Structure preserving reduced order models based on ... · expansions, empirical balancing (e.g.,...

Preview:

Citation preview

1/53

University of Groningen

Structure preserving reduced ordermodels based on balancing fornonlinear systems

Jacquelien M.A. ScherpenUniversity of Groningen

Autumn school Terschelling, 24 September 2009

A-P

DF

Merger D

EM

O : P

urchase from w

ww

.A-P

DF

.com to rem

ove the waterm

ark

Autumn school Terschelling, 24 September 2009 2/53

University of Groningen

Purpose of this presentation

To give a overview of our contributions to positive real andbounded real balancing for nonlinear systems. It can beincluded in the previous nonlinear balancing framework,and structure preservation is obtained for model reduction.Finally, reduction to minimal port-Hamiltonian systems isdiscussed.

Positive and bounded real results are joint work with TudorIonescu, and Kenji Fujimoto. Port-Hamiltonian results arejoint work with Arjan van der Schaft.

Autumn school Terschelling, 24 September 2009 3/53

University of Groningen

Problem setting

Consider an input-output system Σ:u −→ Σ −→ y

and a time-invariant state space realization

x = f(x) + g(x)u

y = h(x)(∗)

Assume that (∗) is a valid state space realization of Σ

about x0.

Autumn school Terschelling, 24 September 2009 4/53

University of Groningen

Problem setting (continued)

Questions treated here:

• If (∗) is minimal and dissipative, can we reformulatethe problem in standard balancing form withobservability and controllability?

• Does model reduction based on the above preservethe dissipativity structure?

• If (∗) is non-minimal and port-Hamiltonian, how canwe obtain a minimal port-Hamiltonian realization?

• Do we need duality between controllability andobservability notions?

Autumn school Terschelling, 24 September 2009 5/53

University of Groningen

Problem setting (continued)

• For linear control systems, rather complete picture,though structure preserving methods still developed.

• Complexity increase: need for nonlinear tools.

Autumn school Terschelling, 24 September 2009 5/53

University of Groningen

Problem setting (continued)

• For linear control systems, rather complete picture,though structure preserving methods still developed.

• Complexity increase: need for nonlinear tools.

Some examples:

• Nonlinear circuits, power systems, fluid systems,MEMS, NEMS .....

• Inflatable space structures

Autumn school Terschelling, 24 September 2009 6/53

University of Groningen

Background

Linear realization theory more or less complete, since 70’s.

• Balanced realizations and Gramians (Moore 1981, . . .).• Balance between past minimal control energy and

generated future output energy .• Hankel operator, Hankel singular values, similarity

invariants, input-output view, tool for model reduction.

Autumn school Terschelling, 24 September 2009 6/53

University of Groningen

Background

Linear realization theory more or less complete, since 70’s.

• Balanced realizations and Gramians (Moore 1981, . . .).• Balance between past minimal control energy and

generated future output energy .• Hankel operator, Hankel singular values, similarity

invariants, input-output view, tool for model reduction.

• Transparent structure, Hankel singular values aremeasure for controllability and observability of state⇒ i.e., in balanced form: if xi is badly controllable and

badly observable then xi is almost “non-minimal”.

Autumn school Terschelling, 24 September 2009 7/53

University of Groningen

Background (continued)

• Unstable systems: normalized right and left coprimefactorizations, (e.g., Meijer, 1990, Ober and McFarlane, 1989)

Autumn school Terschelling, 24 September 2009 7/53

University of Groningen

Background (continued)

• Unstable systems: normalized right and left coprimefactorizations, (e.g., Meijer, 1990, Ober and McFarlane, 1989)

• For nonlinear control systems model reduction oftendone on “ad hoc” basis, i.e., dependent on applicationor by singular perturbation techniques.

Autumn school Terschelling, 24 September 2009 7/53

University of Groningen

Background (continued)

• Unstable systems: normalized right and left coprimefactorizations, (e.g., Meijer, 1990, Ober and McFarlane, 1989)

• For nonlinear control systems model reduction oftendone on “ad hoc” basis, i.e., dependent on applicationor by singular perturbation techniques.

• Proper Orthogonal Decomposition, Karhoenen-Loèveexpansions, empirical balancing (e.g., Lall et. al., 2002)

⇒ data-based linear projection methods.

• “Analytical” methods based on nonlinear extension ofbalancing theory: • Verriest and Gray, 2000

• Scherpen, Fujimoto, Gray, since 1993

Autumn school Terschelling, 24 September 2009 8/53

University of Groningen

Structure preservation

• Both Krylov and balancing method recent interest in(dissi)passivity structure preservation: bounded realand positive real structure, e.g., Antoulas, Sorensen,Brenner, Ha Bin Minh/Trentelman, Meyer, etc. .

• Also interest in physical interpretation, i.e.,port-Hamiltonian structure preservation, RLC structureinterpretation, e.g., Oelof/van der Schaft, Meyer,Polyuga/van der Schaft, etc.

• Control structure preservation from optimizationproblem, e.g., Weiland, ...

Autumn school Terschelling, 24 September 2009 9/53

University of Groningen

Outline

• Review stable linear systems balancing

• Review stable nonlinear systems balancing

• Storage functions in terms of observability andcontrollability functions

• Storage functions and coprime factorizations

• Dissipativity preserving model reduction

• Simple example

• Port-Hamiltonian structure preserving reduction to aminimal system

Autumn school Terschelling, 24 September 2009 10/53

University of Groningen

Stable linear systems

Continuous-time, causal linear input-output systemS : u → y with impulse response H(t).If S is also BIBO stable then the system Hankel operator :

H : Lm2 [0, +∞) → L

p2[0, +∞)

: u → y(t) =

∫ ∞

0

H(t + τ)u(τ) dτ.

Autumn school Terschelling, 24 September 2009 10/53

University of Groningen

Stable linear systems

Continuous-time, causal linear input-output systemS : u → y with impulse response H(t).If S is also BIBO stable then the system Hankel operator :

H : Lm2 [0, +∞) → L

p2[0, +∞)

: u → y(t) =

∫ ∞

0

H(t + τ)u(τ) dτ.

Time flipping operator F : Lm2 [0, +∞) → Lm

2 (−∞, 0]u

t t

u

Autumn school Terschelling, 24 September 2009 10/53

University of Groningen

Stable linear systems

Continuous-time, causal linear input-output systemS : u → y with impulse response H(t).If S is also BIBO stable then the system Hankel operator :

H : Lm2 [0, +∞) → L

p2[0, +∞)

: u → y(t) =

∫ ∞

0

H(t + τ)u(τ) dτ.

Time flipping operator F : Lm2 [0, +∞) → Lm

2 (−∞, 0]u

t t

u

⇒H(u) = S ◦ F(u)

Autumn school Terschelling, 24 September 2009 11/53

University of Groningen

Stable linear systems (continued)

H = OC, with the controllability and observabilityoperators C and O.

H∗H is a self-adjoint compact operator with σi are Hankelsingular values , i.e., σ2

i are eigenvalues of H∗H.

Autumn school Terschelling, 24 September 2009 11/53

University of Groningen

Stable linear systems (continued)

H = OC, with the controllability and observabilityoperators C and O.

H∗H is a self-adjoint compact operator with σi are Hankelsingular values , i.e., σ2

i are eigenvalues of H∗H.

(A,B,C) as. stable state space realization of S of order n.

• σ2i are eigenvalues of MW , where W ≥ 0 and M ≥ 0

are the usual controllability and observabilityGramians fulfilling

AW + WAT = −BBT

AT M + MA = −CT C

Autumn school Terschelling, 24 September 2009 12/53

University of Groningen

Stable linear systems (continued)

(A,B,C) is minimal ⇔ M > 0 and W > 0.

If (A,B,C) is minimal and as. stable, then there exists astate space representation where

Σ := M = W =

σ1 0. . .

0 σn

σ1 ≥ σ2 ≥ ... ≥ σn > 0 Hankel singular values. Thensystem is in balanced form .

Autumn school Terschelling, 24 September 2009 13/53

University of Groningen

Outline

• Review stable linear systems balancing

• Review stable nonlinear systems balancing

• Storage functions in terms of observability andcontrollability functions

• Storage functions and coprime factorizations

• Dissipativity preserving model reduction

• Simple example

• Port-Hamiltonian structure preserving reduction to aminimal system

Autumn school Terschelling, 24 September 2009 14/53

University of Groningen

Review stable nonlinear systems balancing

Smooth system

x = f(x) + g(x)u

y = h(x)

where u ∈ Rm, y ∈ R

p, and x ∈ M (manifold of dim n).

Assumptions:

• f(0) = 0, 0 as. stable eq. point for u = 0, x ∈ X.

• h(0) = 0.

• Controllability function Lc and observability function Lo

smooth and exist.

Autumn school Terschelling, 24 September 2009 15/53

University of Groningen

Energy functions: Gramian extensions

Lc(x0) = min

u ∈ L2(−∞, 0)

x(−∞) = 0, x(0) = x0

1

2

∫ 0

−∞

‖ u(t) ‖2 dt

Minimum amount of control energy necessary to reachstate x0. Lc is the so-called controllability function .

Autumn school Terschelling, 24 September 2009 15/53

University of Groningen

Energy functions: Gramian extensions

Lc(x0) = min

u ∈ L2(−∞, 0)

x(−∞) = 0, x(0) = x0

1

2

∫ 0

−∞

‖ u(t) ‖2 dt

Minimum amount of control energy necessary to reachstate x0. Lc is the so-called controllability function .

Lo(x0) =1

2

∫ ∞

0

‖ y(t) ‖2 dt,x(0)= x0

u(τ)=0, 0≤τ<∞

Output energy generated by state x0.Lo is the so-called observability function.

Autumn school Terschelling, 24 September 2009 16/53

University of Groningen

The controllability and observability function

• In linear case: Lo(x) = 12xT Mx and Lc(x) = 1

2xT W−1x.

• Lyapunov and Hamilton-Jacobi-Bellmann equationscharacterize Lo and Lc.

• Role of observability and controllability for linearsystems is replaced by zero-state observabilityand asymptotic reachability (or anti-stabilizability).

Autumn school Terschelling, 24 September 2009 16/53

University of Groningen

The controllability and observability function

• In linear case: Lo(x) = 12xT Mx and Lc(x) = 1

2xT W−1x.

• Lyapunov and Hamilton-Jacobi-Bellmann equationscharacterize Lo and Lc.

• Role of observability and controllability for linearsystems is replaced by zero-state observabilityand asymptotic reachability (or anti-stabilizability).

• Minimality can be expressed in terms of Fliessexpansions and rank conditions, e.g., Isidori 1995.

• If 0 < Lc(x) < ∞ and 0 < Lo(x) < ∞ for x ∈ X, x 6= 0,then, under appropriate additional assumptions, thesystem is minimal , Scherpen, Gray 2000.

Autumn school Terschelling, 24 September 2009 17/53

University of Groningen

Example

l1

l2m1

m2

θ1

θ2

Frictionless two-link robotarm with actuator (torque)at first link.

System not as. stable. However, associated gradientsystem is as. stable!

Therefore, study of gradient system.Other advantage: reduction based on gradient system isstructure preserving when translated to original system.

Autumn school Terschelling, 24 September 2009 18/53

University of Groningen

Example

Approximately (series approximation, method of Lukes)solving eqs for Lo and Lc (m1 = m2 = l1 = l2 = 1) yields

-0.2-0.1

00.1

0.2-0.2

-0.1

0

0.1

0.2

0

0.001

0.002

0.003

-0.2-0.1

00.1

0.2

-0.2-0.1

0

0.1

0.2-0.2

-0.1

0

0.1

0.2

0

20

40

-0.2-0.1

0

0.1

0.2

Everywhere Lo > 0 and Lc > 0, thus minimal!

However, where Lo and Lc are both close to 0 correspondsto “weakly” obs./contr. subspaces ⇒ almost non-minimal .

Autumn school Terschelling, 24 September 2009 19/53

University of Groningen

Hankel norm

• Hankel norm for linear systems

‖Σ‖2H = max

u∈L2+

‖H(u)‖2

‖u‖2

=

• Hankel norm for nonlinear systems

‖Σ‖2H = max

u∈L2+

‖H(u)‖2

‖u‖2

=

Autumn school Terschelling, 24 September 2009 19/53

University of Groningen

Hankel norm

• Hankel norm for linear systems

‖Σ‖2H = max

u∈L2+

‖H(u)‖2

‖u‖2= max

u∈L2+

〈u,H∗H(u)〉

〈u, u〉

=

• Hankel norm for nonlinear systems

‖Σ‖2H = max

u∈L2+

‖H(u)‖2

‖u‖2= max

u∈L2+

〈u,H∗(H(u), u)〉

〈u, u〉

=

Autumn school Terschelling, 24 September 2009 19/53

University of Groningen

Hankel norm

• Hankel norm for linear systems

‖Σ‖2H = max

u∈L2+

‖H(u)‖2

‖u‖2= max

u∈L2+

〈u,H∗H(u)〉

〈u, u〉

= maxx

xT M x

xT W−1x

• Hankel norm for nonlinear systems

‖Σ‖2H = max

u∈L2+

‖H(u)‖2

‖u‖2= max

u∈L2+

〈u,H∗(H(u), u)〉

〈u, u〉

= maxx

Lo(x)

Lc(x)

Autumn school Terschelling, 24 September 2009 19/53

University of Groningen

Hankel norm

• Hankel norm for linear systems

‖Σ‖2H = max

u∈L2+

‖H(u)‖2

‖u‖2= max

u∈L2+

〈u,H∗H(u)〉

〈u, u〉

= maxx

xT M x

xT W−1x= λmax(H

∗H) = λmax(MW ) = σ21

• Hankel norm for nonlinear systems

‖Σ‖2H = max

u∈L2+

‖H(u)‖2

‖u‖2= max

u∈L2+

〈u,H∗(H(u), u)〉

〈u, u〉

= maxx

Lo(x)

Lc(x)= ???

Autumn school Terschelling, 24 September 2009 20/53

University of Groningen

Balanced realizations (continued)

How to determine ???• For relation with Hankel operator and Hankel norm,

state dependent balanced form does not suffice.

• By considering both eigenstructure of

⋆ differential adjoint (dH(·))∗ (H(·)) and

⋆ full nonlinear Hilbert adjoint H∗(H(u), u),

characterization based on sort of parametrization thatis related to the input value yields form that fill inthe ??? , i.e., give explicit expression for Hankel norm.

Autumn school Terschelling, 24 September 2009 21/53

University of Groningen

Balanced realizations (continued)

• Appropriate assumptions, then there existsneighborhood X of 0 and x = Φ(z) s.t.

Lc(Φ(z)) =1

2

n∑

i=1

z2i

σi(zi)Lo(Φ(z)) =

1

2

n∑

i=1

z2i σi(zi).

In particular, on X, ‖Σ‖H = supz1

Φ(z1,0,...,0)∈X

σ1(z1).

• Singular value functions unique at coordinate axes.

• Tool for balanced structure preserving model reduction.

Autumn school Terschelling, 24 September 2009 21/53

University of Groningen

Balanced realizations (continued)

• Appropriate assumptions, then there existsneighborhood X of 0 and x = Φ(z) s.t.

Lc(Φ(z)) =1

2

n∑

i=1

z2i

σi(zi)Lo(Φ(z)) =

1

2

n∑

i=1

z2i σi(zi).

In particular, on X, ‖Σ‖H = supz1

Φ(z1,0,...,0)∈X

σ1(z1).

• Singular value functions unique at coordinate axes.

• Tool for balanced structure preserving model reduction.

• Discrete time version similar! Fujimoto, Scherpen 2007.

Autumn school Terschelling, 24 September 2009 22/53

University of Groningen

Example (continued)

Gradient system of frictionless two-link robot arm, withm1 = 1, m2 = 10, l1 = 1, l2 = 10

ρ1(s) = 4.2543 × 10−4 + 3.7070 × 10−11s2 − 1.4619 × 10−17s4 + o(s4)

ρ2(s) = 3.7915 × 10−5 − 4.5718 × 10−10s2 − 5.4584 × 10−13s4 + o(s4).

Hankel norm in small neighborhood U of the origin, e.g,U = { z | ‖z‖ ≤ 10 }, (based on 4-th order Taylor seriesapproximation)

‖Σ‖H ≈ sups∈[−10,10]

ρ1(s) = 4.2543 × 10−4.

Autumn school Terschelling, 24 September 2009 23/53

University of Groningen

Outline

• Review stable linear systems balancing

• Review stable nonlinear systems balancing

• Storage functions in terms of observability andcontrollability functions

• Storage functions and coprime factorizations

• Dissipativity preserving model reduction

• Simple example

• Port-Hamiltonian structure preserving reduction to aminimal system

Autumn school Terschelling, 24 September 2009 24/53

University of Groningen

Dissipativity

Consider smooth nonlinear system Σ

x = f(x) + g(x)u

y = h(x) + d(x)u,

x ∈ Rn, u ∈ R

m, y ∈ Rp.

Assumptions:• asymptotically reachable from 0, zero-state detectable.• f(0) = 0, 0 as. stable eq. point for u = 0.• h(0) = 0.

Autumn school Terschelling, 24 September 2009 25/53

University of Groningen

Dissipativity

Definition: Σ is dissipative with respect to supply rates(u, y), if ∃ storage function S : R

n → R, S(x) ≥ 0, suchthat dissipation inequality holds:

S(x0) +

∫ t1

t0

s(u, y)dt ≥ S(x1),

for all x, u and t1 ≥ t0, with x0 = x(t0) and x1 = x(t1).

Differential version:∂S(x)

∂x(f(x) + g(x)u) ≤ s(u, y).

Assumption:There exists ϕ(·), such that s(ϕ(y), y) < 0, ϕ(0) = 0.

Autumn school Terschelling, 24 September 2009 26/53

University of Groningen

Dissipativity

Available storage function of Σ is:

Sa(x0) = sup

u ∈ L2(0,∞)

x(∞) = 0, x(0) = x0

∫ ∞

0

s(u(t), y(t)) dt.

Required supply function of Σ is

Sr(x0) = inf

u ∈ L2(−∞, 0)

x(−∞) = 0, x(0) = x0

∫ 0

−∞

s(u(t), y(t)) dt.

Autumn school Terschelling, 24 September 2009 26/53

University of Groningen

Dissipativity

Available storage function of Σ is:

Sa(x0) = sup

u ∈ L2(0,∞)

x(∞) = 0, x(0) = x0

∫ ∞

0

s(u(t), y(t)) dt.

Required supply function of Σ is

Sr(x0) = inf

u ∈ L2(−∞, 0)

x(−∞) = 0, x(0) = x0

∫ 0

−∞

s(u(t), y(t)) dt.

Lemma : If Σ dissipative w.r.t. s(u, y), then 0 ≤ Sa ≤ Sr

Autumn school Terschelling, 24 September 2009 27/53

University of Groningen

Dissipativity

• Dissipativity of Σ w.r.t. s(u, y) = 12[uT yT ]J

u

y

.

• Define r(x) = [I dT (x)]J

I

d(x)

.

Assumption: r(x) > 0.

Autumn school Terschelling, 24 September 2009 27/53

University of Groningen

Dissipativity

• Dissipativity of Σ w.r.t. s(u, y) = 12[uT yT ]J

u

y

.

• Define r(x) = [I dT (x)]J

I

d(x)

.

Assumption: r(x) > 0.

• Sa is the stabilizing (minimal and Sr is theanti-stabilizing (maximal) solution of aHamilton-Jacobi-Bellman equation.

Autumn school Terschelling, 24 September 2009 28/53

University of Groningen

Dissipativity

Minimal linear system (A,B,C,D). ThenSa(x) = 1

2xT Kminx and Sr = 1

2xT Kmaxx, with Kmin and Kmax

stabilizing and antistabilizing sol. of

AT K + KA +

KB − [0 CT ]J

I

D

[I DT ]J

I

D

−1

·

BT K − [I DT ]J

0

C

−[0 CT ]J

0

C

= 0.

Autumn school Terschelling, 24 September 2009 28/53

University of Groningen

Dissipativity

Minimal linear system (A,B,C,D). ThenSa(x) = 1

2xT Kminx and Sr = 1

2xT Kmaxx, with Kmin and Kmax

stabilizing and antistabilizing sol. of

AT K + KA +

KB − [0 CT ]J

I

D

[I DT ]J

I

D

−1

·

BT K − [I DT ]J

0

C

−[0 CT ]J

0

C

= 0.Specific J : positive real,

bounded real Riccati equa-

tion.

Autumn school Terschelling, 24 September 2009 29/53

University of Groningen

Passivity/positive real

Formulation of positive real balancing (i.e., s(u, y) = uT y) interms of observability and controllability functions:

p = m, J =

0 Im×m

Im×m 0

, r(x) = d(x) + dT (x)

Autumn school Terschelling, 24 September 2009 29/53

University of Groningen

Passivity/positive real

Formulation of positive real balancing (i.e., s(u, y) = uT y) interms of observability and controllability functions:

p = m, J =

0 Im×m

Im×m 0

, r(x) = d(x) + dT (x)

For linear system (A,B,C,D) Riccati equation becomes:

KA + AT K + (KB − CT )R−1(BT K − C) = 0.

Positive real balancing: transformation to equalize anddiagonalize Kmin and Kmax.

Autumn school Terschelling, 24 September 2009 30/53

University of Groningen

Controllability and observability formulation

Consider system Σextended

x = f(x) + g(x)r−1(x)h(x) − g(x)r−1/2(x)u1 + K(x)r−1/2(x)u2

y1 = −r−12 (x)gT (x)

∂T Sa(x)

∂x

y2 = r−12 (x)h(x)

.

Theorem: Under the assumptions mentioned, consider Sa

and Sr of strictly passive Σ = (f(x), g(x), h(x), d(x)), thenSa = Lo, Sr = Lc, Lo observability function and Lc iscontrollability function of Σextended.

Autumn school Terschelling, 24 September 2009 31/53

University of Groningen

Bounded real

Formulation of bounded real balancing (i.e.,s(u, y) = 1

2(||u||2 − ||y||2)) in terms of observability and

controllability functions:

J =

I 0

0 −I

, r(x) = I − dT (x)d(x), c(x) = −dT (x)h(x)

Autumn school Terschelling, 24 September 2009 31/53

University of Groningen

Bounded real

Formulation of bounded real balancing (i.e.,s(u, y) = 1

2(||u||2 − ||y||2)) in terms of observability and

controllability functions:

J =

I 0

0 −I

, r(x) = I − dT (x)d(x), c(x) = −dT (x)h(x)

For linear system (A,B,C,D) Riccati equation becomes:

AK+KA+(KB+CT D)(I−DT D)−1(BT K+DT C)+CT C = 0.

Bounded real balancing on Kmin and Kmax.

Autumn school Terschelling, 24 September 2009 32/53

University of Groningen

Controllability and observability formulation

Consider system Σboundedreal

x = f(x) + g(x)r−1(x)dT (x)h(x) − g(x)r−1/2(x)u1 + K(x)l−1/2(x)u2

y1 = −r−12 (x)gT (x)

∂Sa(x)

∂x

y2 = l−12 (x)h(x), with

∂Sr(x)

∂xK(x) = hT (x).

Theorem: Under the assumptions mentioned, consider Sa

and Sr of strictly bounded real Σ = (f(x), g(x), h(x), d(x)),then Sa = Lo, Sr = Lc, Lo observability function and Lc iscontrollability function of Σboundedreal.

Autumn school Terschelling, 24 September 2009 33/53

University of Groningen

Background

• Linear version of results found in Ober, Jonckheere,recently thesis by Ha Binh Min, Trentelman.

• Linear analysis further developed for model reductionwith error bounds.

• Normalized coprime factorizations for nonlinearsystems by Scherpen, van der Schaft, Paice, Ball

Autumn school Terschelling, 24 September 2009 34/53

University of Groningen

Outline

• Review stable linear systems balancing

• Review stable nonlinear systems balancing

• Storage functions in terms of observability andcontrollability functions

• Storage functions and coprime factorizations

• Dissipativity preserving model reduction

• Simple example

• Port-Hamiltonian structure preserving reduction to aminimal system

Autumn school Terschelling, 24 September 2009 35/53

University of Groningen

A factorization approach

Assumptions and dissipative system, supply rate s(u, y).Consider Σfact:

x = f(x) + g(x)r−1(x)

(gT (x)

∂T Sa(x)

∂x− c(x)

)+ g(x)r−1/2(x)v

y1 = r−1(x)

(gT (x)

∂T Sa(x)

∂x− c(x)

)

y2 = h(x) + d(x)r−1(x)

(gT (x)

∂T Sa(x)

∂x− c(x)

).

Generalization of normalized coprime factorization!

Autumn school Terschelling, 24 September 2009 36/53

University of Groningen

A factorization approach

Theorem:Σfact has controllability and modified observability functionsLc(x) = Sr(x) − Sa(x) > 0 and LJ

o (x) = Sa(x) > 0, with

LMo (x) =

∫ ∞

0

1

2yT My dt, x(0) = x, x(∞) = 0 for M = MT .

Note that minimal energy required to reach a state is givenby difference in required and available storage. Energyobserved at the output is the maximal storage available atthat state.

Autumn school Terschelling, 24 September 2009 37/53

University of Groningen

Outline

• Review stable linear systems balancing

• Review stable nonlinear systems balancing

• Storage functions in terms of observability andcontrollability functions

• Storage functions and coprime factorizations

• Dissipativity preserving model reduction

• Simple example

• Port-Hamiltonian structure preserving reduction to aminimal system

Autumn school Terschelling, 24 September 2009 38/53

University of Groningen

Model reduction using factorization approach

Assumptions:• 0 < Sa < Sr, exist, as. stab. requirement,• Hessians Lo and Lc positive definite.

For system Σ similar to Hankel approach

π2i (s) =

Sa(ξi(s))

Sr(ξi(s))

Define for Σfact:

ρ2i (s) =

LJo (ξi(s))

Lc(ξi(s))

parametrized in s.

Autumn school Terschelling, 24 September 2009 39/53

University of Groningen

Model reduction using factorization approach

Theorem:Assume appropriate assumptions are fulfilled, so that ρi(s)

exist. Then if πi(s) are the axis singular values frombalancing Sa and Sr, then:

πi(s) =ρi(s)√

1 + ρ2i (s)

.

Autumn school Terschelling, 24 September 2009 39/53

University of Groningen

Model reduction using factorization approach

Theorem:Assume appropriate assumptions are fulfilled, so that ρi(s)

exist. Then if πi(s) are the axis singular values frombalancing Sa and Sr, then:

πi(s) =ρi(s)√

1 + ρ2i (s)

.

Theorem:Appropriate assumptions, then there exists coordinatetransformation x = Φ(z) such that:

Sr(Φ(z)) =1

2

n∑

i=1

z2

i

πi(zi)and Sa(Φ(z)) =

1

2

i

z2

i πi(zi), πi(zi) = πi(Φi(z)).

Autumn school Terschelling, 24 September 2009 40/53

University of Groningen

Model reduction using factorization approach

• If πk > πk+1, then split accordingly in Σ1 and Σ2 fortruncation.

• Available storage and required supply preserved!

Autumn school Terschelling, 24 September 2009 40/53

University of Groningen

Model reduction using factorization approach

• If πk > πk+1, then split accordingly in Σ1 and Σ2 fortruncation.

• Available storage and required supply preserved!

Σ1 and Σ2 :

S1a(z

1) = Sa(z1, 0), S1

r (z1) = Sr(z

1, 0) and

S2a(z

2) = Sa(0, z2), S2

r (z2) = Sr(0, z

2).

Autumn school Terschelling, 24 September 2009 41/53

University of Groningen

Model reduction using factorization approach

• Singular value functions of subsystem Σ1 areπi(zi, 0), i = 1, . . . , k and the singular value functions ofsubsystem Σ2 are πj(0, zj), j = k + 1, . . . , n.

• Σ1,2 are dissipative with respect to the supply rates(u, y1,2).

• Similar result for non factorization case.

• Obstacle: s < 0. Not useful for e.g., LQG (HJB)balancing? Still dissipation preserved in "half line"dissipativeness manner, i.e., via one storage function,see Ha Binh Min/Trentelman.

Autumn school Terschelling, 24 September 2009 41/53

University of Groningen

Model reduction using factorization approach

• Singular value functions of subsystem Σ1 areπi(zi, 0), i = 1, . . . , k and the singular value functions ofsubsystem Σ2 are πj(0, zj), j = k + 1, . . . , n.

• Σ1,2 are dissipative with respect to the supply rates(u, y1,2).

• Similar result for non factorization case.

• Obstacle: s < 0. Not useful for e.g., LQG (HJB)balancing? Still dissipation preserved in "half line"dissipativeness manner, i.e., via one storage function,see Ha Binh Min/Trentelman.

Autumn school Terschelling, 24 September 2009 42/53

University of Groningen

Outline

• Review stable linear systems balancing

• Review stable nonlinear systems balancing

• Storage functions in terms of observability andcontrollability functions

• Storage functions and coprime factorizations

• Dissipativity preserving model reduction

• Simple example

• Port-Hamiltonian structure preserving reduction to aminimal system

Autumn school Terschelling, 24 September 2009 43/53

University of Groningen

Electrical circuit

L2L1 R1

V

R

x1 = −x1 + x2 + u

x2 = x1 − x2 − x32

y = −x1 + u = x1 − x2 + 2u.

xi

current in inductor i, i = 1, 2. System strictly positive real,strictly passive from input voltage to voltage over resistors.

Taylor approximations of Sa and Sr yield extended system.

Autumn school Terschelling, 24 September 2009 44/53

University of Groningen

Electrical circuit

Σextended:

x1 = −5

4x1 +

5

4x2 +

1

2u1 +

1

2K(x)u2

x2 = −x1 − x2 − x3

2

y1 = −0.08675x1 + 0.008485x2 − 0.5571926155x3

1+ 1.38155909x2

1x2−

0.950623382x1x2

2+ 0.04010539255x3

2

y2 =1

2x1 −

1

2x2,

with ∂Sr

∂xK(x) = x1 − x2 and

ρ1(s) = 2.506079510 + 69.19812137s2 + o(s4)

ρ2(s) = 0.4508902128 + 0.8176340704s2 + o(s4).

Autumn school Terschelling, 24 September 2009 45/53

University of Groningen

Intermediate conclusions

• Unifying framework for balancing, bounded and positive real

balancing, etc. via Hankel operator approach.

• Both extension and factorization system relations.

• Dissipativity structure preserving methods.

• Application to SMIB system of positive real balancing (8

states), raises new questions important for multi-physics

systems, i.e., what if mechanical part is important from

physics point of view, but not from passivity point of view!

• Missing: good numerical tools for larger scale systems. (NB:

for nonlinear systems 20 states is sometimes already large!)

Autumn school Terschelling, 24 September 2009 46/53

University of Groningen

Outline

• Review stable linear systems balancing

• Review stable nonlinear systems balancing

• Storage functions in terms of observability andcontrollability functions

• Storage functions and coprime factorizations

• Dissipativity preserving model reduction

• Simple example

• Port-Hamiltonian structure preserving reduction toa minimal system

Autumn school Terschelling, 24 September 2009 47/53

University of Groningen

Port-Hamiltonian structure preserving reduction

Consider port-Hamiltonian (PH) system of the form

x = (J(x) − R(x))∂H

∂x(x) + g(x)u

y = g(x)T ∂H

∂x(x)

J(x) = −J(x)T interconnection matrixR(x) = R(x)T ≥ 0 dissipation matrix.

H: Hamiltonian, internal energy of system, u and y ports:

H = uT y −∂T H

∂x(x)R(x)

∂H

∂x(x)

Autumn school Terschelling, 24 September 2009 48/53

University of Groningen

Port-Hamiltonian structure preserving reduction

Properties

• Power preserving interconnection

• Passivity

• H is Lyapunov function, i.e., stability

• Interconnection two pH systems preserves structure,passivity and stability

• Etc.

Autumn school Terschelling, 24 September 2009 48/53

University of Groningen

Port-Hamiltonian structure preserving reduction

Properties

• Power preserving interconnection

• Passivity

• H is Lyapunov function, i.e., stability

• Interconnection two pH systems preserves structure,passivity and stability

• Etc.

How to reduce a non-minimal pH system to a minimal pHsystem?

Autumn school Terschelling, 24 September 2009 49/53

University of Groningen

Port-Hamiltonian structure preserving reduction

• For linear PH systems, see e.g., Polyuga, van derSchaft, 2008.

• Recall: a nonlinear system is minimal (Isidori 95) if thesystem is strongly accessible (controllable for linearsystems) and observable.

First strong accessibility:

• Use nonlinear version of Kalman decomposition, i.e.,under appropriate conditions there exists coordinatessuch that x1 is strongly accessible and x2 is not.

Autumn school Terschelling, 24 September 2009 50/53

University of Groningen

Port-Hamiltonian structure preserving reduction

Suppose that F = J − R, then pH dynamics restricted tostrongly accessible subspace can be written as

x1 =(F11(x

1) − F12(x1)F−1

22 (x1)F21(x1)

)︸ ︷︷ ︸

F (x1)

∂H

∂x1(x1, 0) + g1(x

1)u

y = g1(x1)

∂H

∂x1(x1, 0)

where Fij(x1) = Fij(x

1, 0) for i, j = 1, 2, g1(x1) = g1(x

1, 0),which is again a PH system.

Hamiltonian H(x1, 0). Interconnection and damping F .

Autumn school Terschelling, 24 September 2009 51/53

University of Groningen

Port-Hamiltonian structure preserving reduction

Observability is more complicated. In linear case J11 − R11

stays, and Hamiltonian changes via Schur complement.

In nonlinear case Kalman decomposition can be done, butadditional assumptions are needed:

• If J − R and g are constant matrices, then similar tolinear case.

• If not, then with zero-state observability, via dualitynotions and observability and controllability functions,result can be obtained.

Scherpen and Van der Schaft 2008

Autumn school Terschelling, 24 September 2009 52/53

University of Groningen

General concluding remarks

• Framework for nonlinear balanced realizations inrelation with input-output interpretation from Hankelnorm. Rather complete picture. Extension of linearconcepts were introduced.

Autumn school Terschelling, 24 September 2009 52/53

University of Groningen

General concluding remarks

• Framework for nonlinear balanced realizations inrelation with input-output interpretation from Hankelnorm. Rather complete picture. Extension of linearconcepts were introduced.

• Minimality and PH systems use similar concepts.

Autumn school Terschelling, 24 September 2009 52/53

University of Groningen

General concluding remarks

• Framework for nonlinear balanced realizations inrelation with input-output interpretation from Hankelnorm. Rather complete picture. Extension of linearconcepts were introduced.

• Minimality and PH systems use similar concepts.

• In principle, analytical framework. However,computation of Lo and Lc is “not easy”, (e.g., Krener)and balancing step requires heavy computations,though proofs are constructive. Work in progress.

Autumn school Terschelling, 24 September 2009 52/53

University of Groningen

General concluding remarks

• Framework for nonlinear balanced realizations inrelation with input-output interpretation from Hankelnorm. Rather complete picture. Extension of linearconcepts were introduced.

• Minimality and PH systems use similar concepts.

• In principle, analytical framework. However,computation of Lo and Lc is “not easy”, (e.g., Krener)and balancing step requires heavy computations,though proofs are constructive. Work in progress.

• Many open issues.

Autumn school Terschelling, 24 September 2009 53/53

University of Groningen

1/30

FWN, ITM, IWI

A structure preserving minimalrepresentation of a nonlinearport-Hamiltonian systems

Jacquelien Scherpen

Arjan van der Schaft

University of Groningen

Autumn school Terschelling, 24 September 2009

Autumn school Terschelling, 24 September 2009 2/30

FWN, ITM, IWI

Contribution

Exact model reduction method for a non-observable andnon-strongly accessible port-Hamiltonian system to anobservable and strongly accessible port-Hamiltoniansystem.

The nonlinear version of the Kalman decomposition isinstrumental for the approach. Both descriptions in energyand in co-energy variables are considered, depending onwhich description yields better properties for the reductionstep.

Autumn school Terschelling, 24 September 2009 3/30

FWN, ITM, IWI

Background

• Well-known balanced trunction methods for model order

reduction approximate system by removing the “almost

non-minimal parts” of the state space, e.g. ......

Autumn school Terschelling, 24 September 2009 3/30

FWN, ITM, IWI

Background

• Well-known balanced trunction methods for model order

reduction approximate system by removing the “almost

non-minimal parts” of the state space, e.g. ......

• Order reduction while preserving some energy/power

structure, passivity properties, and/or observability and

controllability properties studied for linear systems in many

papers, e.g., . . . . . .

Autumn school Terschelling, 24 September 2009 3/30

FWN, ITM, IWI

Background

• Well-known balanced trunction methods for model order

reduction approximate system by removing the “almost

non-minimal parts” of the state space, e.g. ......

• Order reduction while preserving some energy/power

structure, passivity properties, and/or observability and

controllability properties studied for linear systems in many

papers, e.g., . . . . . .

• First: exact reduction, i.e., from non-minimal to minimal.

Autumn school Terschelling, 24 September 2009 3/30

FWN, ITM, IWI

Background

• Well-known balanced trunction methods for model order

reduction approximate system by removing the “almost

non-minimal parts” of the state space, e.g. ......

• Order reduction while preserving some energy/power

structure, passivity properties, and/or observability and

controllability properties studied for linear systems in many

papers, e.g., . . . . . .

• First: exact reduction, i.e., from non-minimal to minimal.

• For linear pH systems: from non-controllable /

non-observable to controllable/observable pH system,

Polyuga, Van der Schaft, 2008.

Autumn school Terschelling, 24 September 2009 4/30

FWN, ITM, IWI

Problem setting

Consider port-Hamiltonian (PH) system of the form

x = (J(x) − R(x))∂H

∂x(x) + g(x)u

y = g(x)T∂H

∂x(x)

J(x) = −J(x)T interconnection matrixR(x) = R(x)T ≥ 0 dissipation matrix.

H: Hamiltonian, internal energy of system, u and y ports:

H = uT y −∂T H

∂x(x)R(x)

∂H

∂x(x)

Autumn school Terschelling, 24 September 2009 5/30

FWN, ITM, IWI

Problem setting

Properties

• Power preserving interconnection

• Passivity

• H is Lyapunov function, i.e., stability

• Interconnection two pH systems preserves structure,passivity and stability

• Etc.

Autumn school Terschelling, 24 September 2009 5/30

FWN, ITM, IWI

Problem setting

Properties

• Power preserving interconnection

• Passivity

• H is Lyapunov function, i.e., stability

• Interconnection two pH systems preserves structure,passivity and stability

• Etc.

How to reduce a non-minimal pH system to a minimalpH system?

Autumn school Terschelling, 24 September 2009 6/30

FWN, ITM, IWI

Outline

• Nonlinear systems and minimality

• Reduction to a strongly-accessible pH system

• Reduction to an observable pH system

• Approximate model reduction

• Concluding remarks

Autumn school Terschelling, 24 September 2009 7/30

FWN, ITM, IWI

Nonlinear systems and minimality

Smooth system

x = f(x) + g(x)u

y = h(x)

where u ∈ Rm, y ∈ R

p, and x ∈ M (manifold of dim n).

• Analytic realization (f, g, h) about x0 of formal powerseries is minimal if and only if realization is locallyaccessible and locally observable about x0.

Autumn school Terschelling, 24 September 2009 7/30

FWN, ITM, IWI

Nonlinear systems and minimality

Smooth system

x = f(x) + g(x)u

y = h(x)

where u ∈ Rm, y ∈ R

p, and x ∈ M (manifold of dim n).

• Analytic realization (f, g, h) about x0 of formal powerseries is minimal if and only if realization is locallyaccessible and locally observable about x0.

• Under constant dim. assumption: “nonlinear” Kalmandecomposition for loc. strongly acc./loc. obs. case.

Autumn school Terschelling, 24 September 2009 8/30

FWN, ITM, IWI

Energy functions: Gramian extensions

Lc(x0) = min

u ∈ L2(−∞, 0)

x(−∞) = 0, x(0) = x0

1

2

∫ 0

−∞

‖ u(t) ‖2 dt

Minimum amount of control energy necessary to reachstate x0. Lc is the so-called controllability function .

Autumn school Terschelling, 24 September 2009 8/30

FWN, ITM, IWI

Energy functions: Gramian extensions

Lc(x0) = min

u ∈ L2(−∞, 0)

x(−∞) = 0, x(0) = x0

1

2

∫ 0

−∞

‖ u(t) ‖2 dt

Minimum amount of control energy necessary to reachstate x0. Lc is the so-called controllability function .

Lo(x0) =1

2

∫ ∞

0

‖ y(t) ‖2 dt,x(0)= x0

u(τ)=0, 0≤τ<∞

Output energy generated by state x0.Lo is the so-called observability function.

Autumn school Terschelling, 24 September 2009 9/30

FWN, ITM, IWI

Nonlinear systems and minimality

Assumptions: f(0) = 0, 0 as. stable eq. point for u = 0,x ∈ X, h(0) = 0. Controllability function Lc andobservability function Lo smooth and exist.

Autumn school Terschelling, 24 September 2009 9/30

FWN, ITM, IWI

Nonlinear systems and minimality

Assumptions: f(0) = 0, 0 as. stable eq. point for u = 0,x ∈ X, h(0) = 0. Controllability function Lc andobservability function Lo smooth and exist.

• If 0 < Lc(x) < ∞ and 0 < Lo(x) < ∞ for x ∈ X, x 6= 0,then, under appropriate additional assumptions, thesystem is minimal , Scherpen, Gray 2000.

Autumn school Terschelling, 24 September 2009 10/30

FWN, ITM, IWI

Outline

• Nonlinear systems and minimality

• Reduction to a strongly-accessible pH system

• Reduction to an observable pH system

• Approximate model reduction

• Concluding remarks

Autumn school Terschelling, 24 September 2009 11/30

FWN, ITM, IWI

Reduction to a strongly-accessible pH system

Assume strong acc. distribution C0 const. dim.. Then thereexist local coordinates such that C0 =span{ ∂

∂x1}

x1

x2

=

F11(x) F12(x)

F21(x) F22(x)

∂H

∂x1(x)

∂H

∂x2(x)

+

g1(x)

0

u

y =(g1(x)T 0

)

∂H

∂x1(x)

∂H

∂x2(x)

with F (x) = J(x) − R(x), x1 strongly acc., x2 not.

Autumn school Terschelling, 24 September 2009 12/30

FWN, ITM, IWI

Reduction to a strongly accessible pH system

Assume F22(x1, 0) invertible. Then pH dynamics restricted

to C0 can be written as

x1 =(F11(x

1) − F12(x1)F−1

22 (x1)F21(x1))

︸ ︷︷ ︸F (x1)

∂H

∂x1(x1, 0) + g1(x

1)u

y = g1(x1)

∂H

∂x1(x1, 0)

where Fij(x1) = Fij(x

1, 0) for i, j = 1, 2, g1(x1) = g1(x

1, 0),which is again a pH system.

Hamiltonian H(x1, 0). Interconnection and damping F (x1).

Autumn school Terschelling, 24 September 2009 13/30

FWN, ITM, IWI

Reduction to a stronly accessible pH system

Idea: Since C0 =span{ ∂∂x1},

F21(x)∂H

∂x1(x) + F22(x)

∂H

∂x2(x) = f(x2).

For any value of x2, the x1 sub-system is stronglyaccessible. Plug in x2 = 0, and result is obtained.

Autumn school Terschelling, 24 September 2009 14/30

FWN, ITM, IWI

Outline

• Nonlinear systems and minimality

• Reduction to a strongly-accessible pH system

• Reduction to an observable pH system

• Approximate model reduction

• Concluding remarks

Autumn school Terschelling, 24 September 2009 15/30

FWN, ITM, IWI

Reduction to an observable pH system

Observability is more complicated. In linear caseF11 = J11 − R11 remains, and Hamiltonian changes viaSchur complement.

In nonlinear case Kalman decomposition can be done, butadditional assumptions are needed:

• If J − R and g are in special form, then similar to linearcase.

• If not, then with zero-state observability, via dualitynotions and observability and controllability functions,result can be obtained.

Autumn school Terschelling, 24 September 2009 16/30

FWN, ITM, IWI

Reduction to an observable pH system

Assumptions:

• Observability space O. Observability co-distribution dO

is constant dimensional( ⇒ there exist local coordinates (x1, x2) such thatker dO =span{ ∂

∂x2} ).

• Assume F and g are such that F11, F12, g1, and g2 onlydepend on x1.

•∂2H

(∂x2)2(x) is invertible for all x.

Then, the pH system takes the form

Autumn school Terschelling, 24 September 2009 17/30

FWN, ITM, IWI

Reduction to an observable pH system

x1

x2

=

F11(x1) F12(x

1)

F21(x) F22(x)

∂H

∂x1(x)

∂H

∂x2(x)

+

g1(x1)

g2(x1)

u

y =(gT1 (x1) gT

2 (x1))(

∂H

∂x1(x)

∂H

∂x2(x)

)T

where

F11(x1)

∂H

∂x1(x) + F12(x

1)∂H

∂x2(x) = f(x1)

gT1 (x1)

∂H

∂x1(x) + gT

2 (x1)∂H

∂x2(x) = h(x1)

Autumn school Terschelling, 24 September 2009 18/30

FWN, ITM, IWI

Reduction to an observable pH system

Differentiating to x2 yields that the pH system restricted toobservable part can be written as

x1 = F11(x1)

∂H

∂x1(x1) + g1(x

1)u

y = g1(x1)T ∂H

∂x1(x1)

where∂H

∂x2(x1, x2) = 0

can be solved (at least locally) for x2 as a function x2(x1),determining H(x1) := H(x1, x2(x1)).

Autumn school Terschelling, 24 September 2009 19/30

FWN, ITM, IWI

Reduction to an observable pH system

What if assumption on g and F is not fulfilled? Thenconsider pH system in “physically dual” co-energycoordinates, i.e.,

z =∂H

∂x(x) =: γ(x),

under the assumption that transformation is non-singular.Take H(z) as the full Legendre transform of H(x), i.e.,

H(z) = xT z − H(x),

Autumn school Terschelling, 24 September 2009 20/30

FWN, ITM, IWI

Reduction to an observable pH system

• For system in z coordinates observability functionLo(z) is considered. Then duality (see Fujimoto,

Scherpen, Gray, 2002) via Legendre transform of Lo canbe considered, resulting in characterizations via Lc.

Autumn school Terschelling, 24 September 2009 20/30

FWN, ITM, IWI

Reduction to an observable pH system

• For system in z coordinates observability functionLo(z) is considered. Then duality (see Fujimoto,

Scherpen, Gray, 2002) via Legendre transform of Lo canbe considered, resulting in characterizations via Lc.

• Then strong accessibility result can be applied,resulting in restriction to strongly accessible system.

Autumn school Terschelling, 24 September 2009 20/30

FWN, ITM, IWI

Reduction to an observable pH system

• For system in z coordinates observability functionLo(z) is considered. Then duality (see Fujimoto,

Scherpen, Gray, 2002) via Legendre transform of Lo canbe considered, resulting in characterizations via Lc.

• Then strong accessibility result can be applied,resulting in restriction to strongly accessible system.

• Dual system via Legendre transform of Lc then resultsin a zero-state observable system that has the form ofa co-energy variable presentation.

Autumn school Terschelling, 24 September 2009 20/30

FWN, ITM, IWI

Reduction to an observable pH system

• For system in z coordinates observability functionLo(z) is considered. Then duality (see Fujimoto,

Scherpen, Gray, 2002) via Legendre transform of Lo canbe considered, resulting in characterizations via Lc.

• Then strong accessibility result can be applied,resulting in restriction to strongly accessible system.

• Dual system via Legendre transform of Lc then resultsin a zero-state observable system that has the form ofa co-energy variable presentation.

• Open issue: link co-energy to Hamiltonian.

Autumn school Terschelling, 24 September 2009 21/30

FWN, ITM, IWI

Reduction to an observable pH system

System in co-energy coordinates:

z =

(∂2H

∂z2(z)

)−1 (J(z) − R(z)

)z

+

(∂2H

∂z2(z)

)−1

g(z)u

y = g(z)T z

with J(z) := J(γ−1(z)), R(z) := R(γ−1(z)), andg(z) := g(γ−1(z)).

Split in z1 (zero-observable) and z2 (not zero-observable).

Autumn school Terschelling, 24 September 2009 22/30

FWN, ITM, IWI

Reduction to an observable pH system

Lo(0, z2) = 0. Assume F22(z) invertible and Lo(z

1).Then ∃ coordinates z = ξ(z) s.t. co-energy variabledynamics restricted zero observable part

˙z = Q11(z)(F11(z) − F12(z)

(F22(z)

)−1F21(z)

)

︸ ︷︷ ︸F (z)

z+Q11(z)g1(z)u

y = gT1 (z)z

with Q11(z) = Q11(ξ(z)) =∂2H

(∂x1)2(γ−1(ξ(z))

Fij(z) = Fij(ξ(z)). F (z) + F T (z) ≤ 0.

Autumn school Terschelling, 24 September 2009 23/30

FWN, ITM, IWI

Outline

• Nonlinear systems and minimality

• Reduction to a strongly-accessible pH system

• Reduction to an observable pH system

• Approximate model reduction

• Concluding remarks

Autumn school Terschelling, 24 September 2009 24/30

FWN, ITM, IWI

Approximate model reduction

Can these “Kalman like” decompositions be used forfurther reduction of the model, similar to balancedtruncation, but preserving the pH structure?

Autumn school Terschelling, 24 September 2009 24/30

FWN, ITM, IWI

Approximate model reduction

Can these “Kalman like” decompositions be used forfurther reduction of the model, similar to balancedtruncation, but preserving the pH structure?

• Note that the observability and strong accessibilityreduction methods to a minimal pH systems both yielddifferent minimal pH models.

Hence, immediate generalization of balancedtruncation in this setting is not clear.

Autumn school Terschelling, 24 September 2009 25/30

FWN, ITM, IWI

Approximate model reduction

The observability reduction for special form F and g ismost easily applied for approximation, i.e.,

x1 = F11(x1, x2(x1)) ∂ eH

∂x1 (x1) + g1(x

1, x2(x1))u

y = g1(x1, x2(x1))T ∂ eH

∂x1 (x1)

H = H(x1, x2(x1)), with x2(x1) solution of ∂H∂x2 (x

1, x2) = 0.

Autumn school Terschelling, 24 September 2009 25/30

FWN, ITM, IWI

Approximate model reduction

The observability reduction for special form F and g ismost easily applied for approximation, i.e.,

x1 = F11(x1, x2(x1)) ∂ eH

∂x1 (x1) + g1(x

1, x2(x1))u

y = g1(x1, x2(x1))T ∂ eH

∂x1 (x1)

H = H(x1, x2(x1)), with x2(x1) solution of ∂H∂x2 (x

1, x2) = 0.

Effort constraint reduction, i.e., setting the “effort”

∂H

∂x2= 0.

Autumn school Terschelling, 24 September 2009 26/30

FWN, ITM, IWI

Approximate model reduction

Reduction via strongly accessible method also possible, but

technically more involved.x1

x2

=

F11 F12

F21 F22

e1

e2

+

g1

g2

u

y =(gT1 gT

2

)e1

e2

where e1 = ∂H∂x1 (x), e2 = ∂H

∂x2 (x). Set x2 equal to zero (flowconstraint ). If F22 is invertible, this yields

0 = F21e1 + F22e

2 + g2u ⇒ e2 = −F−122 F21e

1 − F−122 g2u

Autumn school Terschelling, 24 September 2009 27/30

FWN, ITM, IWI

Approximate model reduction

Substitution flow constraint and x2 = 0 yield

x1 =(F11(x

1) − F12(x1)F−1

22 (x1)F21(x1)) ∂H

∂x1(x1, 0)

+(g1(x1, 0) − F12(x

1)F−122 (x1)g2(x

1, 0))u

y = (gT1 (x1, 0) − gT

2 (x1, 0)F−122 (x1)F21(x

1)

−gT2 (x1, 0)F−1

22 (x1)g2(x1, 0)u

with Fij(x1) = Fij(x

1, 0), i, j = 1, 2.

Again pH system (with through-term) provided that

(F12F−122 )T = F−1

22 F21.

Autumn school Terschelling, 24 September 2009 28/30

FWN, ITM, IWI

Outline

• Nonlinear systems and minimality

• Reduction to a strongly-accessible pH system

• Reduction to an observable pH system

• Approximate model reduction

• Concluding remarks

Autumn school Terschelling, 24 September 2009 29/30

FWN, ITM, IWI

Concluding remarks

• Reduction of non-strongly acc. or non-observable pHsystem to strongly acc. or observable pH system,respectively.

• Observability least straightforward, additionalassumptions made, and zero-observability considered.

Autumn school Terschelling, 24 September 2009 29/30

FWN, ITM, IWI

Concluding remarks

• Reduction of non-strongly acc. or non-observable pHsystem to strongly acc. or observable pH system,respectively.

• Observability least straightforward, additionalassumptions made, and zero-observability considered.

Open issues

• Zero-observability pH structure from co-energyrespresentation.

• Which method to use for approximate model reduction.Balance possible?

Autumn school Terschelling, 24 September 2009 30/30

FWN, ITM, IWI

Recommended