Professional Documents
Culture Documents
Port–Hamiltonian Systems
– p. 1/4
CbI: Controllers as Multiport Cyclo–passive Systems
Plant (Σ) and controller (Σc ), with states x ∈ Rn , ζ ∈ Rm , are cyclo–passive, that is,
∃H(·) : Rn → R, Hc : Rm → R, such that
– p. 2/4
Adding the Energies via Power Preserving Interconnection
>
y(·) u + yc> uc = y(·)
>
v.
Problem:
Although Hc (ζ) is free, not clear how to affect x?
Energy functions “coupled" via the generation of invariant spaces.
Another, poorly understood, alternative is to make H c (x, ζ).
– p. 3/4
Invariant Function Method: Basic Idea
C(x, ζ) , F (x) − ζ
– p. 4/4
Preliminaries: Invariant Sets (Manifolds)
x(0) ∈ M ⇒ x(t) ∈ M, ∀t ≥ 0.
4
Fact Given D : Rn → Rn−m , D ∈ C 1 . The set M = {x ∈ Rn |D(x) = 0} is invariant iff
Ḋ|x∈M = 0.
4
Furthermore, the level sets (foliation) Mκ = {x ∈ Rn |D(x) = κ}, κ ∈ R are invariant iff
Ḋ = 0 (♥).
Remarks
For LTI systems, ẋ = Ax, the set spanned by the (real) eigenvectors of A is invariant.
(♥) is a PDE: ∇Df = 0.
The functions D(x) are called conserved quantities (dynamical invariants).
– p. 5/4
CbI for PH Systems with Natural Output y
Given a PH system,
ẋ = F (x)∇H(x) + g(x)u
Σ(u,y) ⇒ Ḣ ≤ u> y
y = g > (x)∇H(x),
– p. 6/4
Energy–Casimir Method
Proposition Assume, ∃ C : Rn → Rm such that the level sets Ωκ , {(x, ζ)|ζ = C(x) + κ}
are invariant, for all κ ∈ R. Then, for all Φ : Rm → R, the function
4
W (x, ζ) = H(x) + Hc (ζ) + Φ(C(x) − ζ).
satisfies
Ẇ ≤ v > y
That is, the system is cyclo–passive w.r.t. W (x, ζ), which (given C) can be shaped selecting
Hc and Φ.
Hence,
Φ̇ = ∇Φ(C˙ − ζ̇) = 0 ⇒ Ẇ = Ḣ + Ḣc .
– p. 7/4
Caveat: Achieving Asymptotic Stability
– p. 8/4
Casimir Functions for PH Systems
We are looking for C such that C˙ − ζ̇ = 0 for all H and Hc . Thus, we get the PDEs
h i F −g
(∇C)> −Im = 0.
g> 0
– p. 9/4
Conditions for CbI
Ẇ ≤ v > y,
4
W (x, ζ) = H(x) + Hc (ζ) + Φ(C(x) − ζ).
Dissipation Obstacle This design not applicable to systems with pervasive dissipation!
– p. 10/4
The Dissipation Obstacle
R∇x Φ(C(x) − ζ) = 0,
for all Φ : Rm → R. Consequently, energy cannot be shaped for coordinates that are
affected by physical damping.
Proof (CbI–PDE) ⇔
F ∇C = −g, g > ∇C = 0.
– p. 11/4
A Power Balance Perspective of the Dissipation Obstacle
RF −1 g = 0,
we see that, if the system admits a Casimir then the dissipated power at the
equilibrium should be zero, i.e., u>
∗ y∗ = 0. That is, we should be able to stabilize the
system extracting a finite amount of energy from the source.
– p. 12/4
Dissipation of PH Systems and Casimirs
Fact For PH systems Casimirs are necessarily “independent of the damped coordinates"
Post-multiplication by ∇C yields
Now, by transposition
∇C > [J(x) + R(x)]∇C = 0
Adding them up
∇C > R(x)∇C = 0 ⇔ R(x)∇C = 0
– p. 13/4
Can General Dynamic Controllers Overcome the Obstacle?
u = −yc , uc = y.
– p. 14/4
cont’
d
C=0
dt
– p. 15/4
Proposition
>
∂C ∂C
(x) (x)
J(x) = Jc (ζ)
∂x ∂x
∂C
R(x) (x) = 0 (DO)
∂x
Rc (ζ) = 0
>
∂C
(x) J(x) = gc (ζ)g > (x)
∂x
Remark Result holds for R(x) ≥ 0. How about pumping energy? (Jeltsema,
WLHMNLC’12).
– p. 16/4
Example without Pervasive Dissipation
1 ϕL
L
L
u C1 1 qC
C1
R2
Energy function
H(qC , φL ) = HE (qC ) + HM (φL ).
– p. 17/4
cont’d
1 2 1 2
W.l.o.g. assume linear elements: H(qC , φL ) = q
2C C
+ φ .
2L L
ˆ C )),
(g > ∇C = 0 ⇒ C = C(q (F ∇C = −q ⇒ C = qC + k).
1 2 1 2
W (qC , φL , ζ) = qC + φ + Hc (ζ) + Φ(qC − ζ),
2C 2L L
– p. 18/4
Example with Pervasive Dissipation
1
ϕL
L
L
u C1 R2 1 q
C1 C
−1
R2
1
Same energy function, but the dissipation has changed: J − R =
−1 0
– p. 19/4
Control by State–Modulated Interconnection with Σ (u,y)
admits a solution for some vector function C : R n → Rm . The PH system Σ(u,y) with the
PH controller Σc and the state–modulated power–preserving interconnection subsystem
u 0 −α(x) y v
ΣSM
I : = + ,
uc α> (x) 0 yc 0
4
W (x, ζ) = H(x) + Hc (ζ) + Φ(C(x) − ζ).
– p. 20/4
CbI with Alternative Port Variables: yPS
verifying
Fd + Fd> ≤ 0 (SY M ),
– p. 21/4
CbI with Σ(u,yPS ) Overcomes the Dissipation Obstacle
Proposition Assume (PO–PDE) admits a solution Fd verifying (SYM) and such that
Fd ∇C = −g (CbIPS − P DE)
for some vector function C : Rn → Rm . Consider the PH system Σ(u,yPS ) coupled with the
PH controller Σc through the power–preserving interconnection subsystem
u 0 −1 yPS v
ΣPS
I : = + .
uc 1 0 yc 0
where
4
WPS (x, ζ) = HPS (x) + Hc (ζ) + Φ(C(x) − ζ),
R
with HPS = (Fd−1 F ∇H)dx.
– p. 22/4
An Example of Stabilization via CbIPS
1 1
Ḣ = − (H 0 )2 + H 0 u.
R R A more physically sensible passive map
is voltage to current, i.e., u → ẋ,
One PH model
1 1
1 1
ẋ = −R H0 + u
ẋ = −R H0 + R
u Σ(u,y PS ) : R
Σ(u,y) : 1 1
1
y PS = −R H0 + u = ẋ,
y = R
H 0, R
– p. 23/4
cont’d
1 0 1
There is no C that satisfies the Casimir conditions: − R C = , 1 C0
R R
= 0, so this
problem can not be solved using CbI.
1
For any x? , the corresponding u? = (H 0 )? and y? = R
(H 0 )? yield the power
1
u? y ? = (H 0 )2? .
R
has an isolated minimum at a given equilibrium point (x ? , 0). For simplicity, take
ζ? = 0.
– p. 24/4
cont’d
Compute
Φ0? = −u?
(∇W PS )? = 0 ⇔ .
(H 0 )? = −u?
c
00 00
Φ > −H
00 00
00 −H Φ
Hc > 00 00 .
H +Φ
Fix a quadratic
1
Hc = (ζ − Cc u? )2 ,
2Cc
with Cc ∈ R+ .
β 2
For Φ we propose Φ(z) = 2
z + γz,
βx? + γ = −u? .
– p. 25/4
cont’d
00
β > −H?
00
1 H β
> − 00? ,
Cc H? + β
00
from where it is easy to see that if H? > 0, we can take β = 0 and the equilibrium
(x? , 0) will be stable, for all Cc > 0, with Lyapunov function W PS (x, ζ) − (W PS )? ,
where
1
W PS (x, ζ) = H(x) + (ζ − Cc u? )2 − u? (x − ζ).
2Cc
– p. 26/4
cont’d
PSfrag replacements
Physical realization
ζ̇ ẋ
v
+
1
Cc
ζ Cc + R +
– u C H 0 (x)
u? – –
– p. 27/4
Control by State–Modulated Interconnection with Σ (u,yPS )
Proposition Assume (PO–PDE) admits a solution Fd verifying (SYM) and such that
g ⊥ Fd ∇C = 0, SM
(CbIPS − P DE)
for some vector function C : Rn → Rm , where g ⊥ ∈ R(n−m)×n is a full rank left annihilator
of g, that is, g ⊥ g = 0 and rank g ⊥ = n − m. The PH system Σ(u,yPS ) with the PH controller
Σc and the state–modulated power–preserving interconnection subsystem
u 0 −α(x) yPS v
ΣSM
I : = + ,
uc α> (x) 0 yc 0
4
WPS (x, ζ) = HPS (x) + Hc (ζ) + Φ(C(x) − ζ).
– p. 28/4
Generating New Cyclo–Passive Port Variables
defines a cyclo–passive pair (u, yN ), with storage function H(x). That is,
Ḣ ≤ u> yN
– p. 29/4
Relationship with yPS
Proposition Assume F is full rank and set Fd = F > , which clearly verifies (PO–PDE) and
(SYM). Select
and
D + S = g > F −> (R − J )F −1 g = −g > F −> F F −1 g = −g > F −> g.
– p. 30/4
CbI With y(·) = yN
gd := F −> (g + 2T ),
and S : Rn → Rm×m as
1 >
S := (gd g + g > gd ).
2
Assume there exists T : Rn → Rn×m such that
rank (g + 2T ) = m,
and
R T
≥ 0, (DC)
T> S
hold.
– p. 31/4
Proposition
1 >
D := (gd g − g > gd ),
2
(g + 2T )⊥ F > ∇C = 0 (N EW − P DE).
– p. 32/4
Proof
Now, (NEW–PDE) and α are equivalent to gd α = ∇C. Replacing in the PDEs above yields
F > gd α = (g + 2T )α
α> gd> gα = α> (S + D)α.
Recalling gd the first equation is verified. The second identity also holds, using the definition
of S and D.
– p. 33/4
Summary of CbI
SM )
(Basic CbIPS ) and (Basic CbIPS
F ∇C = −g, g ⊥ F ∇C = 0.
SM )
(CbIPS ) and (CbIPS
Fd ∇C = −g, g ⊥ Fd ∇C = 0.
– p. 34/4
cont’d
Notation: A → B means that the set of solutions of the PDEs of B is strictly larger than the
PSfrag replacements
one of A, consequently the set of plants to which B is applicable is also strictly larger.
sm sm sm
CbI Basic CbIps CbIps
Question Where is CbINSM ?
– p. 35/4
Comparison of CbI and State–Feedback PBC: Applicability
F −g
(CbI) ∇C =
g> 0
g⊥ F
(EBC) ∇Ha = 0
g⊥ F g>
(CbI SM ) ∇C = 0
g> (Basic IDA)
(Basic CbIPS ) F ∇C = −g
g ⊥ F ∇Ha = 0
(CbIPS ) Fd ∇C = −g plus (PO–PDE)
(F ∇H = Fd ∇HPS )
SM )
(Basic CbIPS (PS)
⊥ g ⊥ Fd ∇Ha = 0
g F ∇C = 0
plus (PO–PDE)
(IDA)
SM )
(CbIPS
g ⊥ Fd ∇C = 0 g ⊥ Fd ∇Ha = g ⊥ (F − Fd )∇H
plus (PO-PDE).
– p. 36/4
ementsImplication Diagram
sm sm sm
CbI Basic CbIps CbIps
– p. 37/4
State–Feedback PBC and CbI: Connections
Dynamic CbI
Dynamic feedback control u = −yc + v = −∇ζ Hc (ζ) + v,
ζ controllers state with energy Hc (ζ) free,
Generate Casimir functions, C, that make Ω = {(x, ζ)|ζ = C(x)} invariant
⇒ For arbitrary Φ
Ḣ(x) + Ḣc (ζ) + Φ(C(x) − ζ) ≤ v > y
State–feedback PBCFind a static state feedback, û(x) that ensures the following.
CbI with regulated sources (EB–PBC):
IDA–PBC:
f (x) + g(x)û(x) = Fd (x)∇Hd (x).
Questions
Is there a connection between the two methods?
What happens if we restrict CbI to the invariant subspace Ω?
Is there an advantage of dynamic extension from minimum assignment viewpoint?
– p. 38/4
Restricting a Dynamic CbI Controller Yields an EB–PBC
d
Ḣ + Hc (C(x)) ≤ v > y.
dt
Furthermore,
d
Hc (C(x)) = −y > ûEB .
dt
4
Proof Define Ha (x) = Hc (C(x))
– p. 39/4
Restricting a CbIPS Controller Yields an IDA–PBC
Proposition Assume the conditions for CbIPS are satisfied. Then, for all Hc : Rm → R, the
state–feedback controller
ûIDA (x) = −∇C Hc (C(x)),
F ∇H + g ûIDA = Fd ∇Hd
4
is satisfied with Hd = HPS + Ha and Ha (x) = Hc (C(x)).
Proof
Conditions for CbIPS :
(PO–PDE) ⇔ F ∇H = Fd ∇HPS ,
(CbIPS –PDE) ⇔ Fd ∇C = −g.
Replacing in the matching equation yields
– p. 40/4
Stabilization via CbI and State–feedback PBC
We have concentrated our attention on the ability of the various PBCs to modify the
energy function, without particular concern to stabilization.
Stability will be ensured if a (desired) strict minimum is assigned to the total energy
function
Proposition In the single input case, the use of a dynamic extension does not provide any
additional freedom for minimum assignment to the corresponding static state–feedback
solutions.
Proof Define
4
W (x, ζ) = H(x) + Hc (ζ) + Φ(C(x) − ζ)
4
Hd (x) = H(x) + Hc (C(x)).
– p. 41/4
cont’d
Compute
∇H + Φ0 ∇C
∇W = , ∇Hd = ∇H + Hc0 ∇C.
Hc0 − Φ0
Now,
∇W? := ∇W (x? , ζ? ) = 0 ⇒ Φ0? = (Hc0 )? ⇒ ∇x W? = (∇Hd )? .
– p. 42/4
cont’d
Now,
>
I ∇C 2 I ∇C ∇2 H + Φ0 ∇2 C + Hc00 ∇C∇C > ∗
∇ W =
0 |∇C|2 0 |∇C|2 ∗ ∗
– p. 43/4
Current Research
1 > −1
S= [g (F + F −> )g] + T > F −1 g + g > F −> T.
2
The order of the dynamic extension is m. There are some advantages for increasing
their number.
Beyond simple nonlinear integrators?
CbI does not help for minimum assignment, but certainly has an impact on
performance and simplicity.
Can CbI be used—as an alternative to the current artificial “perturbation" or “delayed"
framework—to formulate the problem of control over networks?
Key question: Impact of the network topology on the ability to shape the energy, i.e., to
generate the Casimir functions.
– p. 44/4