Anda di halaman 1dari 3

Chapter 1.

Methods for Model Reduction

Quantitatively (and also roughly) with the Henkel singular values,

1= 2.584625 ; 2= 2.579625 ; 3= 1.748808 ; 4= 1.745308 ; 5=


1.248374 ; 6= 1.245874 ;

7= 0.1009331 ; 8= 0.098934

System with all real poles, following Liu and Anderson (1989)

Consider the following fourth order transfer function:

s+ 4
G ( s )=
( s+ 1 )( s+3 ) ( s +5 ) ( s+10) (1.10.4)

Balanced truncation gives better second order model truncation in this


example (see Figure 1.10.5). Combining H2 norm minimization with modal
truncation yields are similar reduced model to that obtained via balanced
realization truncation, see the gain plots of Figure 1.10.5.(In the figure G mt2
denotes the model obtained by truncation and H2-norm optimization.) The
calculation is quite easy( see (1.8.7) and (1.8.8)). The H 2 norm of the
additive error decreases from 1.834647 x 10 -2 to 1.639455 x 10-2 by use of
H2 norm minimization on Cr.

For this example, we can also study the use of Markov parameter or time
moment matching through equation error techniques. If we use the
controllable canonical form for the state space realization of the original
system, then the first k Makrov parameters of the kth orde model given by
(1.5.16) with Cr = C P LT (L P LT))-1 match those of the original system,
provided that the obtained model is controllable (Obinata, 1989).

Figure 1.10.6 Comparison of gain characteristic of addictive error for three


reduced order models (balanced truncation, Markov parameters and time
moment matching)

Set k=3; calculation on (1.10.4) yields Markov parameters of the reduced


model h0 = CrBr = 0; h1 = CrArBr = 0; h2 = CrArBr = 1 which match those of
(1.10.4). in the case of k=2, the resulting reduce order model is
uncontrollable. If we use the special canonical form suggested by
Chidambara (1971), than the first k moments of the kth order model
which is calculated by (1.5.16) with Cr = C P LT (L P LT)-1 match those of the
original system. We can confirm the matching on (1.10.4) with the values;
m0 = -CrAr-1Br = 2.666667 x 10-2, m1=CrAr-2Br = 3.6668 x 10-2. The
matching of Markov parameters in general gives rise to a smaller error in
the high frequency region and the matching of the time moments of
smaller error in the low frequency region.

Figure 1.10.3 comparison of the gain characteristic of the original system


G and four different reduced order system (modal truncation, balanced
truncation, Henkel norm reduction, singular perturbation of balanced
realization)

Figure 1.10.4 Additive error characteristic for four reduced order model
together with error bounds.

Chapter 2

Multiplicative Approximation

2.1 The Multiplicative Approximation Problem

In this section, we will describe what the multiplicative


approximation is, and motivate it. In later section we shall indicate how
the problem solved. We shall find that three is a remarkable
correspondence with earlier results on additive error approximation. There
are actually two related ways of measuring error in the multiplicative
approximation problems, known as relative error and multiplicative error
and we shall consider both possibilities; the reduce model we derive turn
out to be the same.

We begin with a problem statement for the scalar transfer function


case

Approximation Problems using Relative Error and Multiplicative Error

Let G(s) be a transfer function of degree n. A relative error approximation


of G is a transfer function such that.

(s) = G(s) (1- ((s)) (2.1.1)

Where (s) is rational and stable, is minimized.

There is a very closely related problem, which is turn out can be solved in
essentially the same way as the above relative error approximation
problem; one requires

G (s) = (s) (1- ((s)) (2.1.2)

And otherwise the problem statement is the same. This is termed a


multiplicative error approximation problem.
Figure 1.10.2 Additive error characteristic for four reduced order model
together with error bounds.

Dissimilar reduced order models for modal truncation and


balanced realization truncation.

In the above example, it is clear that the approximate models with order
at least four will match well the original system in the low frequency
region, since neglecting the high frequency models does not have
significant effect of the approximation because of the lower peak gain of
the neglected models. Another example is now given, obtained by
modifying the contribution of each mode in the former example. The
sample values are used for (i, i), i=1.4. Set the different values for
ki as k1= 0.05; k2= 0.005; k3= 0.007; k4= 0.004. the four fourth order
models (Gbt : balance truncation; Gmt : model truncation; Goh : optimal
Henkel norm approximation; Gbs : singular perturbation of balanced
realization) are calculated and compared in the gain characteristic with
the original system Figure 1.10.3. Modal truncation tries to retain the two
slower models, while the order three methods seek to copy the first and
the third modes which correspond to the larger Henkel singular values.
The comparison of the additive error is shown in the Figure 1.10.4. Though
optimal Henkel norm approximation enjoys and attractive the norm
bounds for the addictive error, the actual error achieved is not superior to
that for the order three methods in this example. Two methods of
truncation give larger error in the low frequency region; on the other hand
singular perturbation and the optimal Henkel norm approximation give
larger error in the high frequency region because feed through term
appear in reduced models. We have to expect larger error for system in
which each mode has a similar contribution to whole frequency response.
We can see the similarity of the contributions models.

Anda mungkin juga menyukai