Upload
others
View
11
Download
0
Embed Size (px)
Citation preview
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Infinite dimensional port Hamiltoniansystems
Y. Le Gorrec1
1FEMTO-ST, Automatic Control and Micro-Mechatronic Systems Department
CAS Seminar - Topic : PDE
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
1 IntroductionContextOutline
2 Infinite dimensional linear 1-D caseA simple exampleConsidered class of systemsDirac structure and PHS on Hilbert SpaceBoundary control systemsStability, control feedback
3 Model reductionMixed finite element methodSpectral properties
4 Conclusion
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Context
Recent technological progresses and physical knowledge allow togo toward the use of complex systems :
Highly nonlinear.
Involving numerous physical domains and possibleheterogeneity.
With distributed parameters or organized in network.
New issue for system control theory
Modelling step is important → the physical properties can beadvantageously used for analysis, control or simulation purposes
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Context
Recent technological progresses and physical knowledge allow togo toward the use of complex systems :
Highly nonlinear.
Involving numerous physical domains and possibleheterogeneity.
With distributed parameters or organized in network.
New issue for system control theory
Modelling step is important → the physical properties can beadvantageously used for analysis, control or simulation purposes
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Context
Example 1 : Ionic Polymer Metal CompositeParcours
Activitésd’enseignement
Activités derecherche
Projetsd’intégration•ENSMM•FEMTO-ST/AS2M/SAMMI
Candidature au Poste de Professeur des Universités PU 61 - 0843, 2008 ENSMM p. 15/15
! Exemple de projet support: ProjetFranco-Japonais SAKURA
! Mise en oeuvre pratique - Pilote
Electromechanical system.
3 scales : Polymer-electrode interface, diffusion in thepolymer, beam bending.
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Context
Example 2 : Nanotweezer for DNA manipulation
Multiphysic system.
Infinite dimensional system.
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Context
Example 3 : Adsorption process
Parcours
Activitésd’enseignement
Activités derecherche•Motivation•Développementsthéoriques
•Projets -Encadrements- Publications
Projetsd’intégration
Candidature au Poste de Professeur des Universités PU 61 - 0843, 2008 ENSMM p. 9/15
Procédé d’adsorption par modulation de pression (A. Baaiu PHD)
Objectif : séparation par adsorption
(coll. IFP)
A
B
A B
B
! Système hétérogène
multi-échelles, régi par
thermodynamique irréversible
! EDP non linéaires
Microporous crystal
Bidispersepellet
rp z
rc
L!25 cm
R !1,24 mm
R !1 µmc
p
Extra granular phase
Macropore scale
Micropore scale
R !1 cmint
Multiscale heterogeneous system.
Dynamic behavior driven by irreversible thermodynamic laws
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Context
Example 3 : Adsorption process
Parcours
Activitésd’enseignement
Activités derecherche•Motivation•Développementsthéoriques
•Projets -Encadrements- Publications
Projetsd’intégration
Candidature au Poste de Professeur des Universités PU 61 - 0843, 2008 ENSMM p. 9/15
Procédé d’adsorption par modulation de pression (A. Baaiu PHD)
Objectif : séparation par adsorption
(coll. IFP)
A
B
A B
B
! Système hétérogène
multi-échelles, régi par
thermodynamique irréversible
! EDP non linéaires
Microporous crystal
Bidispersepellet
rp z
rc
L!25 cm
R !1,24 mm
R !1 µmc
p
Extra granular phase
Macropore scale
Micropore scale
R !1 cmint
Multiscale heterogeneous system.
Considered phenomena :
Fluid scale : convection, dispersion.Pellet scale : diffusion (Stephan-Maxwell).Microscopic scale : Knudsen law.
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Outline
A unified approach : the port Hamiltonian framework
Emphasis the geometric structure related to power exchanges powerful for analysis, model reduction and control.
Today : Infinite dimensional linear 1-D case and discretization
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Outline
A unified approach : the port Hamiltonian framework
Emphasis the geometric structure related to power exchanges powerful for analysis, model reduction and control.
Today : Infinite dimensional linear 1-D case and discretization
!
!"#$$
%&'$!()$
%&'$
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Outline
A unified approach : the port Hamiltonian framework
Emphasis the geometric structure related to power exchanges powerful for analysis, model reduction and control.
Today : Infinite dimensional linear 1-D case and discretization
Other works :
Infinite dimensional linear 2-D, 3-D cases (joint work with B.Maschke and H. Zwart).
Infinite dimensional port Hamiltonian systems 2D, 3D cases(cf A.v.d. Schaft and B. Maschke).
Infinite dimensional NL systems : Passivity and PHS vs.Stability and Riemann Invariants (Joint work : V. DosSantos, B. Maschke).
Non linear control : IDA-PBC, Entropy based control ofChemical reactors (Joint work : F. Couenne).
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
A simple example
1 IntroductionContextOutline
2 Infinite dimensional linear 1-D caseA simple exampleConsidered class of systemsDirac structure and PHS on Hilbert SpaceBoundary control systemsStability, control feedback
3 Model reductionMixed finite element methodSpectral properties
4 Conclusion
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
A simple example
Vibrating string :!
!"#"
u(t,z)
The classical modelling is based on the wave equation : Newton’slaw + Hooke’s law (restoring force proportional to the deformation)
∂2u(z , t)
∂t2=
1
µ(z)
∂
∂z
(T (z)
∂u(z , t)
∂z
)
The structure of the model is not apparent. How to choose theboundary conditions ? ? ?
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
A simple example
Let choose as state variables the energy variables :
the strain ε = ∂u(z,t)∂z
the elastic momentum p = µ(z)v(z , t)
The total energy is given by : H(ε, p) = U(ε) + K (p)
U(ε) is the elastic potential energy :
U(ε) =
∫ b
a
1
2T (z)
(∂u(z , t)
∂z
)2
=
∫ b
a
1
2T ε(z , t)2
where T (z) denotes the elastic modulus.
K (v) is the kinetic co-energy :
K (p) =
∫ b
a
1
2µ(z)v (z , t)2 =
∫ b
a
1
2
1
µ(z)p2(z , t)
where µ(z) denotes the string mass.
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
A simple example
The vector of fluxes is given by :
β =
(v(t, z)σ
)where v(z , t) is the velocity and σ(z , t) = T (z)ε(z , t) the stress.The vector of fluxes β may be expressed in term of the generatingforces :
β =
(0 11 0
)︸ ︷︷ ︸
( δHδεδHδp
)︸ ︷︷ ︸
canonical generatinginerdomain coupling forces
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
A simple example
From the conservation laws :
∂
∂t
(εp
)+
∂
∂z
(vσ
)= 0
Consequently
∂
∂t
(εp
)= − ∂
∂z
(0 11 0
)( δHδεδHδp
)and
PDEs :
∂
∂t
(εp
)=
(0 − ∂
∂z
− ∂∂z 0
)( δHδεδHδp
)⇔ ∂2u(z , t)
∂t2= σ2∂
2u(z , t)
∂z2
+BC
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
A simple example
Underlying structure :
∂
∂t
(εp
)︸ ︷︷ ︸ =
(0 − ∂
∂z
− ∂∂z 0
)︸ ︷︷ ︸
(T (z) 0
0 1µ(z)
)(εp
)︸ ︷︷ ︸
f J = matrix e = drivingdifferential operator force
Hamiltonian operator J is skew-symmetric only for functionwith compact domain strictly in Z :∫ b
a
(e1 e2
)J(
e′1e′2
)+(
e′1 e′2)J(
e1
e2
)= − [e1e
′2 + e2e
′1]
ba
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
A simple example
Power balance equation :
ddt H(ε, p) =
∫ ba
(δHδε
∂ε∂t + δH
δp∂p∂t
)dz
= −∫ ba
(δHδε
∂∂z
δHδp + δH
δp∂∂z
δHδε
)dz
= −[δHδε
δHδp
]ba
If driving forces are zero at the boundary, the total energy isconserved, else there is a flow of power at the boundary. Definetwo port boundary variables as follows :(
f∂e∂
)=
( δHδεδHδp
)|a,b
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
A simple example
The linear space D 3 (f1, f2, e1, e2, f∂ , e∂)(f1e2
)=
(0 − ∂
∂z
− ∂∂z 0
)(e1
e2
)(
f∂e∂
)=
(e1
e2
)|a,b
defines a Dirac structure :D = D⊥ with respect to the symmetricpairing : ∫ b
ae1f1dz +
∫ b
ae2f2dz + [f∂e∂ ]ba
Port Hamiltonian system(∂
∂tα,δH
δα, f∂ , e∂
)∈ D
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Considered class of systems
1 IntroductionContextOutline
2 Infinite dimensional linear 1-D caseA simple exampleConsidered class of systemsDirac structure and PHS on Hilbert SpaceBoundary control systemsStability, control feedback
3 Model reductionMixed finite element methodSpectral properties
4 Conclusion
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Considered class of systems
Considered class of systems :
∂x
∂t(t, z) = (J − GRSG∗R)L(z)x(t, z), x(0, z) = x0(z),
m(ffp
)= Je
(eep
)=
(J GR
−G∗R 0
)(eep
)with ep = Sfp where S is a coercive operator
„ffp
«∈ F ,
„eep
«∈ E and E = F = L2((a, b),Rn)× L2((a, b),Rn)
Covers models of : beams, wave, plates, (with or without damping)and also systems of diffusion/convection, chemical reactors ...
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Dirac structure and PHS on Hilbert Space
1 IntroductionContextOutline
2 Infinite dimensional linear 1-D caseA simple exampleConsidered class of systemsDirac structure and PHS on Hilbert SpaceBoundary control systemsStability, control feedback
3 Model reductionMixed finite element methodSpectral properties
4 Conclusion
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Dirac structure and PHS on Hilbert Space
The system is defined by :
f = J e
• Let the space of flow variables, F , and the space of effort
variables,E , be real Hilbert spaces.• Define the space of bond variables as B = F × E endowed by thenatural inner product⟨b1, b2
⟩=⟨f 1, f 2
⟩F+⟨e1, e2
⟩E , b1 =
(f 1, e1
), b2 =
(f 2, e2
)∈ B.
In order to define a Dirac structure, let us moreover endow thebond space B with a canonical symmetrical pairing, i.e., a bilinearform defined as follows :⟨b1, b2
⟩+
=⟨f 1, rE,Fe2
⟩F+⟨e1, rF,E f
2⟩E , b
1 =(f 1, e1
), b2 =
(f 2, e2
)∈ B.
(1)
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Dirac structure and PHS on Hilbert Space
Denote by D⊥ the orthogonal subspace to D with respect to thesymmetrical pairing :
D⊥ =
b ∈ B|⟨b, b′
⟩+
= 0 for all b′ ∈ D. (2)
Definition : A Dirac structure D on the bond space B = F×Eis a subspace of B which is maximally isotropic with respect tothe canonical symmetrical pairing, i.e.,
D⊥ = D. (3)
(fe
)∈ D ⇐⇒ Power conservation
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Dirac structure and PHS on Hilbert Space
PHS Definition based on Dirac structure and Hamiltonianfunction (total energy of the system).
Definition : Let B = E ×F be the bound space defined above andconsider the Dirac structure D and the Hamilonian function H(x)with x the energy variables. Define the flow variables, f ∈ F as thetime variation of the energy variables and the effort variables e ∈ Eas the variational derivative of H(x). The system
(f , e) =
(∂x
∂t,δHδx
)∈ D
is a Port Hamiltonian system with total energy H(x)
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Dirac structure and PHS on Hilbert Space
Parametrization :
J e =N∑
i=0
P(i)d ie
dz i(z) z ∈ [a, b] ,
where e ∈ HN((a, b); Rn) and P(i), i = 0, . . . ,N, is a n × n realmatrix with PN non singular and Pi = PT
i (−1)i+1. Let define
Q =
0BBB@P1 P2 · · · PN
−P2 −P3 · · · 0... · · ·
. . ....
(−1)N−1PN 0 · · · 0
1CCCABack to the Vibrating string
∂
∂t
„εp
«| z =
„0 −1−1 0
«| z
∂∂z
T (z) 0
0 1µ(z)
!„εp
«| z
f P1 e
,Q = P1
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Dirac structure and PHS on Hilbert Space
We define the symmetrical pairing (not depending on J ) of andthe port variables associated with J .Let F = E = L2((a, b); Rn)× RnN and define B = F × E with thefollowing canonical symmetrical pairing :⟨(
f 1, f 1∂ , e
1, e1∂
) (f 2, f 2
∂ , e2, e2
∂
)⟩+
= 〈e1, f 2〉L2 + 〈e2, f 1〉L2 − 〈e1∂ , f
2∂ 〉 − 〈e2
∂ , f1∂ 〉,
Definition : The port variables (e∂ , f∂) ∈ RnN associated with Jare defined by :
„f∂e∂
«= Rext
0BBBBBBBBBB@
e(b)...
dN−1edzN−1 (b)
e(a)...
dN−1edzN−1 (a)
1CCCCCCCCCCA, Rext =
1√
2
`Q −QI I
´
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Dirac structure and PHS on Hilbert Space
The subspace DJ of B defined as
DJ =
8>><>>:0BB@
ff∂ee∂
1CCA ˛ e ∈ HN((a, b); Rn),J e = f ,
„f∂e∂
«= Rext
0BB@e(b)
...
∂N−1z e(a)
1CCA9>>=>>;
is a Dirac structure, that means that D = D⊥.
Back to the Vibrating string
∂
∂t
(εp
)︸ ︷︷ ︸ =
(0 −1−1 0
)︸ ︷︷ ︸ ∂
∂z
(T (z)ε
1µ(z)p
)︸ ︷︷ ︸
f P1 e
,Q = P1
„f∂e∂
«=
1√
2
„P1 −P1
I I
«„e(b)e(a)
«=
1√
2
0BBB@T (a)ε(a)− T (b)ε(b)
p(a)µ(a)− p(b)µ(b)
T (a)ε(a) + T (b)ε(b)p(a)µ(a)
+ p(b)µ(b)
1CCCA
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Boundary control systems
1 IntroductionContextOutline
2 Infinite dimensional linear 1-D caseA simple exampleConsidered class of systemsDirac structure and PHS on Hilbert SpaceBoundary control systemsStability, control feedback
3 Model reductionMixed finite element methodSpectral properties
4 Conclusion
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Boundary control systems
Boundary Control Systems :Theorem : Let W be a nN × 2nN full rank matrix. The system
x(t) = JLx(t)
u(t) = Bx(t) = W
(f∂(t)e∂(t)
)is a boundary control system, where AW = (JL)kerB is the
generator of a contraction semigroup on L2 ((a, b) ,Rn) if and onlyif
W ΣW T ≥ 0 where Σ =
(0 II 0
)
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Boundary control systems
Let define the linear mapping C : HN((a, b),Rn)→ RnN as
Cx(t) := W
(f∂(t)e∂(t)
)and the output as y(t) = Cx(t), then for u ∈ C 2((0,∞); RnN) andx(0)− Bu(0) ∈ D(JW ) the following balance equation is satisfied :
1
2
d
dt‖x(t)‖2 =
(uT (t) yT (t)
)PW
(u(t)y(t)
).
where PW ,fW =
(W ΣW T W ΣW
W T ΣW T W ΣW T
)−1
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Boundary control systems
Using W = S (I + V , I − V ) :
V = 0
x(t) = J x(t),u(t) = 1
2 (f∂(t) + e∂(t))y(t) = 1
2 (f∂(t)− e∂(t))=⇒
boundary control system, withthe associated semigroupa contraction12
ddt ‖x(t)‖2 = ‖u(t)‖2 − ‖y(t)‖2.
V = I
x(t) = J x(t)u(t) = f∂(t)y(t) = −e∂(t)
=⇒boundary control system, withthe associated semigroupunitary12
ddt ‖x(t)‖2 = u(t)T y(t)
Vibrating string :
V = 0 ⇒ u =1√2
(T (a)ε(a)
p(a)µ(a)
)and y =
1√2
(−T (b)ε(b)
− p(b)µ(b)
)
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Stability, control feedback
1 IntroductionContextOutline
2 Infinite dimensional linear 1-D caseA simple exampleConsidered class of systemsDirac structure and PHS on Hilbert SpaceBoundary control systemsStability, control feedback
3 Model reductionMixed finite element methodSpectral properties
4 Conclusion
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Stability, control feedback
Control using feedback at the boundary!
"#$%!&''#!()*+$,!
!!
!" #" $"
%""""""""&"
Open loop system Closed loop systemx = JLxu = Wimp
(f∂e∂
)y = Cimp
(f∂e∂
)
x = JLxr = (Wimp + αCimp)
(f∂e∂
)y = Cimp
(f∂e∂
)
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Stability, control feedback
The closed loop system described byx = JLxr = (Wimp + αCimp)
(f∂e∂
)y = Cimp
(f∂e∂
)is a boundary control system. Furthermore, the operatorAs = JL|D(As) generates a contraction semigroup onX = L2 ((a, b); Rn) where
D(As) =
x ∈ D(J )|
(f∂e∂
)∈ kerW
and W = (Wimp + αCimp) is a full rank nN × 2nN matrix
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Stability, control feedback
Theorem : (Assumption (λ− As)−1 : X → X is a compactoperator for λ > 0). Then the system described by (VV T = 0) :
x = J x
r = (Wimp + αCimp)
(f∂e∂
)y = Cimp
(f∂e∂
)with r = 0 and α > 0 is globally asymptotically stable. For anyx(0) ∈ X the unique (classical or weak) solution x(t) = T (t)x(0)of the closed loop system asymptotically approaches to zero, i.e.
lim∞‖x(t)‖X = 0
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Stability, control feedback
1 IntroductionContextOutline
2 Infinite dimensional linear 1-D caseA simple exampleConsidered class of systemsDirac structure and PHS on Hilbert SpaceBoundary control systemsStability, control feedback
3 Model reductionMixed finite element methodSpectral properties
4 Conclusion
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Stability, control feedback
Let us restrict the previous class to :
∂x
∂t(t, z) = P1
∂z
∂t(Lx)(t, z) + (P0 − G0)Lx(t, z)
Let consider a BCS as defined previously, assuming thatu(t) = 0, ∀t ≥ 0. Then the system is exponentially stable if
either ‖ (Lx(b)) ‖2R ≤ k1 (〈αy , y〉R + 〈G0Lx(t),Lx(t)〉R)
or ‖ (Lx(a)) ‖2R ≤ k1 (〈αy , y〉R + 〈G0Lx(t),Lx(t)〉R)
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Stability, control feedback
As state variables we choose
x1 = ∂w∂z − φ : shear displacement,
x2 = ρ∂w∂t : transverse momentum distribution,
x3 = ∂φ∂z : angular displacement,
x4 = Iρ∂φ∂t : angular momentum distribution.
Then the model of the beam can be rewritten as
∂
∂t
0BB@x1
x2
x3
x4
1CCA =
26640 1 0 01 0 0 00 0 0 10 0 1 0
3775| z
P1
∂
∂z
0BBB@K x11ρ
x2
EI x31Iρ
x4
1CCCA+
26640 0 0 −10 0 0 00 0 0 01 0 0 0
3775| z
P0
0BBB@K x11ρ
x2
EI x31Iρ
x4
1CCCA| z
x
.
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Stability, control feedback
One can define the boundary port variables :
„f∂e∂
«=
1√
2
»P1 −P1
I I
–„(Lx)(b)(Lx)(a)
«=
1√
2
0BBBBBBBBBB@
(ρ−1x2)(b)− (ρ−1x2)(a)(Kx1)(b)− (Kx1)(a)
(I−1ρ x4)(b)− (I−1
ρ x4)(a)(EIx3)(b)− (EIx3)(a)(Kx1)(b) + (Kx1)(a)
(ρ−1x2)(b) + (ρ−1x2)(a)(EIx3)(b) + (EIx3)(a)
(I−1ρ x4)(b) + (I−1
ρ x4)(a)
1CCCCCCCCCCA.
(4)
Let us consider stabilization by applying velocity feedback i.e.following BC :
1ρ(a) x2(a) = 0, 1
Iρ(a)x4(a) = 0,
K (b)x1(b, t) = −α11
ρ(b)x2(b, t), EI (b)x3(b, t) = −α21
Iρ(b) x4(b)
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Stability, control feedback
Input mapping :
W =1√
2
2664−1 0 0 0 0 1 0 0
0 0 −1 0 0 0 0 1α1 1 0 0 1 α1 0 0
0 0 α2 1 0 0 1 α2
3775 then W ΣW T = 2
26640 0 0 00 0 0 00 0 α1 00 0 0 α2
3775As output we can choose
y =
0BBB@−K(a)x1(a)−(EI )(a)x3(a)
1ρ(b)
x2(b)1
Iρ(b)x4(b)
1CCCA , with fW =1√
2
26640 1 0 0 −1 0 0 00 0 0 1 0 0 −1 01 0 0 0 0 1 0 00 0 1 0 0 0 0 1
3775 .
Then
P−1W ,W =
[2α I
I 0
],PW ,W =
[0 II −2α
]
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Stability, control feedback
Energy balance :
d
dtE (t) =
d
dt‖x(t)‖2
L = 〈u(t), y(t)〉U − 〈αy(t), y(t)〉R
where
〈αy(t), y(t)〉R = α1|(ρ−1x2)(b, t)|2 + α2|(I−1x4)(b, t)|2
Then
‖ (Lx(b)) ‖2R = |(kx1)(b)|2 + |(ρ−1x2)(b)|2 + |(EIx3)(b)|2 + |(I−1
ρ x4)(b)|2= (α2
1 + 1)|(ρ−1x2)(b, t)|2 + (α22 + 1)|(I−1
ρ x4)(b)|2≤ κ〈αy(t), y(t)〉R
⇒ Stability
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Mixed finite element method
1 IntroductionContextOutline
2 Infinite dimensional linear 1-D caseA simple exampleConsidered class of systemsDirac structure and PHS on Hilbert SpaceBoundary control systemsStability, control feedback
3 Model reductionMixed finite element methodSpectral properties
4 Conclusion
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Mixed finite element method
Back to the example of the adsorption column :
Extragranular phase
Adsorption column
Pellet
Macropore
Phase
Flowing fluid
Classical formumlation.
∂q(z,t)∂t
+ ∂N∂z
+ αFads = 0∂qp(z,t)
∂t− Fads = 0
(5)
q mole density in the fluide phase, qp in the adsorbed phase, total fluxN = Nconv + Ndisp = vq − D ∂µ
∂z, Fads = k1a(µ− µp)
with Dankwert boundary conditions N|0 = vqin et ∂q∂z|L = 0
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Mixed finite element method
Back to the example of the adsorption column :
Classical formumlation.
∂q(z,t)∂t
+ ∂N∂z
+ αFads = 0∂qp(z,t)
∂t− Fads = 0
(5)
with Dankwert boundary conditions N|0 = vqin et ∂q∂z|L = 0
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Mixed finite element method
Back to the example of the adsorption column :
Classical formumlation.
∂q(z,t)∂t
+ ∂N∂z
+ αFads = 0∂qp(z,t)
∂t− Fads = 0
(5)
with Dankwert boundary conditions N|0 = vqin et ∂q∂z|L = 0
Port Hamiltonian formulation2664fqFdisp
αfqp
Fads
3775 =
26640 d 0 −1d 0 0 00 0 0 11 0 −1 0
37752664µNµp
Nads
3775 (6)
with constitutive relations
N = −D ∗ Fdisp + v ∗ q = −D ∗ Fdisp + vbµ
Nads = α a k1 Fads
µ = bq(7)
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Mixed finite element method
Back to the example of the adsorption column :
Classical formumlation.
∂q(z,t)∂t
+ ∂N∂z
+ αFads = 0∂qp(z,t)
∂t− Fads = 0
(5)
with Dankwert boundary conditions N|0 = vqin et ∂q∂z|L = 0
Dirac structure : interconnection structure2664fqFdisp
αfqp
Fads
3775 =
26640 d 0 −1d 0 0 00 0 0 11 0 −1 0
37752664µNµp
Nads
3775 (6)
with the boundary port variables
„f∂e∂
«=
0BB@N(0)N(L)µ(0)−µ(L)
1CCA
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Mixed finite element method
DA R10
adsF
dN
t
q
pellet
accumulationinterconnection
dissipation
d
N
0
L
0N
LNConstitutive
thermodynamical law
Constitutive dispersion
law
)(qf )(dfNdispN
0
d
dconvN
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Mixed finite element method
DA R10
adsF
dN
t
q
pellet
accumulationinterconnection
dissipation
d
N
0
L
0N
LNConstitutive
thermodynamical law
Constitutive dispersion
law
)(qf )(dfNdispN
0
d
dconvN
Goal of the discretization scheme : preserve the structure
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Mixed finite element method
Remark 3.1 The Hodge star operator ! is a linear operator mapping p forms on V to (n" p)forms i.e. :
! : !p(V ) # !n!p(V )
In cartesian coordinates, consider the functions g(z) and the 1-form f(z) = g(z)dz then !f(z) =g(z). qL
s and qL being 1-forms, !qLs and !qL are 0-forms.
(3) Closure equation associated with the di!usion - We use Knudsen law [19] to presentthe di"usion in the adsorbed phase (microporous scale). That is to say we only consider thefriction exerted by the solid on the adsorbed species. So the constitutive relation representingthe di"usion is:
fmic2 = "Dmic ! qL
RT! emic
2 (15)
where Dmic is the di"usion constant.
4 Model reduction based on geometrical properties
4.1 General concept
As previously mentioned, the port based modelling of the adsorption column is based on basicelements having well defined energetic behavior as already depicted in Figure 3. The dis-tributed aspect of this port based model is essentially supported by the Dirac structure whichlinks power exchanges within the spatial domain and through the boundaries. The proposeddiscretization method consists in splitting the initial structured infinite dimensional model intoN finite dimensional sub-models (finite elements) with the same energetic behavior (cf Figure 4).Furthermore, the support functions used for the interpolation of both e"ort and flow variablesare di"erent to have enough degrees of freedom to guarantee the conservation of the structuralproperties.
e1
ab
f1
ab
e2
ab
f2
ab-f
1
ab
e1
ab -e2
ab
f2
ab
e1
a
e1
b
f2
a
f2
b
DabC Rab
a b
0
Rmic
a b
Power flow exchanged with
the following volume
Power flow exchanged with the previous volume
0 1
Figure 4: Principle of the spatial discretization
The interconnection between these sub-models is done using the power conjugate boundaryport variables. They correspond to the energy flowing across the boundary of one submodel tothe boundary of the next sub-model. To each submodel is associated the same generic structure(and consequently parts of the global mass and energy balances) as the global structure, thedi"erence lying in the fact that submodels are finite dimensional i.e. the Dirac structure Dab
on Figure 4 is finite dimensional and the reduced e"ort and flow variables (eab, fab) are no morespatially distributed.
For simplicity reasons superscript mic will be omitted in the remaining of the section. Letus now explain the reduction scheme.
7
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Mixed finite element method
The variables dN = Nd and dµ = µd are one-forms and are approximated onΩab by :
Nd(t, z) = Nabd (t) wNd
ab (z)
e2(t, z) = µabd (t) wµd
ab (z)(7)
where the support one-forms wNdab and wµd
ab satisfy :R
ΩabwNd
ab =R
Ωabwµd
ab = 1.The variables µ and N are zero-forms and are approximated on Ωab by :
µ(t, z) = µ(a)wµa (z) + µ(b)wµ
b (z),
N(t, z) = N(a)wNa (z) + N(b)wN
b (z)(8)
where the support zero-forms wµa ,wµ
b ,wNa et wN
b satisfy :
wµa (a) = 1, wµ
a (b) = 0, wµb (a) = 0, wµ
b (b) = 1wN
a (a) = 1, wNa (b) = 0, wN
b (a) = 0, wNb (b) = 1,
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Mixed finite element method
The approximated variables must verify the relation induced by theinterconnection structure :
Nd = dNµd = dµ
=⇒
8>><>>:Nab
d = N(b)− N(a) = Nb∂(t) − Na
∂(t)
µabd = µ(b)− µ(a) = − µb
∂(t) − µa∂(t)
Write the net power using the previous interpolation formulae :
Pnetab = µab Nab
d + µabd Nab +
hµb∂N
b∂ − µa
∂Na∂
iand identify with the real one, we obtain for µab and Nab :
µab = αabµa∂(t)− (1− αab)µb
∂(t)
Nab = (1− αab)Na∂(t) + αabN
b∂(t)
where the parameter αab =R
Ωabwµ
a wNdab ∈ [0, 1].
Relations between fab =ˆNab
d Nab Na∂ Nb
∂
˜T, eab = [µab µab
d µa∂ µ
b∂ ]T define
again a finite dimensional Dirac structure.
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Mixed finite element method
Objective : compute Nabdisp from µab
d such that the power behavior of the
dissipation element(dispersion law) is preserved : Nabdispµ
abd =
R b
aµd ∧ Ndisp as
well as the constitutive law of the dissipative element. Let us write theapproximate instantaneous power
PabR = −KabD(µab
d (t))2 (9)
with Kab depending on forms w only.
Nabdisp =
∂PabR
∂µabd
= −2KabDµabd (t) (10)
Objective : compute µab with respect to qab.
q(t, z), q and Nd lies in the same space =⇒ q(t, z) = qab(t)wNdab (z)
The energy of the element and its approximation on [a, b] are given by :
Gab =R t
0
`Rab
q(t, z)µ(t, z)´dt, g ab =
Z t
0
qab[bqab(t)| z µab
]wNdab (z)dt =
Z t
0
qabµabdt
dG ab
dt= qabµ
ab ⇒ dG ab
dqab= µab (11)
So the thermodynamic relations linking the gibbs density, concentration andchemical potentials are preserved by discretization.
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Mixed finite element method
Global schemePrincipe de la discrétisation structurée (Figure reprise deBaaiu2007)
!"#
$%&'%
(%&)(%
$*&)$%
(*&(+,-./()0((
$ 1 ( 1
$ 2 ( 2
!"#$%345
(%345
$*345
(*345$ 4 ( 4
6$ 46% ( 46%
!"#$%3%5
(%3%5
$*3%5
(*3%5$ % ( %
$ 7 ( 787&2
8%&2/ 8
806%&2/306%5 8
80&2/0 8
846%&2/346%5 8
84&2/4 8&1
!"#$%305
(%305
$*305
(*305$ 0/% ( 0/%
6$ 0 ( 0
9:
9:
$-;<=$3+,-)0;0,->(<,-;0?<$>@'<>25@,<;0$
@,<;0$$-;<=$3+,-)0;0,->(<,-;0?<$>@'<>15
PENG HUI (LAGEP-UCBL) Fevrier 20 / 48
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Mixed finite element method
Simulations
With 10 meshes
0 20 40 60 80 1000.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
1.1
Time: sec
Con
cent
ratio
n of
com
pone
nt Q
: mol
/Vol
ume
Finite difference with N=10
0 20 40 60 80 1000.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
1.1
Time: sec
Con
cent
ratio
n of
com
pone
nt Q
: mol
/Vol
ume
Structural method N=10
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Spectral properties
Analysis on simple cases :
Hyperbolic system (like undamped wave equation) :(f1f2
)=
(0 − ∂
∂z
− ∂∂z 0
)(e1
e2
)Parabolic system (like diffusion) :(
f1e2
)=
(0 − ∂
∂z
− ∂∂z 0
)(e1
f2
)
Question
What about spectral properties with respect to closure equations
Case 1 : f = ∂x∂t , e = δxH,H =
∫ ba
(αx2
1 + βx22
)dz
Case2 : f1 = ∂x1∂t , e1 = bx1, f2 = De2
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Spectral properties
Case 1 : The associated PDE (without closure equation) is givenby : (
x1
x1
)=
(0 − ∂
∂z
− ∂∂z 0
)(x1
x2
)+ BC
The spectrum associated to this operator is defined by :(ϕ1
1 ϕ12
ϕ21 ϕ2
2
)(λ1 00 λ2
)=
(0 − ∂
∂z
− ∂∂z 0
)(ϕ1
1 ϕ12
ϕ21 ϕ2
2
)+BC
Now let us consider the closure equation
(e1
e2
)=
(δx1Hδx1H
)with : H(x1, x2) =
∫ ba
(αx2
1 + βx22
)dz , then(
ψ11 ψ1
2
ψ21 ψ2
2
)(µ1 00 µ2
)=
(0 − ∂
∂z
− ∂∂z 0
)(2β 00 2α
)(ψ1
1 ψ12
ψ21 ψ3
2
)
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Spectral properties
Finally we obtain by identification :
=
0@ ϕ11 ϕ1
2
ϕ21 ϕ2
2
1A︷ ︸︸ ︷( √2β 0
0√
2α
)(ψ1
1 ψ12
ψ21 ψ2
2
)=
0@ λ1 00 λ2
1A︷ ︸︸ ︷(µ1
2√αβ
0
0 µ2
2√αβ
)
=
(0 − ∂
∂z
− ∂∂z 0
)=
0@ ϕ11 ϕ1
2
ϕ21 ϕ2
2
1A︷ ︸︸ ︷( √2β 0
0√
2α
)(ψ1
1 ψ12
ψ21 ψ2
2
)
Geometrical transformation :
Homothecy of the spectrum with factor 2√αβ
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Spectral properties
0.2 0.1 0 0.1 0.21
0.8
0.6
0.4
0.2
0
0.2
0.4
0.6
0.8
1
Real part
Imaginaty part
Spectrum of the canonical system
0.2 0.1 0 0.1 0.21
0.8
0.6
0.4
0.2
0
0.2
0.4
0.6
0.8
1
Real part
Imaginary part
Specrum of the hyperbolic system
Homothety
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Spectral properties
Case 2 : In a same way(f1e2
)=
(0 − ∂
∂z
− ∂∂z 0
)(e1
f2
)(12)
with the constraintf2 = De2 (13)
With the hypothesis of separation of variables (12) can berewritten with x = ψ0 expµt, e1 = bx :(
µ 00 b
)(ψ0∂ψ0∂z
)=
(0 − ∂
∂z
− ∂∂z 0
)(b 00 Db
)(ψ0∂ψ0∂z
)+ BC
(14)The eigenvalues and eigenfunctions satisfy :(
µ 00 1
)(ψ0
e2
)=
(0 ∂
∂z∂∂z 0
)(b 00 D
)(ψ0
e2
)
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Spectral properties
which leads finally to( √b 0
0√
Dµ
)(ψ0ee2
µ
)µ1/2√
bD=
(0 − ∂
∂z
− ∂∂z 0
)( √b 0
0√
Dµ
)(ψ0ee2
µ
)
which has to be identified to :(ϕ1
ϕ2
)λ =
(0 − ∂
∂z
− ∂∂z 0
)(ϕ1
ϕ2
)
Geometrical transformation :
Dilation and rotation of the spectrum
λk = |λk |e±i π2 −→ µk = Db|λk |2e±iπ
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Spectral properties
−5 −4 −3 −2 −1 0 1 2 3 4 5−0.8
−0.6
−0.4
−0.2
0
0.2
0.4
0.6
0.8
Real part
Imag
inar
y pa
rt
Canonical system associated toStokes−Dirac structure Parabolic system
Dilatation and rotation
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Spectral propertiesPrincipe de la discrétisation structurée (Figure reprise deBaaiu2007)
!"#
$%&'%
(%&)(%
$*&)$%
(*&(+,-./()0((
$ 1 ( 1
$ 2 ( 2
!"#$%345
(%345
$*345
(*345$ 4 ( 4
6$ 46% ( 46%
!"#$%3%5
(%3%5
$*3%5
(*3%5$ % ( %
$ 7 ( 787&2
8%&2/ 8
806%&2/306%5 8
80&2/0 8
846%&2/346%5 8
84&2/4 8&1
!"#$%305
(%305
$*305
(*305$ 0/% ( 0/%
6$ 0 ( 09:
9:
$-;<=$3+,-)0;0,->(<,-;0?<$>@'<>25@,<;0$
@,<;0$$-;<=$3+,-)0;0,->(<,-;0?<$>@'<>15
PENG HUI (LAGEP-UCBL) Fevrier 20 / 48
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Spectral properties
Let first consider one element of discretization.
Input : input mass flow
Output : output effort
This leads to :(f ab1
eab2
)=
(0 1
α− 1α 0
)(eab
1
f ab2
)+
(0 − 1
α− 1α 0
)(e∂b
f∂a
)(
f∂b
e∂a
)=
(0 1
α1α 0
)(eab
2
f ab2
)+
(0 α−1
α1−αα 0
)(e∂b
f∂a
)This element is interconnected with the next one using continuityof flow and equality of efforts : e∂
0b = e∂
1a and f∂
0b = −f∂
1a
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Spectral properties
In the case α = 1f 11
e12...
f N1
eN2
=
(0 M−MT 0
)
e11
f 12...
eN1
f N2
+ g
(e∂
1N
f∂00
)
(f∂
1N
e∂00
)= gT
e1
1
f 12...
eN1
f N2
with M =
1 −1 0 0 · · ·0 1 −1 0 · · ·0
. . .. . .
. . .. . .
0 · · · 0 1 −1
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Spectral properties
The spectrum associated with the Dirac structure :[0 dd 0
](eq
ep
)= λ
(eq
ep
)is computed using :
deq
dz2 = λ2eq, eq = Aeλz + Be−λz and
ep = Aeλz − Be−λz . With homogeneous boundary conditionsep(0) = eq(L) = 0 :[
1 −1eλL e−λL
](AB
)= 0
And
λk =2k + 1
2Lπi
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Spectral properties
The interconnection of n elements gives rise to the following statematrix :
A =
[0 M−MT 0
](15)
with
M =
1 −1 0 0 · · ·0 1 −1 0 · · ·0
. . .. . .
. . .. . .
0 · · · 0 1 −1
(16)
The eigenvalues of the finite dimensional system are given by :
λk =n
L
√2 cos
(2k + 1
2n + 1π
)− 2
if k n then :
λk '2k + 1
2Lπi
= first k eigenvalues of the infinite dimensional system
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
1 IntroductionContextOutline
2 Infinite dimensional linear 1-D caseA simple exampleConsidered class of systemsDirac structure and PHS on Hilbert SpaceBoundary control systemsStability, control feedback
3 Model reductionMixed finite element methodSpectral properties
4 Conclusion
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Conclusion :
For 1D case :
A powerful tool to highlight physical properties
Parametrization of all admissible BC for a large class ofsystems
Simple tools (matrix conditions) to check stability
Model reduction
Preserve energetic properties and the geometric structure
Link the solutions of very different systems (parabolic andhyperbolic)
Numerically efficient
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Future work :
Many extensions :
Generalization to 2D-3D linear systems.
Stabilization of a class of non linear systems.
Extensions to irreversible thermodynamic systems.
Control using infinite dimensional Port Hamiltonian Systems(Imersion/reduction + Casimir functions).
Model reduction of complex interconnected systems.
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Advertisement :
JDA-JNA course
ISEM lecture notes → book.
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
Y. Le Gorrec, H. Zwart and B. Maschke,
Dirac structures and Boundary Control Systems associated withSkew-Symmetric Differential Operators
SIAM Journal on Control and Optimization, Vol : 44 Issue 5, pages1864-1892, 2005.
J. A. Villegas, H. Zwart, Y. Le Gorrec, B. Maschke,
Exponential Stability of a Class of Boundary Control Systems.
IEEE Transactions on Automatic Control, Volume : 54, Issue : 1,pp : 142-147, January 2009.
H. Zwart, Y. Le Gorrec, B. Maschke and J. Villegas,
Well-posedness and regularity of hyperbolic boundary controlsystems on a one-dimensional spatial domain .
ESAIM Control, Optimisation and Calculus of Variations, Volume16, pp. 1077-1093, 2010.
Introduction Infinite dimensional linear 1-D case Model reduction Conclusion
A Baaiu, F. Couenne, L. Lefevre, Y. Le Gorrec and M. Tayakout,
Structure-preserving infinite dimensional model reduction.Application to Adsorption Processes.
Journal of Process Control - Volume 19, Issue 3, Pages 394-404,March 2009.
V. Dos Santos, Y. Le Gorrec, B. Maschke,
A Hamiltonian perspective to the stabilization of systems of twoconservation laws .
Networks and Heterogeneous Media - American Institute ofMathematical Sciences, Volume 4, Number 2, pp. 249-266, June2009.