (Berkeley : University of California).
WILSON, R G., 1974, Ph.D. Dissertation (Edmonton: University of Alberta).
WILSON, R. G., FISHER, D. G., and SEDORO, D. E., 1974, A.I.Ch.E. Jl, 20, 1131.
Downloaded by [RMIT University Library] at 22:49 19 June 2016

You might also like