optimal steady-state control - nrel · optimal steady-state control problem statement. given: 1. a...
TRANSCRIPT
![Page 1: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/1.jpg)
Optimal Steady-State Control (with Application to Secondary Frequency Control of Power Systems)
John W. Simpson-Porco
NREL Workshop on Opt./ Control
Golden, CO, USA
April 22, 2019
![Page 2: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/2.jpg)
Collaborators
Enrique Mallada Liam S. P. Lawrence
John’s Hopkins Univ. MASc 2019, Waterloo
Talk based on:
Lawrence, JWSP, Mallada: The optimal steady-state control problem (Arxiv preprint, revision pending . . . )
1 / 27
1
![Page 3: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/3.jpg)
Control Systems 101
Prototypical feedback control problem is tracking and disturbance rejection in the presence of model uncertainty
w
PlantController r e −
How is the reference r being determined?
u y
2 / 27
![Page 4: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/4.jpg)
Control Systems 101
Prototypical feedback control problem is tracking and disturbance rejection in the presence of model uncertainty
w
PlantController r e −
How is the reference r being determined?
u y
2 / 27
![Page 5: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/5.jpg)
Feedforward Optimization of Large-Scale Systems
r1 r2 rn
minimize z,r
f (z) + g(r)
subject to (z , r) ∈ C(model, pred.)
z1 wn
wmeas
zp
C1
P1 w1
C2
P2 w2z2
Cn
Pn zn
Physical Coupling P0
(i.e., Infrastructure Dynamics)
wp
3 / 27
![Page 6: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/6.jpg)
Feedback Optimization of Large-Scale Systems
C1
P1 w1z1
C2
P2 w2z2
Cn
Pn wnzn
Physical Coupling P0
(i.e., Infrastructure Dynamics)
wpzp
yp wmeas
r1 r2 rn
ξk+1 = f (ξk , model, pred., meas.)
rk = h(ξk , model, pred., meas.)
4 / 27
![Page 7: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/7.jpg)
Optimal Steady-State Control Problem Statement
Given: 1 a dynamic system model with
a class of external disturbances w(t) a model uncertainty specification (e.g., parametric)
2 a vector of outputs y ∈ Rp of system to be optimized 3 an optimization problem in y
Design, if possible, a feedback controller such that
1 closed-loop is (robustly) well-posed and internally stable 2 the regulated output tracks its optimal value
lim t→∞
y(t) − y ?(t) = 0 , ∀disturb, ∀ uncertainties
5 / 27
![Page 8: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/8.jpg)
Optimal Steady-State Control Problem Statement
Given: 1 a dynamic system model with
a class of external disturbances w(t) a model uncertainty specification (e.g., parametric)
2 a vector of outputs y ∈ Rp of system to be optimized 3 an optimization problem in y
Design, if possible, a feedback controller such that
1 closed-loop is (robustly) well-posed and internally stable 2 the regulated output tracks its optimal value
lim t→∞
y(t) − y ?(t) = 0 , ∀disturb, ∀ uncertainties
5 / 27
![Page 9: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/9.jpg)
Optimal Steady-State Control Problem Statement
Given: 1 a dynamic system model with
a class of external disturbances w(t) a model uncertainty specification (e.g., parametric)
2 a vector of outputs y ∈ Rp of system to be optimized 3 an optimization problem in y
Design, if possible, a feedback controller such that
1 closed-loop is (robustly) well-posed and internally stable 2 the regulated output tracks its optimal value
lim t→∞
y(t) − y ?(t) = 0 , ∀disturb, ∀ uncertainties
5 / 27
![Page 10: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/10.jpg)
Optimal Steady-State Control Problem Statement
Given: 1 a dynamic system model with
a class of external disturbances w(t) a model uncertainty specification (e.g., parametric)
2 a vector of outputs y ∈ Rp of system to be optimized 3 an optimization problem in y
Design, if possible, a feedback controller such that
1 closed-loop is (robustly) well-posed and internally stable 2 the regulated output tracks its optimal value
lim t→∞
y(t) − y ?(t) = 0 , ∀disturb, ∀ uncertainties
5 / 27
![Page 11: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/11.jpg)
Optimal Steady-State Control Problem Statement
Given: 1 a dynamic system model with
a class of external disturbances w(t) a model uncertainty specification (e.g., parametric)
2 a vector of outputs y ∈ Rp of system to be optimized 3 an optimization problem in y
Design, if possible, a feedback controller such that
1 closed-loop is (robustly) well-posed and internally stable 2 the regulated output tracks its optimal value
lim t→∞
y(t) − y ?(t) = 0 , ∀disturb, ∀ uncertainties
5 / 27
![Page 12: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/12.jpg)
Optimal Steady-State Control Problem Statement
Given: 1 a dynamic system model with
a class of external disturbances w(t) a model uncertainty specification (e.g., parametric)
2 a vector of outputs y ∈ Rp of system to be optimized 3 an optimization problem in y
Design, if possible, a feedback controller such that
1 closed-loop is (robustly) well-posed and internally stable 2 the regulated output tracks its optimal value
lim t→∞
y(t) − y ?(t) = 0 , ∀disturb, ∀ uncertainties
5 / 27
![Page 13: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/13.jpg)
Optimal Steady-State Control Problem Statement
Given: 1 a dynamic system model with
a class of external disturbances w(t) a model uncertainty specification (e.g., parametric)
2 a vector of outputs y ∈ Rp of system to be optimized 3 an optimization problem in y
Design, if possible, a feedback controller such that
1 closed-loop is (robustly) well-posed and internally stable 2 the regulated output tracks its optimal value
lim t→∞
y(t) − y ?(t) = 0 , ∀disturb, ∀ uncertainties
5 / 27
![Page 14: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/14.jpg)
LTI-Convex OSS Control: Setup Overview
1 Uncertain (possibly unstable) LTI dynamics
x = A(δ)x + B(δ)u + Bw (δ)w
ym = Cm(δ)x + Dm(δ) + Qm(δ)w
y = C (δ)x + D(δ)u + Q(δ)w
δ = parametric uncertainty, w = const. disturbances
ym = system measurements available for feedback
y = system states/inputs to be optimized
a steady-state convex optimization problem 2
y ?(w , δ) = argmin {f (y , w) : y ∈ C(w , δ)}y∈Rp
6 / 27
![Page 15: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/15.jpg)
LTI-Convex OSS Control: Setup Overview
1 Uncertain (possibly unstable) LTI dynamics
x = A(δ)x + B(δ)u + Bw (δ)w
ym = Cm(δ)x + Dm(δ) + Qm(δ)w
y = C (δ)x + D(δ)u + Q(δ)w
δ = parametric uncertainty, w = const. disturbances
ym = system measurements available for feedback
y = system states/inputs to be optimized
a steady-state convex optimization problem 2
y ?(w , δ) = argmin y∈Rp
{f (y , w) : y ∈ C(w , δ)}
6 / 27
![Page 16: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/16.jpg)
LTI-Convex OSS Control: Setup I
0 = A(δ)x + B(δ)u + Bw (δ)w Forced equilibria (x , u, y) satisfy
y = C (δ)x + D(δ)u + Q(δ)w
This defines an affine set of achievable steady-state outputs
Y (w , δ) = (offset vector) + V (δ)
Note: Due to
selection of variables y ∈ Rp to be optimized, and/or
structure of state-space matrices (A, B, C , D)
1
2
it may be that Y (w , δ) ⊂ Rp
constraint y ∈ Y (w , δ) cannot be ignored!!
7 / 27
![Page 17: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/17.jpg)
LTI-Convex OSS Control: Setup I
0 = A(δ)x + B(δ)u + Bw (δ)w Forced equilibria (x , u, y) satisfy
y = C (δ)x + D(δ)u + Q(δ)w
This defines an affine set of achievable steady-state outputs
Y (w , δ) = (offset vector) + V (δ)
Note: Due to
selection of variables y ∈ Rp to be optimized, and/or
structure of state-space matrices (A, B, C , D)
1
2
it may be that Y (w , δ) ⊂ Rp
constraint y ∈ Y (w , δ) cannot be ignored!!
7 / 27
![Page 18: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/18.jpg)
LTI-Convex OSS Control: Setup II Desired regulated output y?(w , δ) solution to
minimize y∈Rp
f0(y , w) (convex cost)
subject to y ∈ Y (w , δ) (equilibrium)
Hy = Lw (engineering equality)
Jy ≤ Mw (engineering inequality)
Equilibrium constraints ensure compatibility between the plant and the optimization problem
=⇒ guarantees a steady-state exists s.t. y = y?(w , δ).
We want to track optimal output y?(w , δ)
8 / 27
![Page 19: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/19.jpg)
LTI-Convex OSS Control: Setup II Desired regulated output y?(w , δ) solution to
minimize y∈Rp
f0(y , w) (convex cost)
subject to y ∈ Y (w , δ) (equilibrium)
Hy = Lw (engineering equality)
Jy ≤ Mw (engineering inequality)
Equilibrium constraints ensure compatibility between the plant and the optimization problem
=⇒ guarantees a steady-state exists s.t. y = y?(w , δ).
We want to track optimal output y?(w , δ)
8 / 27
![Page 20: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/20.jpg)
LTI-Convex OSS Control: Setup II Desired regulated output y?(w , δ) solution to
minimize y∈Rp
f0(y , w) (convex cost)
subject to y ∈ Y (w , δ) (equilibrium)
Hy = Lw (engineering equality)
Jy ≤ Mw (engineering inequality)
Equilibrium constraints ensure compatibility between the plant and the optimization problem
=⇒ guarantees a steady-state exists s.t. y = y?(w , δ).
We want to track optimal output y?(w , δ)
8 / 27
![Page 21: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/21.jpg)
LTI-Convex OSS Control: Setup II Desired regulated output y?(w , δ) solution to
minimize y∈Rp
f0(y , w) (convex cost)
subject to y ∈ Y (w , δ) (equilibrium)
Hy = Lw (engineering equality)
Jy ≤ Mw (engineering inequality)
Equilibrium constraints ensure compatibility between the plant and the optimization problem
=⇒ guarantees a steady-state exists s.t. y = y?(w , δ).
We want to track optimal output y?(w , δ)
8 / 27
![Page 22: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/22.jpg)
LTI-Convex OSS Control: Setup II Desired regulated output y?(w , δ) solution to
minimize y∈Rp
f0(y , w) (convex cost)
subject to y ∈ Y (w , δ) (equilibrium)
Hy = Lw (engineering equality)
Jy ≤ Mw (engineering inequality)
Equilibrium constraints ensure compatibility between the plant and the optimization problem
=⇒ guarantees a steady-state exists s.t. y = y?(w , δ).
We want to track optimal output y?(w , δ)
8 / 27
![Page 23: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/23.jpg)
Optimality Models for OSS Control
An optimality model filters the available measurements to robustly produce a proxy error � for the true tracking error e = y?(w , δ) − y
w
u ym �
ξ = OM state Plant Optimality Model
Steady-state requirement: if the plant and optimality model are both in equilibrium and � = 0, then y = y?(w , δ).
Driving � to zero (+ internal stability) drives y to y?
9 / 27
![Page 24: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/24.jpg)
Optimality Models for OSS Control
An optimality model filters the available measurements to robustly produce a proxy error � for the true tracking error e = y?(w , δ) − y
w
Plant ymu
Optimality Model �
ξ = OM state
Steady-state requirement: if the plant and optimality model are both in equilibrium and � = 0, then y = y?(w , δ).
Driving � to zero (+ internal stability) drives y to y?
9 / 27
![Page 25: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/25.jpg)
Optimality Models for OSS Control
An optimality model filters the available measurements to robustly produce a proxy error � for the true tracking error e = y?(w , δ) − y
w
Plant ymu
Optimality Model �
ξ = OM state
Steady-state requirement: if the plant and optimality model are both in equilibrium and � = 0, then y = y?(w , δ).
Driving � to zero (+ internal stability) drives y to y?
9 / 27
![Page 26: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/26.jpg)
Optimality Models for OSS Control
An optimality model filters the available measurements to robustly produce a proxy error � for the true tracking error e = y?(w , δ) − y
w
Plant ymu
Optimality Model �
ξ = OM state
Steady-state requirement: if the plant and optimality model are both in equilibrium and � = 0, then y = y?(w , δ).
Driving � to zero (+ internal stability) drives y to y?
9 / 27
![Page 27: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/27.jpg)
Big Picture for OSS Control Optimality model reduces OSS control to output regulation
Plant Optimality Model
Stabilizing Controller
Integral Control
ymu
w
η
� ξ
uom
Optimality Model: creates proxy error signal �
Integral Control: integrates �
Stabilizing Controller: stabilizes closed-loop system
10 / 27
![Page 28: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/28.jpg)
Big Picture for OSS Control Optimality model reduces OSS control to output regulation
Plant Optimality Model
Stabilizing Controller
Integral Control
ymu
w
η
� ξ
uom
Optimality Model: creates proxy error signal �
Integral Control: integrates �
Stabilizing Controller: stabilizes closed-loop system
10 / 27
![Page 29: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/29.jpg)
Big Picture for OSS Control Optimality model reduces OSS control to output regulation
Plant Optimality Model
Stabilizing Controller
Integral Control
ymu
w
η
� ξ
uom
Optimality Model: creates proxy error signal �
Integral Control: integrates �
Stabilizing Controller: stabilizes closed-loop system
10 / 27
![Page 30: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/30.jpg)
Big Picture for OSS Control Optimality model reduces OSS control to output regulation
Plant Optimality Model
Stabilizing Controller
Integral Control
ymu
w
η
� ξ
uom
Optimality Model: creates proxy error signal �
Integral Control: integrates �
Stabilizing Controller: stabilizes closed-loop system
10 / 27
![Page 31: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/31.jpg)
Optimality Model Details I
Can we implement an optimality model that is robust against δ?
minimize y ∈Rp
f0(y , w)
subject to y ∈ Y (w , δ) = (offset) + V (δ)
Hy = Lw
Jy ≤ Mw
Optimality condition:
? rf0(y , w) + JTν? ⊥ (V (δ) ∩ null(H))
possibly depends on uncertain parameter δ.
11 / 27
![Page 32: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/32.jpg)
Optimality Model Details I
Can we implement an optimality model that is robust against δ?
minimize y ∈Rp
f0(y , w)
subject to y ∈ Y (w , δ) = (offset) + V (δ)
Hy = Lw
Jy ≤ Mw
Optimality condition:
? rf0(y , w) + JTν? ⊥ (V (δ) ∩ null(H))
possibly depends on uncertain parameter δ.
11 / 27
![Page 33: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/33.jpg)
Optimality Model Details I
Can we implement an optimality model that is robust against δ?
minimize y ∈Rp
f0(y , w)
subject to y ∈ Y (w , δ) = (offset) + V (δ)
Hy = Lw
Jy ≤ Mw
Optimality condition:
? rf0(y , w) + JTν? ⊥ (V (δ) ∩ null(H))
possibly depends on uncertain parameter δ.
11 / 27
![Page 34: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/34.jpg)
Optimality Model Details II When can an optimality model encode the gradient KKT condition?
? rf0(y , w) + JTν? ⊥ (V (δ) ∩ null(H))
Robust Feasible Subspace Property
V (δ) ∩ null(H) is independent of δ
V (δ1) V (δ2)
null(H)
12 / 27
![Page 35: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/35.jpg)
Optimality Model Details II When can an optimality model encode the gradient KKT condition?
? rf0(y , w) + JTν? ⊥ (V (δ) ∩ null(H))
Robust Feasible Subspace Property
V (δ) ∩ null(H) is independent of δ
V (δ1) V (δ2)
null(H)
12 / 27
![Page 36: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/36.jpg)
Optimality Model Details II When can an optimality model encode the gradient KKT condition?
? rf0(y , w) + JTν? ⊥ (V (δ) ∩ null(H))
Robust Feasible Subspace Property
V (δ) ∩ null(H) is independent of δ
V (δ1) V (δ2)
null(H)
12 / 27
![Page 37: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/37.jpg)
Optimality Model Details III If Robust Feasible Subspace property holds, then
ν = ϕ(ν, Jy − Mw) � � Hy − Lw
� = T0 T(rf0(y , w) + JTν)
range(T0)
= V(δ) ∩ null(H)
(Design freedom!)
is an optimality model for the LTI-Convex OSS Control Problem.
Comments: 1 T0
Tz extracts component of z in subspace V(δ) ∩ null(H):
�2 = 0 ⇐⇒ rf0(y , w) + JTν ⊥ V (δ) ∩ null(H)
2 Flexibility: different equivalent formulations of optimization problem yield different optimality models
13 / 27
![Page 38: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/38.jpg)
Optimality Model Details III If Robust Feasible Subspace property holds, then
ν = ϕ(ν, Jy − Mw) � � Hy − Lw
� = T0 T(rf0(y , w) + JTν)
range(T0)
= V(δ) ∩ null(H)
(Design freedom!)
is an optimality model for the LTI-Convex OSS Control Problem.
Comments: 1 T0
Tz extracts component of z in subspace V(δ) ∩ null(H):
�2 = 0 ⇐⇒ rf0(y , w) + JTν ⊥ V (δ) ∩ null(H)
2 Flexibility: different equivalent formulations of optimization problem yield different optimality models
13 / 27
![Page 39: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/39.jpg)
Optimality Model Details III If Robust Feasible Subspace property holds, then
ν = ϕ(ν, Jy − Mw) � � Hy − Lw
� = T0 T(rf0(y , w) + JTν)
range(T0)
= V(δ) ∩ null(H)
(Design freedom!)
is an optimality model for the LTI-Convex OSS Control Problem.
Comments: 1 T0
Tz extracts component of z in subspace V(δ) ∩ null(H):
�2 = 0 ⇐⇒ rf0(y , w) + JTν ⊥ V (δ) ∩ null(H)
2 Flexibility: different equivalent formulations of optimization problem yield different optimality models
13 / 27
![Page 40: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/40.jpg)
What happens if RFS does not hold?
14 / 27
![Page 41: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/41.jpg)
Stabilizer Details I Can we actually stabilize this thing?
Goal: stabilize cascade of plant → optimality model → integrators 1
Plant Optimality Model
Integral Control
Stabilizing Controller
w
ym �
η
2
3
Can prove closed-loop stable =⇒ OSS control problem solved For QP OSS control, can prove cascade is stabilizable iff
plant stabilizable/detectable optimization problem has a unique solution engineering constraints not redundant with equilibrium constraints T0 has full column rank
15 / 27
![Page 42: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/42.jpg)
Stabilizer Details I Can we actually stabilize this thing?
Goal: stabilize cascade of plant → optimality model → integrators 1
Plant Optimality Model
Integral Control
Stabilizing Controller
w
ym �
η
2 Can prove closed-loop stable =⇒ OSS control problem solved 3 For QP OSS control, can prove cascade is stabilizable iff
plant stabilizable/detectable optimization problem has a unique solution engineering constraints not redundant with equilibrium constraints T0 has full column rank
15 / 27
![Page 43: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/43.jpg)
Stabilizer Details I Can we actually stabilize this thing?
Goal: stabilize cascade of plant → optimality model → integrators 1
Plant Optimality Model
Integral Control
Stabilizing Controller
w
ym �
η
2 Can prove closed-loop stable =⇒ OSS control problem solved 3 For QP OSS control, can prove cascade is stabilizable iff
plant stabilizable/detectable optimization problem has a unique solution engineering constraints not redundant with equilibrium constraints T0 has full column rank
15 / 27
![Page 44: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/44.jpg)
Stabilizer Details I Can we actually stabilize this thing?
Goal: stabilize cascade of plant → optimality model → integrators 1
Plant Optimality Model
Integral Control
Stabilizing Controller
w
ym �
η
2
3
Can prove closed-loop stable =⇒ OSS control problem solved For QP OSS control, can prove cascade is stabilizable iff
plant stabilizable/detectable optimization problem has a unique solution engineering constraints not redundant with equilibrium constraints T0 has full column rank
15 / 27
![Page 45: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/45.jpg)
Stabilizer Details I Can we actually stabilize this thing?
Goal: stabilize cascade of plant → optimality model → integrators 1
Plant Optimality Model
Integral Control
Stabilizing Controller
w
ym �
η
2
3
Can prove closed-loop stable =⇒ OSS control problem solved For QP OSS control, can prove cascade is stabilizable iff
plant stabilizable/detectable optimization problem has a unique solution engineering constraints not redundant with equilibrium constraints T0 has full column rank
15 / 27
![Page 46: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/46.jpg)
Stabilizer Details I Can we actually stabilize this thing?
Goal: stabilize cascade of plant → optimality model → integrators 1
Plant Optimality Model
Integral Control
Stabilizing Controller
w
ym �
η
2
3
Can prove closed-loop stable =⇒ OSS control problem solved For QP OSS control, can prove cascade is stabilizable iff
plant stabilizable/detectable optimization problem has a unique solution engineering constraints not redundant with equilibrium constraints T0 has full column rank
15 / 27
![Page 47: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/47.jpg)
Stabilizer Details I Can we actually stabilize this thing?
Goal: stabilize cascade of plant → optimality model → integrators 1
Plant Optimality Model
Integral Control
Stabilizing Controller
w
ym �
η
2
3
Can prove closed-loop stable =⇒ OSS control problem solved For QP OSS control, can prove cascade is stabilizable iff
plant stabilizable/detectable optimization problem has a unique solution engineering constraints not redundant with equilibrium constraints T0 has full column rank
15 / 27
![Page 48: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/48.jpg)
Stabilizer Details II Optimality model contains monotone nonlinearity rf0(y) . . .
Plant Optimality Model
Stabilizing Controller
Integral Control
ymu
w
η
� ξ
uom
In theory: full-order robustly stabilizing controller design
In practice: low-gain integral control u = −K η if open-loop stable, or any heuristic, e.g., linearize and do H2 design
Stabilizer design options: 1
2
Closed-loop stability analysis:
Robust stability (e.g., IQC-based) or time-scale separation
16 / 27
1
![Page 49: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/49.jpg)
Stabilizer Details II Optimality model contains monotone nonlinearity rf0(y) . . .
Plant Optimality Model
Stabilizing Controller
Integral Control
ymu
w
η
� ξ
uom
In theory: full-order robustly stabilizing controller design
In practice: low-gain integral control u = −K η if open-loop stable, or any heuristic, e.g., linearize and do H2 design
Stabilizer design options: 1
2
Closed-loop stability analysis:
Robust stability (e.g., IQC-based) or time-scale separation
16 / 27
1
![Page 50: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/50.jpg)
Stabilizer Details II Optimality model contains monotone nonlinearity rf0(y) . . .
Plant Optimality Model
Stabilizing Controller
Integral Control
ymu
w
η
� ξ
uom
In theory: full-order robustly stabilizing controller design
In practice: low-gain integral control u = −K η if open-loop stable, or any heuristic, e.g., linearize and do H2 design
Stabilizer design options: 1
2
Closed-loop stability analysis:
Robust stability (e.g., IQC-based) or time-scale separation
16 / 27
1
![Page 51: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/51.jpg)
Big Picture for OSS Control Optimality model reduces OSS control to output regulation
Plant Optimality Model
Stabilizing Controller
Integral Control
ymu
w
η
� ξ
uom
Optimality Model: creates proxy error signal �
Integral Control: integrates �
Stabilizing Controller: stabilizes closed-loop system
17 / 27
![Page 52: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/52.jpg)
Application: Inexact Reference Tracking
Want minimum error asymptotic tracking of a (possibly infeasible) reference signal subject to actuator limits, e.g.
minimize ym,u
kym − rk∞
subject to (ym, u) ∈ Y (w , δ)
u ≤ u ≤ u
If reference feasible, then exact tracking possible
Could promote sparsity in steady-state control actions
18 / 27
![Page 53: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/53.jpg)
1
Application: Frequency Control of Power Systems
Regulate frequency of an interconnected AC power system in presence of unknown disturbances (locally balance supply and demand)
Balancing Authority
Secondary Control
ΔPref
ΔPu
(Δω, Δptie)
2 Modern challenges / opportunities: variation due to RES =⇒ need fast control
inverter-based resources =⇒ fast actuation
new sensing, comm., comp. =⇒ new architectures
19 / 27
![Page 54: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/54.jpg)
1
Application: Frequency Control of Power Systems
Regulate frequency of an interconnected AC power system in presence of unknown disturbances (locally balance supply and demand)
Balancing Authority
Secondary Control
ΔPref
ΔPu
(Δω, Δptie)
2 Modern challenges / opportunities: variation due to RES =⇒ need fast control
inverter-based resources =⇒ fast actuation
new sensing, comm., comp. =⇒ new architectures
19 / 27
![Page 55: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/55.jpg)
2
Application: Frequency Control of Power Systems
1 Regulate frequency of an interconnected AC power system in presence of unknown disturbances (locally balance supply and demand)
Balancing Authority
Secondary Control
ΔPref
ΔPu
(Δω, Δptie)
Modern challenges / opportunities: variation due to RES =⇒ need fast control
inverter-based resources =⇒ fast actuation
new sensing, comm., comp. =⇒ new architectures
19 / 27
![Page 56: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/56.jpg)
Key insights into frequency control problem
1 For discussion, small-signal network of machines + turbine/gov
Δθi = Δωi ,
Mi Δ ωi = − Xn
j=1 Tij (Δθi − Δθj ) − Di Δωi + ΔPm,i + ΔPu,i
Ti Δ Pm,i = −ΔPm,i − R−1 d,i Δωi + ΔPref
i .
2 Model internally stable, DC gain analysis yields h iX1 ΔPrefΔωss = +ΔPu,iiβ
i
where β = P (Di + R−1) is frequency stiffness. i d,i
3 Lots of flexibility in choice of ΔPref !
20 / 27
![Page 57: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/57.jpg)
Optimal Allocation of Secondary Resources
ΔPu
ΔPref Power System
Δω
Allocate reserves ΔPref subject to frequency regulation i
Xn Ci (ΔPrefminimize )i
ΔPref ∈Rn i=1
subject to F Δω = 0
This OSS problem satisfies the robust feasible subspace property =⇒ can construct (several) different optimality models!
21 / 27
![Page 58: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/58.jpg)
Optimal Allocation of Secondary Resources
ΔPu
ΔPref Power System
Δω
Allocate reserves ΔPref subject to frequency regulation i
minimize ΔPref ∈Rn
Xn
i=1 Ci (ΔPref
i )
subject to F Δω = 0
This OSS problem satisfies the robust feasible subspace property =⇒ can construct (several) different optimality models!
21 / 27
![Page 59: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/59.jpg)
OSS Framework Recovers Recent Controllers 1 Distributed Averaging PI Control Xn
aij (rCi (Pref ) −rCj (P
ref�i = Δωi − ))i jj=1
ηi = �i
Pref = Stabilizeri (�i , ηi , ωi )i
Note: many architecture variations possible
AGC (stylized version)
Prefη = k · Δωcc , i = (rCi )−1(η)
2
3 Gather-and-broadacst (Dorfler & Grammatico)
nX Prefη =
1 Δωi , = (rCi )
−1(η)i n i=1
22 / 27
![Page 60: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/60.jpg)
3
OSS Framework Recovers Recent Controllers 1 Distributed Averaging PI Control Xn
aij (rCi (Pref ) −rCj (P
ref�i = Δωi − ))i jj=1
ηi = �i
Pref = Stabilizeri (�i , ηi , ωi )i
Note: many architecture variations possible
AGC (stylized version)
Prefη = k · Δωcc , i = (rCi )−1(η)
2
Gather-and-broadacst (Dorfler & Grammatico)
nX Prefη =
1 Δωi , = (rCi )
−1(η)i n i=1
22 / 27
![Page 61: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/61.jpg)
Conclusions
Robust feedback optimization of dynamic systems
Optimality model reduces OSS problem to output reg. problem
uom
Optimal Steady-State (OSS) Control framework
1
2
Plant Optimality Model
Stabilizing Controller
Integral Control
ymu
w
η
� ξ
Many pieces of theory wide open . . .
1
2
Decentralized, hierarchical, competitive, . . .
Performance improvement (e.g., feedforward, anti-windup)
23 / 27
![Page 62: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/62.jpg)
Conclusions
Robust feedback optimization of dynamic systems
Optimality model reduces OSS problem to output reg. problem
uom
Optimal Steady-State (OSS) Control framework
1
2
Plant Optimality Model
Stabilizing Controller
Integral Control
ymu
w
η
� ξ
Many pieces of theory wide open . . .
1 Decentralized, hierarchical, competitive, . . .
2 Performance improvement (e.g., feedforward, anti-windup)
23 / 27
![Page 63: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/63.jpg)
Details in paper on arXiv
https://arxiv.org/abs/1810.12892
Enrique Mallada Liam S. P. Lawrence
John’s Hopkins Univ. University of Waterloo
24 / 27
![Page 65: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/65.jpg)
appendix
![Page 66: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/66.jpg)
Feedforward vs. Feedback Optimization
Property Feedforward Feedback
Setpoint Quality ≈ Optimal ≈ Optimal
High-Fidelity Model Crucial Not crucial
Robustness No Yes
Feedback Design/Analysis Unchanged More difficult
Computational Effort Moderate ???
MPC: high computational effort, difficult analysis ⇒ Alternatives?
Compared to MPC, if we give a bit on trajectory optimality, can we can gain a lot on ease of design, analysis, and implementation?
Here is a first cut of such an approach.
![Page 67: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/67.jpg)
Feedforward vs. Feedback Optimization
Property Feedforward Feedback
Setpoint Quality ≈ Optimal ≈ Optimal
High-Fidelity Model Crucial Not crucial
Robustness No Yes
Feedback Design/Analysis Unchanged More difficult
Computational Effort Moderate ???
MPC: high computational effort, difficult analysis ⇒ Alternatives?
Compared to MPC, if we give a bit on trajectory optimality, can we can gain a lot on ease of design, analysis, and implementation?
Here is a first cut of such an approach.
![Page 68: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/68.jpg)
Feedforward vs. Feedback Optimization
Property Feedforward Feedback
Setpoint Quality ≈ Optimal ≈ Optimal
High-Fidelity Model Crucial Not crucial
Robustness No Yes
Feedback Design/Analysis Unchanged More difficult
Computational Effort Moderate ???
MPC: high computational effort, difficult analysis ⇒ Alternatives?
Compared to MPC, if we give a bit on trajectory optimality, can we can gain a lot on ease of design, analysis, and implementation?
Here is a first cut of such an approach.
![Page 69: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/69.jpg)
Feedforward vs. Feedback Optimization
Property Feedforward Feedback
Setpoint Quality ≈ Optimal ≈ Optimal
High-Fidelity Model Crucial Not crucial
Robustness No Yes
Feedback Design/Analysis Unchanged More difficult
Computational Effort Moderate ???
MPC: high computational effort, difficult analysis ⇒ Alternatives?
Compared to MPC, if we give a bit on trajectory optimality, can we can gain a lot on ease of design, analysis, and implementation?
Here is a first cut of such an approach.
![Page 70: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/70.jpg)
Feedforward vs. Feedback Optimization
Property Feedforward Feedback
Setpoint Quality ≈ Optimal ≈ Optimal
High-Fidelity Model Crucial Not crucial
Robustness No Yes
Feedback Design/Analysis Unchanged More difficult
Computational Effort Moderate ???
MPC: high computational effort, difficult analysis ⇒ Alternatives?
Compared to MPC, if we give a bit on trajectory optimality, can we can gain a lot on ease of design, analysis, and implementation?
Here is a first cut of such an approach.
![Page 71: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/71.jpg)
Feedforward vs. Feedback Optimization
Property Feedforward Feedback
Setpoint Quality ≈ Optimal ≈ Optimal
High-Fidelity Model Crucial Not crucial
Robustness No Yes
Feedback Design/Analysis Unchanged More difficult
Computational Effort Moderate ???
MPC: high computational effort, difficult analysis ⇒ Alternatives?
Compared to MPC, if we give a bit on trajectory optimality, can we can gain a lot on ease of design, analysis, and implementation?
Here is a first cut of such an approach.
![Page 72: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/72.jpg)
Feedforward vs. Feedback Optimization
Property Feedforward Feedback
Setpoint Quality ≈ Optimal ≈ Optimal
High-Fidelity Model Crucial Not crucial
Robustness No Yes
Feedback Design/Analysis Unchanged More difficult
Computational Effort Moderate ???
MPC: high computational effort, difficult analysis ⇒ Alternatives?
Compared to MPC, if we give a bit on trajectory optimality, can we can gain a lot on ease of design, analysis, and implementation?
Here is a first cut of such an approach.
![Page 73: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/73.jpg)
Is OSS Control just a standard tracking problem?
w
PlantController y?(w , δ) e
−
We want y to track y?(w , δ), but two problems: 1
2
?unmeasured components of w change y
y? depends on uncertainty δ (relevant if Y ⊂ Rp)
Standard tracking approach infeasible for quickly varying w(t), or large uncertainties δ, or particular
choices of regulated outputs
u y
![Page 74: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/74.jpg)
Is OSS Control just a standard tracking problem?
w
PlantController y?(w , δ) e
−
We want y to track y?(w , δ), but two problems: 1 unmeasured components of w change y?
2 y? depends on uncertainty δ (relevant if Y ⊂ Rp)
Standard tracking approach infeasible for quickly varying w(t), or large uncertainties δ, or particular
choices of regulated outputs
u y
![Page 75: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/75.jpg)
Is OSS Control just a standard tracking problem?
w
PlantController y?(w , δ) e
−
We want y to track y?(w , δ), but two problems: 1 unmeasured components of w change y?
2 y? depends on uncertainty δ (relevant if Y ⊂ Rp)
Standard tracking approach infeasible for quickly varying w(t), or large uncertainties δ, or particular
choices of regulated outputs
u y
![Page 76: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/76.jpg)
Towards an internal model principle . . .
� � Hy − Lw
� = TT0 rf0(y , w) range(T0) = V(δ) ∩ null(H)
uom
Plant Optimality Model
Stabilizing Controller
Integral Control
ymu
w
η
� ξ
Interpretation: Exact robust asymptotic optimization achieved if loop incorporates a model of the optimal set of
the optimization problem
![Page 77: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/77.jpg)
Slide on EOA Approach . . .
![Page 78: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/78.jpg)
Example 1: Necessity of Equilibrium Constraints Consider the OSS control problem:
Dynamics: 1
� x1
x2
�
=
� −1 0 1 −1
� � x1
x2
�
+
� 1 −1
�
u +
� 1 1
�
w
y =
� x1
u
�
2 Optimization problem:
minimize y∈R2
g(y) := 1 2y 2 1 +
1 2y 2 2
What happens if we omit the equilibrium constraints?
η = rf0(y)
u = −K η
![Page 79: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/79.jpg)
Example 1: Necessity of Equilibrium Constraints (cont.)
![Page 80: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/80.jpg)
Example 2: Necessity of Robust Feasible Subspace Consider the OSS control problem:
Dynamics: 1
� x1
x2
�
=
� −1 − δ 0 1 + δ −1
� � x1
x2
�
+
� 1 −1
�
u +
� 1 1
�
w
y =
� x1
u
�
2 Optimization problem:
minimize y∈R2
1 2y 2 1 +
1 2y 2 2
subject to y ∈ Y (w , δ) = y(w , δ) + V (δ)
�� �� 1
We can show V (δ) = span ⇒ V (δ) dependent on δ. δ
![Page 81: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/81.jpg)
Example 2: Necessity of Robust Feasible Subspace (cont.)
We apply our scheme anyway supposing δ = 0
Optimality model + integral control yields. . .
If δ = 0.5 in the true plant If δ = 0 in the true plant
⇒ achieve sub-optimal cost of ⇒ achieve optimal cost of 0.1538. 0.1599.
![Page 82: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/82.jpg)
1
Robust Output Subspace Optimality Model
If furthermore V (δ) itself is independent of δ, then
µ = Hy − Lw
ν = max(ν + Jy − Mw , 0) − ν
� = R0 T(rf0(y , w) + HT µ + JTν) (Design freedom!)
range R0 = V(δ)
is also an optimality model for the LTI-Convex OSS Control Problem.
Can take R0 = I if V (δ) = Rp, which holds if � � A C
B D
has full row rank ⇐⇒ No transmission zeros
at s = 0
Again, different equivalent formulations of optimization problem give different optimality models
2
![Page 83: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/83.jpg)
1
Robust Output Subspace Optimality Model
If furthermore V (δ) itself is independent of δ, then
µ = Hy − Lw
ν = max(ν + Jy − Mw , 0) − ν
� = R0 T(rf0(y , w) + HT µ + JTν) (Design freedom!)
range R0 = V(δ)
is also an optimality model for the LTI-Convex OSS Control Problem.
Can take R0 = I if V (δ) = Rp, which holds if � � A C
B D
has full row rank ⇐⇒ No transmission zeros
at s = 0
Again, different equivalent formulations of optimization problem give different optimality models
2
![Page 84: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/84.jpg)
1
Robust Output Subspace Optimality Model
If furthermore V (δ) itself is independent of δ, then
µ = Hy − Lw
ν = max(ν + Jy − Mw , 0) − ν
� = R0 T(rf0(y , w) + HT µ + JTν) (Design freedom!)
range R0 = V(δ)
is also an optimality model for the LTI-Convex OSS Control Problem.
Can take R0 = I if V (δ) = Rp, which holds if � � A C
B D
has full row rank ⇐⇒ No transmission zeros
at s = 0
Again, different equivalent formulations of optimization problem give different optimality models
2
![Page 85: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/85.jpg)
OSS Control in the Literature
The OSS controller architecture found throughout the literature on real-time optimization.
Problem [Nelson and Mallada ’18]
Design a feedback controller to drive the system
x(t) = Ax(t) + B(u(t) + w)
ym(t) = Cx(t) + D(u(t) + w)
to the solution of the optimization problem
minimize x∈Rn
f (x) .
![Page 86: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/86.jpg)
Optimality Model
OSS Control in the Literature (cont.)
u + w , ym Observer
x −rf0 �
PI u
Controller Design The optimality model is an observer with gradient output
x = (A − LC )x + (B − LD)(u + w) + Lym
� = −rf0(x) .
A PI controller serves as internal model and stabilizer
eI = � , u = KI eI + Kp� .
![Page 87: Optimal Steady-State Control - NREL · Optimal Steady-State Control Problem Statement. Given: 1. a dynamic system model with. a class of external disturbances w(t) a model uncertainty](https://reader033.vdocument.in/reader033/viewer/2022060414/5f12a6309efdc23f9b18c3de/html5/thumbnails/87.jpg)
OSS Control in the Literature (cont.)
Observer −rf0 PI u + w , ym x � u
Optimality Model
Controller Design The optimality model is an observer with gradient output
˙x = (A − LC )x + (B − LD)(u + w) + Lym
� = −rf0(x) .
A PI controller serves as internal model and stabilizer
eI = � , u = KI eI + Kp� .