Lecture: Model reduction

Automatic Control 2
Model reduction

Prof. Alberto Bemporad

University of Trento

Academic year 2010-2011

Prof. Alberto Bemporad (University of Trento) Automatic Control 2 Academic year 2010-2011 1 / 17
 Lecture: Model reduction Model reduction

Systems reduction

The complexity of the control law often depends on the order of the system
(for example in state-space methods like dynamic compensation)
For control design purposes, can we approximate the model with another
model of reduced order that preserves the original transfer function as much
as possible ?
We already know that uncontrollable and unobservable modes do not affect
the transfer function. They can be eliminated by operating a canonical
decomposition
Can we try to eliminate other modes that are weakly uncontrollable and/or
weakly observable, and get a numerically well-conditioned lower-order
state-space realization ?

Model reduction and balanced transformations answer the above questions

Prof. Alberto Bemporad (University of Trento) Automatic Control 2 Academic year 2010-2011 2 / 17
 Lecture: Model reduction Model reduction

Unbalanced realizations and scaling

Consider the linear system
 
− 12
– ™
x1 (k + 1) x1 (k)
   −6 
0 10
 x (k + 1) = + u(k)
− 32 x2 (k) 106

2 0
— x (k)
 
”
1
10−6

 y(k) = 106
x2 (k)
The state x1 is “weakly” reachable, but “very”observable
The state x2 is “very” reachable, but “weakly” observable
Let’s rescale the system by operating the change of coordinates
 6 
10 0
z= x
0 10−6
The system expressed in new coordinates is numerically balanced
  – 1 ™
z1 (k + 1) z1 (k)

−2
  
0 1
 z (k + 1) = + u(k)
2
(k)

2 0 − z 2 1
3
— z (k)

”
1

 y(k) = 1 1
z2 (k)
Prof. Alberto Bemporad (University of Trento) Automatic Control 2 Academic year 2010-2011 3 / 17
 Lecture: Model reduction Model reduction

Grammians

Consider the discrete-time linear system

x(k + 1) = Ax(k) + Bu(k)



y(k) = Cx(k) + Du(k)

Since now on we assume that matrix A is asymptotically stable

The controllability Grammian for discrete-time systems is the matrix
X∞
MATLAB
Wc ¬ Aj BB0 (A0 )j
Wc = gram(sys,’c’)
j=0

The observability Grammian for discrete-time systems is the matrix

∞
X
MATLAB
Wo ¬ (A0 )j C0 CAj
Wc = gram(sys,’o’)
j=0

Similar definitions exist for continuous-time systems

For P
discrete-time systems the controllability Grammian is related to the cost of minimum energy control:
∞
min j=0 u2 (j) = x(0)0 Wc−1 x(0). The observability Grammian to the output energy of the free response:
P∞ 2
j=0 y (j) = x(0)0 Wo x(0). The Grammians solve the Lyapunov equations Wc = AWc A0 + BB0 and
Wo = A0 Wo A + C0 C, respectively
Prof. Alberto Bemporad (University of Trento) Automatic Control 2 Academic year 2010-2011 4 / 17
 Lecture: Model reduction Model reduction

Balanced state-space realizations

Definition
A state-space realization is called balanced if the Grammians Wc and Wo
are equal and diagonal
 σ 0 ... 0 
1

 0 σ2 ... 0 
Wc = Wo = Σ, Σ =  .. .. . . .. 
. . . .
0 ... 0 σn

A procedure to derive the transformation matrix T such that the equivalent

state-space form à = T −1 AT, B̃ = T −1 B, C̃ = CT, D̃ = D is balanced is
described in [1]
the procedure is implemented in the MATLAB function balreal
MATLAB
where σ = [σ12 . . . σn2 ]0
[sysb,σ,T −1 ,T ] = balreal(sys)
[1] A.J. Laub, M.T. Heath, C.C. Paige, R.C. Ward, “Computation of System Balancing Transformations and
Other Applications of Simultaneous Diagonalization Algorithms,” IEEE Trans. Automatic Control, vol. 32,
pp. 115-122, 1987
Prof. Alberto Bemporad (University of Trento) Automatic Control 2 Academic year 2010-2011 5 / 17
 Lecture: Model reduction Model reduction

Model reduction
Once the system is in balanced form we can easily reduce the order of the
model by eliminating the states associated with small σi ’s

Σ1 0|
 
W̃c = W̃o = Σ2  Σ1
−0 Σ| 2

   
Ã11 Ã| 12 B̃1
z(k + 1) = z(k) + u(k)
−
Ã21 Ã| 22 −
B̃2
” —
y(k) = C̃1 −
C̃2 z(k)

MATLAB
rsys = modred(sys,elim,’del’)

z1 (k + 1) = Ã11 z1 (k) + B̃1 u(k) elim = indices of the states

y(k) = C̃1 z1 (k) to be eliminated
A similar idea applies to continuous-time systems
Wednesday, May 26, 2010

Prof. Alberto Bemporad (University of Trento) Automatic Control 2 Academic year 2010-2011 6 / 17
 Lecture: Model reduction Model reduction

Example

s3 +11s2 +36s+26
Transfer function: G(s) = s4 +14.6s3 +74.96s2 +153.7s+99.65
State-space realization in canonical reachability form:
 0 1 0 0
 0
A= 0
0
0
0
1
0
0
1 , B = 0 , C = [ 26 36 11 1 ] , D = 0
0
−99.65 −153.7 −74.96 −14.6 1

Controllability and observability Grammians:

0.0000 0.0000 −0.0001 0.0000 84.6757 122.1890 37.4774 3.3919
   
Wc = 0.0000
−0.0001
0.0001
−0.0000
−0.0000
0.0006
−0.0006
0.0000 Wo = 122.1890
37.4774
179.5744
55.2945
55.2945
17.0406
5.0110
1.5448
0.0000 −0.0006 0.0000 0.0408 3.3919 5.0110 1.5448 0.1401

After balancing, we get

MATLAB
 0.1394 » sys=ss(A,B,C,D);
0 0 0

» [sysb,sigma,Tinv,T] = balreal(sys);
W̃c = W̃o = 0
0
0.0095
0
0
0.0006
0
0 » Wc=diag(sigma);
0 0 0 0.0000 » Wo=Wc

Prof. Alberto Bemporad (University of Trento) Automatic Control 2 Academic year 2010-2011 7 / 17
 Lecture: Model reduction Model reduction

Example

After applying the transformation matrix T we get

 −3.601 0.8212 −0.6163 −0.05831   −1.002 
à = −0.6163 −1.027 −5.914 −1.127 , B̃ = −0.08612
−0.8212 −0.593 1.027 0.09033 −0.1064
0.05831 0.09033 1.127 −4.492 0.008112
C̃ = [ −1.002 0.1064 −0.08612 −0.008112 ] , D̃ = 0

Let’s eliminate states z3 , z4 and get a model of reduced order 2

The transfer function of the reduced-order model is
0.9926s + 0.7297
Gb (s) =
s2 + 4.194s + 2.81

Prof. Alberto Bemporad (University of Trento) Automatic Control 2 Academic year 2010-2011 8 / 17
 Lecture: Model reduction Model reduction

Example (cont’d)
Let’s compare the frequency responses of G and Gb : they are almost
indistinguishable !
Bode diagram of G and G
b

−10

−15

−20
In general, the smaller are the
Magnitude (dB)

−25

−30
removed singular values σi with
−35

−40 respected to the ones we keep,

−45
0
G
Gb
the more similar are the
responses of the original and
Phase (deg)

−45

reduced-order models
−90
−1 0 1 2
10 10 10 10
Frequency (rad/sec)

Note that after balancing the states completely loose their physical meaning
The original state x can be recovered (approximately)
”T T — from the reduced state
z1 using the transformation matrix T = T11 12
21 T22
”T —
x = T21 z1 ← x is treated here as an output of the reduced-order model
11

The reduced state z1 can be estimated by a state observer (the pair Ã1 , C̃1 is
observable by construction)
Prof. Alberto Bemporad (University of Trento) Automatic Control 2 Academic year 2010-2011 9 / 17
 Lecture: Model reduction Model reduction

Matched DC gain method for model reduction

Consider again the complete model

x1 (k + 1) = A11 x1 (k) + A12 x2 (k) + B1 u(k)

x2 (k + 1) = A21 x1 (k) + A22 x2 (k) + B2 u(k)
y(k) = C1 x1 (k) + C2 x2 (k)

Assume the dynamics of x2 are “infinitely fast”: x2 (k + 1) ≈ x2 (k)

We can eliminate the states contained in x2 by substituting

x2 (k) = (I − A22 )−1 (A21 x1 (k) + B2 u(k))

therefore obtaining

x1 (k + 1) = (A11 + A12 (I − A22 )−1 A21 )x1 (k) + (B1 + A12 (I − A22 )−1 B2 )u(k)
y(k) = (C1 + C2 (I − A22 )−1 A21 )x1 (k) + C2 (I − A22 )−1 B2 u(k)

A similar idea can be applied in continuous-time, by setting ẋ2 (t) ≈ 0

x2 (t) = A−1
22 (A21 x1 (t) + B2 u(t))

Prof. Alberto Bemporad (University of Trento) Automatic Control 2 Academic year 2010-2011 10 / 17
 Lecture: Model reduction Model reduction

Matched DC gain method for model reduction

Property
The matched DC-gain method preserves the DC gain of the
original full-order model

Proof:
Simply observe that for both the original and the reduced-order model in
steady-state x1 , x2 depend on u in the same way ƒ

MATLAB elim is the vector of state indices to

rsys=modred(sys,elim,’mdc’) eliminate

Prof. Alberto Bemporad (University of Trento) Automatic Control 2 Academic year 2010-2011 11 / 17
 Lecture: Model reduction Model reduction

Example (cont’d)
Consider again the state-space realization
 0 1 0 0
 0
A= 0
0
0
0
1
0
0
1 , B = 0 , C = [ 26 36 11 1 ] , D = 0
0
−99.65 −153.7 −74.96 −14.6 1

Let’s eliminate the states x3 and x4 using the matched DC-gain method
We get a 2nd order model whose transfer function is
0.1467s2 + 0.4803s + 0.3469
Gm (s) =
s2 + 2.05s + 1.329

Bode diagram of G and Gr

−10

−15
Magnitude (dB)

−20

−25

−30
Emphasis here is on matching at low
−35

−40
0
frequencies (DC gain in particular!)
G

DC gains: G(0) = Gm (0) = 0.2609

Gr
Phase (deg)

−45

−90
−1 0 1 2
10 10 10 10
Frequency (rad/sec)

Prof. Alberto Bemporad (University of Trento) Automatic Control 2 Academic year 2010-2011 12 / 17
 Lecture: Model reduction Model reduction

Example (cont’d)

Step responses of G, Gr, Gm

0.35
G
Gr
Gm
0.3

0.25

Original DC gain: G(0) = 0.2609

Amplitude

0.2

0.15
DC gain of Gm (s): Gm (0) = 0.2609
0.1
DC gain of Gb (s): Gb (0) = 0.2597
0.05

0
0 1 2 3 4 5 6
Time (sec)

The matched DC-gain method is only good to capture the DC gain exactly
Model reduction via balanced transformation provides best match

Prof. Alberto Bemporad (University of Trento) Automatic Control 2 Academic year 2010-2011 13 / 17
 Lecture: Model reduction Model reduction

Comments on model reduction

How good is the reduced-order model should be judged on the performance

of the original system in closed-loop with a controller based on the reduced
model

A good reduced-order model provides very good closed-loop performance

and a low-order dynamic control law at the same time

Let’s see an example ...

Prof. Alberto Bemporad (University of Trento) Automatic Control 2 Academic year 2010-2011 14 / 17
 Lecture: Model reduction Model reduction

Example (cont’d)
Consider the LQR performance index
∞
X
min y2 (k) + ρu2 (k)
k=0

LQR controller based on complete model (A, B, C, D):

1
u(t) = Kx(t) + Hr(t), H=
C(−A − BK)−1 B + D

LQR controller based on reduced model (A1 , B1 , C1 , D1 ):

 z1 
z2 = T −1 x
u(t) = Kb z1 (t) + Hb r(t), Hb = C(−A−BK [I 10]T −1 )−1 B+D
b

LQR controller based on the model reduced by the matched DC gain method:
x1 =
1 0 0 0
0100 x
1
u(t) = Km x1 (t) + Hm r(t), Hm = C(−A−BK [I 0])−1 B+D m

Prof. Alberto Bemporad (University of Trento) Automatic Control 2 Academic year 2010-2011 15 / 17
 Lecture: Model reduction Model reduction

Example (cont’d)

Closed−loop step response of G with different LQR controllers

=
1.4
K
K
u Kx
b

= [ −29.5403 −43.0578 −13.2412 −1.2001 ] x

Km
1.2

1 ” —
u = Kb 0 T −1 x
Amplitude

0.8

= [ −29.5145 −43.0825 −13.2468 −1.1999 ] x

0.6
” —
0.4
u = Km I 0 x
= [ −17.5373 −21.4324 0 0 ] x
0.2

0
0 0.5 1 1.5 2 2.5 3 3.5
Time (sec)

” —
Note the similarity between controllers K and Kb 0 T −1

Prof. Alberto Bemporad (University of Trento) Automatic Control 2 Academic year 2010-2011 16 / 17
 Lecture: Model reduction Model reduction

English-Italian Vocabulary

model reduction riduzione dell’ordine del modello

Grammian gramiano

Translation is obvious otherwise.

Prof. Alberto Bemporad (University of Trento) Automatic Control 2 Academic year 2010-2011 17 / 17