You are on page 1of 25

university-logo

FEL3100 Multivariable Feedback Control


Lecture 8: LMIs, Model Reduction and Summary [Ch.
11-12]
Elling W. Jacobsen, Automatic Control Lab, KTH
Lecture 8: LMIs, Model Reduction and Summary FEL3100 MIMO Control
university-logo
Todays program

A very brief introduction to LMI problems

A brief introduction to model reduction

A very brief summary of the course

Information about exam.


Lecture 8: LMIs, Model Reduction and Summary FEL3100 MIMO Control
university-logo
Linear Matrix Inequalities
Linear Matrix Inequality (LMI)
F
0
+F
1
x
1
+F
2
x
2
+. . . F
m
x
m
> 0
where
x = [x
1
. . . x
m
] is a real vector
F
i
, i = 0, m are symmetric real matrices
An LMI imposes a convex constraint on x

feasibility problem: nd some x that satises LMI

optimization problem: minimize c


T
x subject to LMI
Old problem, efcient solvers now available
Lecture 8: LMIs, Model Reduction and Summary FEL3100 MIMO Control
university-logo
Example 1: Linear stability problem
LTI system

x = Ax(t )

Lyapunov function V(x) = x


T
Px > 0 with

V(x) < 0 iff
P = P
T
> 0, A
T
P +PA < 0

Corresponds to an LMI feasibility problem (just stack the


two inequalities into one big matrix)
Lecture 8: LMIs, Model Reduction and Summary FEL3100 MIMO Control
university-logo
Example 2: Linear robust stability problem
Polytopic LTV system

x = A(t )x(t ) , A(t ) Co{A


1
, . . . , A
L
}

Lyapunov function exist iff


P = P
T
> 0 , A
T
i
P +PA
i
< 0, i = 1, . . . L

LMI feasibility problem


Lecture 8: LMIs, Model Reduction and Summary FEL3100 MIMO Control
university-logo
Optimization problems: H

-norm
Consider LTI system

x = Ax(t ) +Bw(t )
z(t ) = Cx(t ) +Dw(t )
The H

norm of G
yw
is equivalent to solving
min s.t .
_
_
A
T
P +PA PB C
T
B
T
P I D
T
C D I
_
_
< 0, P > 0
i.e., minimization subject to LMI.

Computing upper bound on structured singular value via


min
D
(D
1
ND) is another example that can be cast as an
optimization problem with LMI constraint.

Solutions to Riccati equations, e.g., in H

-optimal control,
can be obtained via LMI feasibility problem.
Lecture 8: LMIs, Model Reduction and Summary FEL3100 MIMO Control
university-logo
LMIs in contol - the Essence

Many problems in optimal and robust control can be cast


as LMI problems convex optimization problems for
which efcient algorithms exist (e.g., interior point
methods)

Matlab: LMI toolbox and LMI feasibility and optimization


solvers in Robust Control toolbox.
Lecture 8: LMIs, Model Reduction and Summary FEL3100 MIMO Control
university-logo
The LTI Model Reduction Problem
Given minimal state-space model (A, B, C, D)

x = Ax(t ) +Bu(t ) x R
n
,u R
m
y(t ) = Cx(t ) +Du(t ) y R
l
or, as input-output model
Y(s) =
_
C(sI A)
1
B +D

. .
G(s)
U(s)

model reduction: we seek a model

x
r
= A
r

x(t ) +B
r
u(t ) x
r
R
k
,u R
m
y(t ) = C
r
x
r
(t ) +D
r
u(t ) y R
l
G
a
(s) = C
r
(sI A
r
)
1
B
r
+D
r
with k < n, such that the predicted input-output behavior is close
in some sense, e.g., G G
a

is small
Lecture 8: LMIs, Model Reduction and Summary FEL3100 MIMO Control
university-logo
Why Model Reduction?

Reduced computational complexity


time for dynamic simulation is approximately proportional to
n
3
(if A dense)
in particular, important for real time applications, e.g.,
controllers

Controller synthesis methods typically yield controllers that have


order at least equal to model order, usually signicantly higher.
Thus, to obtain low order controller
reduce model order prior to control design, or
reduce controller order after design
A multitude of model reduction methods. Here we will consider those
most commonly employed in linear control theory.
Lecture 8: LMIs, Model Reduction and Summary FEL3100 MIMO Control
university-logo
Truncation and Residualization
Divide state vector x into two vectors x
1
and x
2
of dimension k and
n k, respectively

x
1
= A
11
x
1
(t ) +A
12
x
2
(t ) +B
1
u(t )

x
2
= A
21
x
1
(t ) +A
22
x
2
(t ) +B
2
u(t )
y(t ) = C
1
x
1
(t ) +C
2
x
2
(t ) +Du(t )
We aim at removing the state vector x
2
, i.e., obtain a kth order model
from an nth order model

Truncation: let x
2
= 0, i.e., remove x
2
from state-space model

Residualization: let

x
2
= 0, i.e., x
2
becomes an algebraic
variable which depends on x
1
and u
Lecture 8: LMIs, Model Reduction and Summary FEL3100 MIMO Control
university-logo
Truncation
With x
2
= 0 we get
(A
r
, B
r
, C
r
, D
r
) = (A
11
, B
1
, C
1
, D)

Simply removing a number of states makes little sense in the


general case

Consider rst transforming (A, B, C, D) into Jordan form and


arrange the states so that x
2
correspond to the fastest modes,
i.e., largest magnitude eigenvalues

If the Jordan form is diagonal (distinct eigenvalues


i
) then
G(s) =
n

i =1
c
i
b
T
i
s
i

Removing the n k fastest modes then yields the model error


G(s) G
a
(s) =
n

i =k+1
c
i
b
T
i
s
i
G G
a

i =k+1
(c
i
b
T
i
)
|Re(
i
)|
note: must assume stable G(s)
Lecture 8: LMIs, Model Reduction and Summary FEL3100 MIMO Control
university-logo
Truncation contd

The H

error bound
n

i =k+1
(c
i
b
T
i
)
|Re(
i
)|
depends not only on eigenvalues of fast modes, but also on the
terms (residues) c
i
b
T
i
, i.e., the effect of inputs u on x
2
and effect
of x
2
on outputs y

At =
G
a
(i ) = G(i ) = D
Thus, no error at innite frequency
Lecture 8: LMIs, Model Reduction and Summary FEL3100 MIMO Control
university-logo
Residualization
With

x
2
= 0 we get (assume A
22
invertible)
x
2
(t ) = A
1
22
A
21
x
1
(t ) A
1
22
B
2
u(t )
and elimination of x
2
from partitioned model then yields

x
1
(t ) = (A
11
A
12
A
1
22
A
21
)x
1
(t ) + (B
1
A
12
A
1
22
B
2
)u(t )
y(t ) = (C
1
C
2
A
1
22
A
21
)x
1
(t ) + (D C
2
A
1
22
B
2
)u(t )

Thus, the reduced model (A


r
, B
r
, C
r
, D
r
) =
(A
11
A
12
A
1
22
A
21
, B
1
A
12
A
1
22
B
2
, C
1
C
2
A
1
22
A
21
, DC
2
A
1
22
B
2
)

Corresponds to a singular perturbation method if A transformed


to Jordan form rst

At zero frequency
G
a
(0) = G(0)
follows from the fact that

x
2
0 at steady-state
Lecture 8: LMIs, Model Reduction and Summary FEL3100 MIMO Control
university-logo
Comments on Truncation and Residualization

Truncation gives best approximation at high frequencies

Residualization gives best approximation at low frequencies

The two methods are related through the bilinear transformation


s
1
s

Both methods can in principle give rise to arbitrarily large model


reduction errors since total effect of states on input-output
behavior not necessarily related to the speed of response

Should be combined with some method that ensures relatively


small overall effect of removed states on input-output behavior
balancing
Lecture 8: LMIs, Model Reduction and Summary FEL3100 MIMO Control
university-logo
The Controllability Gramian

The state space model (A, B, C, D) has an impulse response


from u(t ) to x(t ) given by
X(t ) = e
At
B

A quantication of the size of the impulse response is


P(t ) =
_
t
0
X()X
T
()d =
_
t
0
e
A
BBe
A
T

Dene the Controllability Gramian P as


P = lim
t
P(t )

The controllability gramian can be computed from the Lyapunov


equation
AP +PA
T
+BB
T
= 0

P is a quantitative measure for controllability of the different


states. Essentially, measures the effect of the inputs on the
different states
Lecture 8: LMIs, Model Reduction and Summary FEL3100 MIMO Control
university-logo
The Observability Gramian

The state space model (A, B, C, D) with input u(t ) = 0 and initial
state x(0) = x

has the output


y(t ) = Ce
At
x

The energy of the output


_
t
0
y
T
()y()d = x
T
_
t
0
e
A
T

C
T
Ce
A
d
. .
Q(t )
x

Dene the Observability Gramian Q as


Q = lim
t
_
t
0
e
A
T

C
T
Ce
A
d

The observability gramian can be computed from the Lyapunov


equation
A
T
Q +QA +C
T
C = 0

Q is a quantitative measure for observability of the different


states. Essentially, measures the effect of states on outputs
Lecture 8: LMIs, Model Reduction and Summary FEL3100 MIMO Control
university-logo
Balanced Realizations

Consider the nth order stable state space model (A, B, C, D) with
controllability and observability gramians P and Q, respectively

We seek a similarity transformation of the states x


b
(t ) = Tx(t )
so that the transformed state space model

x
b
= TAT
1
x
b
(t ) +TBu(t )
y(t ) = CT
1
x
b
(t ) +Du(t )
has controllability and observability gramians
P = Q = diag(
1
, . . . ,
n
);
i
=
_

i
(PQ)
where the Hankel singular values
1
>
2
> . . . >
n

Each state x
bi
in the balanced realization is as observable as it is
controllable, and
i
is a measure of how controllable/observable
it is

A state with a relatively small


i
has a relatively small effect on
the input-output behavior and can hence be removed without
signicantly affecting the predicted input-output behavior
Lecture 8: LMIs, Model Reduction and Summary FEL3100 MIMO Control
university-logo
Balanced Truncation and Residualization

Consider the balanced realization (A, B, C, D) of G(s) with


partitioning
A =
_
A
11
A
12
A
21
A
22
_
, B =
_
B
1
B
2
_
, C =
_
C
1
C
2
_
P = Q =
_

1
0
0
2
_
where
1
= diag(
1
, . . . ,
k
) and
2
= diag(
k+1
, . . . ,
n
)

A balanced truncation or residualization retaining the k states


corresponding to
1
will both have model reduction error
G G
k
a

2
n

i =k+1

i
twice the sum of the tail

May in principle include frequency dependent weighting to


emphasize certain frequency ranges, However, this introduces
extra states and it is furthermore usually non-trivial to choose
weigths that give the desired result
Lecture 8: LMIs, Model Reduction and Summary FEL3100 MIMO Control
university-logo
Optimal Hankel Norm Approximation

The Hankel norm of a transfer-matrix E(s)


E(s)
H
=
1
=
_
(PQ)
i.e., equals the maximum Hankel singular value of E(s)

Optimal Hankel norm approximations seeks to minimize


G G
k
a

H
for a given order k of the reduced order model

For stable square G(s) the optimal Hankel norm kth order
approximation can be directly computed and has Hankel norm
error
G G
k
a

H
=
k+1

The optimal Hankel norm is independent of the D-matrix of G


k
a
The minimum -norm of the error is
min
D
G G
k
a

i =k+1

i
i.e., sum of the tails only
Lecture 8: LMIs, Model Reduction and Summary FEL3100 MIMO Control
university-logo
Unstable models
Balanced truncation and residualization and optimal Hankel norm
approximations applies to stable G(s) only. Two tricks to deal with
unstable models
1. Separate out unstable part before performing model reduction of
stable part
G(s) = G
u
(s) +G
s
(s) G
a
(s) = G
u
(s) +G
sa
(s)
2. Consider coprime factorization of G(s)
G(s) = M
1
(s)N(s)
with M(s) and N(s) stable. Apply model reduction to [M(s) N(s)]
and use
G
a
(s) = M
1
a
(s)N
a
(s)
Lecture 8: LMIs, Model Reduction and Summary FEL3100 MIMO Control
university-logo
Model Reduction in Matlab

modreal: truncation or residualization

slowfast: slow/fast mode decomposition

balreal: balanced realization

hankelmr: optimal Hankel norm approximation

stabproj: decompose into stable and antistable parts

ncfmr: balanced model truncation for normalized coprime


factors
Homework 8: test it out yourself! (no hand in).
use e.g., rss(n) to generate a random stable state-space
model with n states.
compare step and frequency responses of different
reduced models
Lecture 8: LMIs, Model Reduction and Summary FEL3100 MIMO Control
university-logo
Very Brief Course Summary

Lecture 1:

introduction: never forget that (i) feedback control has


hard limitations, imposed by the system itself and (ii)
models are always uncertain

classic SISO feedback control: shaping the open loop

Lecture 2: Limitations in SISO systems

conicts and trade-offs: S +T = 1, Bode sensitivity integral

fundamental limitations; RHP zeros and poles, time delays

scaling dependent limitations

Lecture 3: Introduction to MIMO systems

poles and zeros

generalized Nyquist Theorem

directionality; SVD, singular values for gain

the weighted H

-norm

ill-conditioning and input uncertainty


Lecture 8: LMIs, Model Reduction and Summary FEL3100 MIMO Control
university-logo

Lecture 4: Limitations in MIMO feedback

extension of SISO results

zero and pole directions

Bode integral: trade-off involves frequencies as well as


directions

limitations from uncertainty

Lecture 5: uncertainty and robust stability

uncertainty set: discs on Nyquist plot

RS from small gain theorem: the M- -loop

SISO: RS + NP RP

Lecture 6: RS and RP in MIMO systems

MIMO: perturbation typically a structured matrix: need


for analysis of RS

RP can be cast as ctitious RS problem; analyze using

MIMO: RS + NP = RP
Lecture 8: LMIs, Model Reduction and Summary FEL3100 MIMO Control
university-logo

Lecture 7: controller design and synthesis

synthesis: min
K
P
m
, m = 2,

H
2
- generalized LQG

- worst case approach

controllers from solving Riccati equations: observer + state


feedback

-synthesis for RP; DK-iterations

Glover-McFarlane: classic loopshaping for performance,


followed by robust stabilization
Lecture 8: LMIs, Model Reduction and Summary FEL3100 MIMO Control
university-logo
The Final Moment
Exam:

1-day take home exam, open book.

Allowed aids: course book(s), lecture slides, matlab,


calculator

Not allowed: old exams, exercises, solutions

available between April 4-April 30

send an email to jacobsen@ee.kth.se with the date at


which you want to take it (give at least 3 days notice)

All exercises must be approved prior to exam.


Lecture 8: LMIs, Model Reduction and Summary FEL3100 MIMO Control