Reduced Basis Method for Parametric

Transcription

1 1/24 Reduced Basis Method for Parametric H 2 -Optimal Control Problems NUMOC2017 Andreas Schmidt, Bernard Haasdonk Institute for Applied Analysis and Numerical Simulation, University of Stuttgart andreas.schmidt@mathematik.uni-stuttgart.de 22 nd June, 2017

2 Outline 1 Introduction 2 Reduced Basis Approximation 3 Numerical Results 4 Conclusion and Outlook 2/24

3 Introduction 3/24

4 Basic Overview A general control problem Goal: Set up a general framework for realistic control problems. Noise w 1 (t) Control u(t) System x(t) State z(t) Performance y(t) Measurement Noise w 2 (t) In a more mathematical way: LTI System x(t) = Ax(t) +B u u(t) +B w w 1 (t) y(t) = C y x(t) +D yw w 2 (t) z(t) = C z x(t) +D zu u(t) x(0) = x 0. State equation Measurements Performance output 4/24

5 Basic Overview A general control problem Goal: Set up a general framework for realistic control problems. Noise w 1 (t) Control u(t) System x(t) State z(t) Performance y(t) Measurement Noise w 2 (t) In a more mathematical way: LTI System x(t) = Ax(t) +B u u(t) +B w w 1 (t) y(t) = C y x(t) +D yw w 2 (t) z(t) = C z x(t) +D zu u(t) x(0) = x 0. State equation Measurements Performance output 4/24

6 Basic Overview A general control problem Goal: Set up a general framework for realistic control problems. Noise w 1 (t) Control u(t) System x(t) State z(t) Performance y(t) Measurement Noise w 2 (t) In a more mathematical way: LTI System x(t) = Ax(t) +B u u(t) +B w w 1 (t) y(t) = C y x(t) +D yw w 2 (t) z(t) = C z x(t) +D zu u(t) x(0) = x 0. State equation Measurements Performance output 4/24

7 Basic Overview A general control problem Goal: Set up a general framework for realistic control problems. Noise w 1 (t) Control u(t) System x(t) State z(t) Performance y(t) Measurement Noise w 2 (t) In a more mathematical way: LTI System x(t) = A(µ)x(t) +B u (µ)u(t) +B w (µ)w 1 (t) y(t) = C y (µ)x(t) +D yw (µ)w 2 (t) z(t) = C z (µ)x(t) +D zu (µ)u(t) x(0) = x 0 (µ). State equation Measurements Performance output 4/24

8 The Control Problem What should the controller do? w(t) System z(t) u(t) y(t) Controller Control objective Try to find a controller that maps the noisy measurements y(t) to control signals u(t) such that certain goals are achieved: System is stable Effect of noise w(t) on output z(t) is minimized in H 2 norm 1 G 2 = 2π G(iω) 2 F dω, where G(s) = C (si A ) 1 B 5/24

9 The Control Problem What should the controller do? w(t) u(t) Control objective System Controller y(t) z(t) Controller: x K = A K x K +B K y u = C K x K Closed-loop LTI: h = A h + Bw z = C h Try to find a controller that maps the noisy measurements y(t) to control signals u(t) such that certain goals are achieved: System is stable Effect of noise w(t) on output z(t) is minimized in H 2 norm 1 G 2 = 2π G(iω) 2 F dω, where G(s) = C (si A ) 1 B 5/24

10 The Control Problem What should the controller do? w(t) u(t) Control objective System Controller y(t) z(t) Controller: x K = A K x K +B K y u = C K x K Closed-loop LTI: h = A h + Bw z = C h Try to find a controller that maps the noisy measurements y(t) to control signals u(t) such that certain goals are achieved: System is stable Effect of noise w(t) on output z(t) is minimized in H 2 norm 1 G 2 = 2π G(iω) 2 F dω, where G(s) = C (si A ) 1 B 5/24

11 The Control Problem What should the controller do? w(t) u(t) Control objective System Controller y(t) z(t) Controller: x K = A K x K +B K y u = C K x K Closed-loop LTI: h = A h + Bw z = C h Try to find a controller that maps the noisy measurements y(t) to control signals u(t) such that certain goals are achieved: System is stable Effect of noise w(t) on output z(t) is minimized in H 2 norm 1 G 2 = 2π G(iω) 2 F dω, where G(s) = C (si A ) 1 B 5/24

12 The Control Problem What should the controller do? w(t) u(t) Control objective System Controller y(t) z(t) Controller: x K = A K x K +B K y u = C K x K Closed-loop LTI: h = A h + Bw z = C h Try to find a controller that maps the noisy measurements y(t) to control signals u(t) such that certain goals are achieved: System is stable Effect of noise w(t) on output z(t) is minimized in H 2 norm 1 G 2 = 2π G(iω) 2 F dω, where G(s) = C (si A ) 1 B 5/24

13 The Control Problem The solution Theorem 1 ( [Zhou, Doyle, Glover 1996]) Let P,Q R n n be the unique stabilizing solutions to the algebraic Riccati equations (AREs) A T P +PA PB u (D T zud zu ) 1 B T u P +C T z C z = 0 (1) AQ +QA T QCy T (D yw Dyw) T 1 C y Q +B w B T w = 0. (2) Define the control gain K := (D T zud zu ) 1 B T u P and the observer gain L := QC T y (D yw D T yw) 1. Then the H 2 optimal controller is given by u(t) := Kx K (t), (3) where x K (t) R n is the solution to the observer-equation: x K (t) = (A BK)x K (t) +L[y(t) C y x K (t)]. (4) K. Zhou, J. Doyle and K. Glover. Robust and Optimal Control /24

14 The Control Problem The solution Theorem 1 ( [Zhou, Doyle, Glover 1996]) Let P,Q R n n be the unique stabilizing solutions to the algebraic Riccati equations (AREs) A T P +PA PB u (D T zud zu ) 1 B T u P +C T z C z = 0 (1) AQ +QA T QCy T (D yw Dyw) T 1 C y Q +B w B T w = 0. (2) Define the control gain K := (D T zud zu ) 1 B T u P and the observer gain L := QC T y (D yw D T yw) 1. Then the H 2 optimal controller is given by u(t) := Kx K (t), (3) where x K (t) R n is the solution to the observer-equation: x K (t) = (A BK)x K (t) +L[y(t) C y x K (t)]. (4) K. Zhou, J. Doyle and K. Glover. Robust and Optimal Control /24

15 The Control Problem The solution Theorem 1 ( [Zhou, Doyle, Glover 1996]) Let P,Q R n n be the unique stabilizing solutions to the algebraic Riccati equations (AREs) A T P +PA PB u (D T zud zu ) 1 B T u P +C T z C z = 0 (1) AQ +QA T QCy T (D yw Dyw) T 1 C y Q +B w B T w = 0. (2) Define the control gain K := (D T zud zu ) 1 B T u P and the observer gain L := QC T y (D yw D T yw) 1. Then the H 2 optimal controller is given by u(t) := Kx K (t), (3) where x K (t) R n is the solution to the observer-equation: x K (t) = (A BK)x K (t) +L[y(t) C y x K (t)]. (4) K. Zhou, J. Doyle and K. Glover. Robust and Optimal Control /24

16 Reduced Basis Approximation 7/24

17 Reduced Basis Method in a Nutshell Calculation is split in offline and online phase Online efficiency through parameter separability Basis generation (usually) by Greedy procedure Many applications: Elliptic/parabolic PDEs, variational inequalities, dyn. systems, Kalman filter,... A. Patera and G. Rozza. Reduced Basis Approximation and a Posteriori Error Estimation for Parametrized Partial Differential Equations. To appear in (tentative) MIT Pappalardo Graduate Monographs in Mechanical Engineering B. Haasdonk. Reduced Basis Methods for Parametrized PDEs A Tutorial Introduction for Stationary and Instationary Problems. SimTech Preprint. Chapter to appear in P. Benner, A. Cohen, M. Ohlberger and K. Willcox (eds.): "Model Reduction and Approximation: Theory and Algorithms", SIAM, Philadelphia, IANS, University of Stuttgart, Germany, /24

18 Reduced Basis Method in a Nutshell Calculation is split in offline and online phase Online efficiency through parameter separability Basis generation (usually) by Greedy procedure Many applications: Elliptic/parabolic PDEs, variational inequalities, dyn. systems, Kalman filter,... A. Patera and G. Rozza. Reduced Basis Approximation and a Posteriori Error Estimation for Parametrized Partial Differential Equations. To appear in (tentative) MIT Pappalardo Graduate Monographs in Mechanical Engineering B. Haasdonk. Reduced Basis Methods for Parametrized PDEs A Tutorial Introduction for Stationary and Instationary Problems. SimTech Preprint. Chapter to appear in P. Benner, A. Cohen, M. Ohlberger and K. Willcox (eds.): "Model Reduction and Approximation: Theory and Algorithms", SIAM, Philadelphia, IANS, University of Stuttgart, Germany, /24

19 Reduced Basis Method in a Nutshell Calculation is split in offline and online phase Online efficiency through parameter separability Basis generation (usually) by Greedy procedure Many applications: Elliptic/parabolic PDEs, variational inequalities, dyn. systems, Kalman filter,... A. Patera and G. Rozza. Reduced Basis Approximation and a Posteriori Error Estimation for Parametrized Partial Differential Equations. To appear in (tentative) MIT Pappalardo Graduate Monographs in Mechanical Engineering B. Haasdonk. Reduced Basis Methods for Parametrized PDEs A Tutorial Introduction for Stationary and Instationary Problems. SimTech Preprint. Chapter to appear in P. Benner, A. Cohen, M. Ohlberger and K. Willcox (eds.): "Model Reduction and Approximation: Theory and Algorithms", SIAM, Philadelphia, IANS, University of Stuttgart, Germany, /24

20 Reduced Basis Method in a Nutshell Calculation is split in offline and online phase Online efficiency through parameter separability Basis generation (usually) by Greedy procedure Many applications: Elliptic/parabolic PDEs, variational inequalities, dyn. systems, Kalman filter,... A. Patera and G. Rozza. Reduced Basis Approximation and a Posteriori Error Estimation for Parametrized Partial Differential Equations. To appear in (tentative) MIT Pappalardo Graduate Monographs in Mechanical Engineering B. Haasdonk. Reduced Basis Methods for Parametrized PDEs A Tutorial Introduction for Stationary and Instationary Problems. SimTech Preprint. Chapter to appear in P. Benner, A. Cohen, M. Ohlberger and K. Willcox (eds.): "Model Reduction and Approximation: Theory and Algorithms", SIAM, Philadelphia, IANS, University of Stuttgart, Germany, /24

21 Reduction of the ARE How to do that? Recall: The ARE is a nonlinear matrix-valued equation for X(µ) R n n : A(µ) T X(µ) +X(µ)A(µ) X(µ)B(µ)R(µ) 1 B(µ) T X(µ) +C(µ) T C(µ) = 0. SVD/Eigenvalue decomposition X(µ) = V ΣV T σi Index i Projection leads to small N n dimensional ARE Idea: [S., Haasdonk 2017] Approximate X(µ) X(µ) := V X X N (µ)v T X A N (µ) T X N (µ)+x N (µ)a N (µ) X N (µ)b N (µ)r(µ) 1 B(µ) T X N (µ)+c N (µ) T C N (µ) = 0. A. Schmidt and B. Haasdonk. Reduced basis approximation of large scale parametric algebraic Riccati equations. ESAIM: Control, Optimisation and Calculus of Variations (Feb. 2017). V. Simoncini. Analysis of the Rational Krylov Subspace Projection Method for Large-Scale Algebraic Riccati Equations. SIAM Journal on Matrix Analysis and Applications (Jan. 2016). 9/24

22 Reduction of the ARE How to do that? Recall: The ARE is a nonlinear matrix-valued equation for X(µ) R n n : A(µ) T X(µ) +X(µ)A(µ) X(µ)B(µ)R(µ) 1 B(µ) T X(µ) +C(µ) T C(µ) = 0. SVD/Eigenvalue decomposition X(µ) = V ΣV T σi Index i Projection leads to small N n dimensional ARE Idea: [S., Haasdonk 2017] Approximate X(µ) X(µ) := V X X N (µ)v T X A N (µ) T X N (µ)+x N (µ)a N (µ) X N (µ)b N (µ)r(µ) 1 B(µ) T X N (µ)+c N (µ) T C N (µ) = 0. A. Schmidt and B. Haasdonk. Reduced basis approximation of large scale parametric algebraic Riccati equations. ESAIM: Control, Optimisation and Calculus of Variations (Feb. 2017). V. Simoncini. Analysis of the Rational Krylov Subspace Projection Method for Large-Scale Algebraic Riccati Equations. SIAM Journal on Matrix Analysis and Applications (Jan. 2016). 9/24

23 Reduction of the ARE How to do that? Recall: The ARE is a nonlinear matrix-valued equation for X(µ) R n n : A(µ) T X(µ) +X(µ)A(µ) X(µ)B(µ)R(µ) 1 B(µ) T X(µ) +C(µ) T C(µ) = 0. SVD/Eigenvalue decomposition X(µ) = V ΣV T Z(µ)Z(µ) T σi Index i Projection leads to small N n dimensional ARE Idea: [S., Haasdonk 2017] Approximate X(µ) X(µ) := V X X N (µ)v T X A N (µ) T X N (µ)+x N (µ)a N (µ) X N (µ)b N (µ)r(µ) 1 B(µ) T X N (µ)+c N (µ) T C N (µ) = 0. A. Schmidt and B. Haasdonk. Reduced basis approximation of large scale parametric algebraic Riccati equations. ESAIM: Control, Optimisation and Calculus of Variations (Feb. 2017). V. Simoncini. Analysis of the Rational Krylov Subspace Projection Method for Large-Scale Algebraic Riccati Equations. SIAM Journal on Matrix Analysis and Applications (Jan. 2016). 9/24

24 Reduction of the ARE How to do that? Recall: The ARE is a nonlinear matrix-valued equation for X(µ) R n n : A(µ) T X(µ) +X(µ)A(µ) X(µ)B(µ)R(µ) 1 B(µ) T X(µ) +C(µ) T C(µ) = 0. SVD/Eigenvalue decomposition X(µ) = V ΣV T Z(µ)Z(µ) T σi Index i Projection leads to small N n dimensional ARE Idea: [S., Haasdonk 2017] Approximate X(µ) X(µ) := V X X N (µ)v T X A N (µ) T X N (µ)+x N (µ)a N (µ) X N (µ)b N (µ)r(µ) 1 B(µ) T X N (µ)+c N (µ) T C N (µ) = 0. A. Schmidt and B. Haasdonk. Reduced basis approximation of large scale parametric algebraic Riccati equations. ESAIM: Control, Optimisation and Calculus of Variations (Feb. 2017). V. Simoncini. Analysis of the Rational Krylov Subspace Projection Method for Large-Scale Algebraic Riccati Equations. SIAM Journal on Matrix Analysis and Applications (Jan. 2016). 9/24

25 Reduction of the ARE How to do that? Recall: The ARE is a nonlinear matrix-valued equation for X(µ) R n n : A(µ) T X(µ) +X(µ)A(µ) X(µ)B(µ)R(µ) 1 B(µ) T X(µ) +C(µ) T C(µ) = 0. SVD/Eigenvalue decomposition X(µ) = V ΣV T Z(µ)Z(µ) T σi Index i Projection leads to small N n dimensional ARE Idea: [S., Haasdonk 2017] Approximate X(µ) X(µ) := V X X N (µ)v T X A N (µ) T X N (µ)+x N (µ)a N (µ) X N (µ)b N (µ)r(µ) 1 B(µ) T X N (µ)+c N (µ) T C N (µ) = 0. A. Schmidt and B. Haasdonk. Reduced basis approximation of large scale parametric algebraic Riccati equations. ESAIM: Control, Optimisation and Calculus of Variations (Feb. 2017). V. Simoncini. Analysis of the Rational Krylov Subspace Projection Method for Large-Scale Algebraic Riccati Equations. SIAM Journal on Matrix Analysis and Applications (Jan. 2016). 9/24

26 Low-Rank Factor Greedy Algorithm Algorithm 1: Low-Rank Factor Greedy Algorithm (LRFG) for the basis generation. Data: Initial basis matrix V 0, training set P train P, tolerance ε, inner tolerance tol i [0,1], error indicator (V,µ) Set V := V 0. while max µ Ptrain (V,µ) > ε do µ := argmax µ Ptrain (V,µ) Solve the full dimensional ARE for the low rank factor Z(µ ) Set Z := (I n VV T )Z(µ ) Ẑ = POD(Z,tol i ) Extend the current basis matrix V = (V,Ẑ) Return V. 10/24

27 Low-Rank Factor Greedy Algorithm Algorithm 2: Low-Rank Factor Greedy Algorithm (LRFG) for the basis generation. Data: Initial basis matrix V 0, training set P train P, tolerance ε, inner tolerance tol i [0,1], error indicator (V,µ) Set V := V 0. while max µ Ptrain (V,µ) > ε do µ := argmax µ Ptrain (V,µ) Solve the full dimensional ARE for the low rank factor Z(µ ) Set Z := (I n VV T )Z(µ ) Ẑ = POD(Z,tol i ) Extend the current basis matrix V = (V,Ẑ) Return V. 10/24

28 Low-Rank Factor Greedy Algorithm Algorithm 3: Low-Rank Factor Greedy Algorithm (LRFG) for the basis generation. Data: Initial basis matrix V 0, training set P train P, tolerance ε, inner tolerance tol i [0,1], error indicator (V,µ) Set V := V 0. while max µ Ptrain (V,µ) > ε do µ := argmax µ Ptrain (V,µ) Solve the full dimensional ARE for the low rank factor Z(µ ) Set Z := (I n VV T )Z(µ ) Ẑ = POD(Z,tol i ) Extend the current basis matrix V = (V,Ẑ) Return V. (W,µ) := R( P(µ)) F C(µ) T Q(µ)C(µ) F 10/24

29 Low-Rank Factor Greedy Algorithm Algorithm 4: Low-Rank Factor Greedy Algorithm (LRFG) for the basis generation. Data: Initial basis matrix V 0, training set P train P, tolerance ε, inner tolerance tol i [0,1], error indicator (V,µ) Set V := V 0. while max µ Ptrain (V,µ) > ε do µ := argmax µ Ptrain (V,µ) Solve the full dimensional ARE for the low rank factor Z(µ ) Set Z := (I n VV T )Z(µ ) Ẑ = POD(Z,tol i ) Extend the current basis matrix V = (V,Ẑ) Return V. (W,µ) := R( P(µ)) F C(µ) T Q(µ)C(µ) F 10/24

30 Low-Rank Factor Greedy Algorithm Algorithm 5: Low-Rank Factor Greedy Algorithm (LRFG) for the basis generation. Data: Initial basis matrix V 0, training set P train P, tolerance ε, inner tolerance tol i [0,1], error indicator (V,µ) Set V := V 0. while max µ Ptrain (V,µ) > ε do µ := argmax µ Ptrain (V,µ) Solve the full dimensional ARE for the low rank factor Z(µ ) Set Z := (I n VV T )Z(µ ) Ẑ = POD(Z,tol i ) Extend the current basis matrix V = (V,Ẑ) Return V. (W,µ) := R( P(µ)) F C(µ) T Q(µ)C(µ) F POD(X,tol i ) returns first l SVs, such that: li=1 σ 2 i tol i (> 0.99) rank(x) i=1 σ 2 i 10/24

31 What is the overall idea? The observer equation The state estimation involves an n-dimensional ODE: x K (t;µ) = (A(µ) B(µ)K(µ))x K (t;µ) +L(µ)[y(t) C y (µ)x K (t;µ)]. Apply POD by calculating snapshots for some µ P and inputs y(t): X := [x K (t 1 ;µ 1 ),x K (t 2 ;µ 1 ),...,x K (t l ;µ P )] Figure: Some bases for an advection diffusion equation. S. Volkwein. Lecture Notes: Proper Orthogonal Decomposition: Theory and Reduced-Order Modelling /24

32 What is the overall idea? The observer equation The state estimation involves an n-dimensional ODE: x K (t;µ) = (A(µ) B(µ)K(µ))x K (t;µ) +L(µ)[y(t) C y (µ)x K (t;µ)]. Apply POD by calculating snapshots for some µ P and inputs y(t): X := [x K (t 1 ;µ 1 ),x K (t 2 ;µ 1 ),...,x K (t l ;µ P )] Figure: Some bases for an advection diffusion equation. S. Volkwein. Lecture Notes: Proper Orthogonal Decomposition: Theory and Reduced-Order Modelling /24

33 The Reduced Controller Or: Why is now everything faster? Build three different bases V P,V Q and V xk (independently, offline). See[S., Haasdonk 2016]. Definition 2 (Reduced controller) Let P NP and Q NQ be the solutions of the reduced AREs and define the reduced gains K := (D T zud zu ) 1 B T u P, L := QC T y (D yw D T yw) 1, where and the control signal P = V P P NP V T P, Q = V Q Q NQ V T Q, û(t) := K x(t), where x(t) R N x satisfies the reduced observer equation x = V T x K (A B K LC y )V xk x +V T x K Ly A. Schmidt and B. Haasdonk. Reduced basis method for H2 optimal feedback control problems. IFAC-PapersOnLine (2016). 2nd IFAC Workshop on Control of Systems Governed by Partial Differential Equations CPDE /24

34 The Reduced Controller Or: Why is now everything faster? Build three different bases V P,V Q and V xk (independently, offline). See[S., Haasdonk 2016]. Definition 2 (Reduced controller) Let P NP and Q NQ be the solutions of the reduced AREs and define the reduced gains K := (D T zud zu ) 1 B T u P, L := QC T y (D yw D T yw) 1, where and the control signal P = V P P NP V T P, Q = V Q Q NQ V T Q, û(t) := K x(t), where x(t) R N x satisfies the reduced observer equation x = V T x K (A B K LC y )V xk x +V T x K Ly A. Schmidt and B. Haasdonk. Reduced basis method for H2 optimal feedback control problems. IFAC-PapersOnLine (2016). 2nd IFAC Workshop on Control of Systems Governed by Partial Differential Equations CPDE /24

35 Numerical Results 13/24

36 Distributed Control of Damped 2D Wave Consider for Ω := [0,1] 2 the following damped wave-equation with parameters µ 1,µ 2 [0.1,2] [1,100]. f tt 0.1 f + µ 1 f t = 10 1 Ωu u +3 1 Ωw w 1, t > 0, f (0,ξ;µ) = sin(ξ 1 π)sin(ξ 2 π), ξ Ω f t (0,ξ;µ) = 0, ξ Ω, s(t;µ) = 1 f (t,ξ;µ)dξ +0.05w 2 (t). Ω y Ω y Consider two performance outputs: o 1 (t) = 1 µ 2 f (t,ξ;µ)dξ, Ω y Ω y Ω w Ω u Ω y Output o1(t) o 2 (t) = 0.1u(t) Uncontrolled Controlled /24

37 Distributed Control of Damped 2D Wave Basis generation results The equations are discretized in space by using FD: n = 800. We construct the three bases: V P,V Q and V x. The observer system was simulated with y(t) = sin(t 2 ) with t [0,2π]. t full [s] t red [s] Basis size Feedback ARE P Observer ARE Q State Estimation ODE x Table: Basis generation results. Overall RB approximation Overall controller is ODE of dimension 25 instead of 800 and the speed up for the AREs is in the magnitude of (using care). By using parameter separability, e.g. A(µ) = Q A i=1 θ i(µ)a i the online simulation can be implemented independent of n. 15/24

38 Distributed Control of Damped 2D Wave Approximation of control signal Setting: Full basis for observer gain basis and state estimation basis x Control u(t) Full N P =5 N P =10 N P = Time t 16/24

39 Distributed Control of Damped 2D Wave Approximation of control signal Setting: Full basis for feedback and observer gain basis Control u(t) 2 Full 4 N x =5 N x =10 N x = Time t 17/24

40 Distributed Control of Damped 2D Wave Approximation of the H 2 -norm Relative maximum error over test set P test with P test = 50. G( ;µ) Ĝ( ;µ) 2 max. (5) µ P test G( ;µ) 2 Table: Relative error in the transfer functions from w to z in the closed loop systems. N Q N P e e e e e e e e e e e e e e e e-04 Surprise: No instabilities occured! See famous one-page article[doyle 78]. J. Doyle. Guaranteed margins for LQG regulators. Automatic Control, IEEE Transactions on (Aug. 1978). 18/24

41 Conclusion and Outlook 19/24

42 Conclusion and Outlook We have seen: Model order reduction for H 2 optimal control problems. Full parametric and realistic control setup Expensive due to two AREs and state estimation Model reduction for AREs and the state estimation Large speed-up What next? Stability considerations and error estimation Robustification? Same approach for H -control problems 20/24

43 Conclusion and Outlook We have seen: Model order reduction for H 2 optimal control problems. Full parametric and realistic control setup Expensive due to two AREs and state estimation Model reduction for AREs and the state estimation Large speed-up What next? Stability considerations and error estimation Robustification? Same approach for H -control problems 20/24

44 Thank you for your attention! Questions? Comments? 21/24

45 References I J. Doyle. Guaranteed margins for LQG regulators. Automatic Control, IEEE Transactions on 23.4 (Aug. 1978). B. Haasdonk. Reduced Basis Methods for Parametrized PDEs A Tutorial Introduction for Stationary and Instationary Problems. SimTech Preprint. Chapter to appear in P. Benner, A. Cohen, M. Ohlberger and K. Willcox (eds.): "Model Reduction and Approximation: Theory and Algorithms", SIAM, Philadelphia, IANS, University of Stuttgart, Germany, A. Patera and G. Rozza. Reduced Basis Approximation and a Posteriori Error Estimation for Parametrized Partial Differential Equations A. Schmidt and B. Haasdonk. Reduced basis method for H2 optimal feedback control problems. IFAC-PapersOnLine 49.8 (2016). 2nd IFAC Workshop on Control of Systems Governed by Partial Differential Equations CPDE A. Schmidt and B. Haasdonk. Reduced basis approximation of large scale parametric algebraic Riccati equations. ESAIM: Control, Optimisation and Calculus of Variations (Feb. 2017). V. Simoncini. Analysis of the Rational Krylov Subspace Projection Method for Large-Scale Algebraic Riccati Equations. SIAM Journal on Matrix Analysis and Applications 37.4 (Jan. 2016). 22/24

46 References II S. Volkwein. Lecture Notes: Proper Orthogonal Decomposition: Theory and Reduced-Order Modelling K. Zhou, J. Doyle and K. Glover. Robust and Optimal Control /24

47 The Control Problem Mathematical model of the controller Consider the open-loop system x = Ax +B u u +B w w z = C z x +D zu u y = C y x +D yw w Ansatz for controller w(t) u(t) System Controller y(t) z(t) x K = A K x K +B K y u = C K x K The interconnection (i.e. the closed-loop system) for h T = (x,x K ) can be written as h = A h + Bw, z = C h. 24/24

48 The Control Problem Mathematical model of the controller Consider the open-loop system x = Ax +B u u +B w w z = C z x +D zu u y = C y x +D yw w Ansatz for controller w(t) u(t) System Controller y(t) z(t) x K = A K x K +B K y u = C K x K The interconnection (i.e. the closed-loop system) for h T = (x,x K ) can be written as h = A h + Bw, z = C h. 24/24

49 The Control Problem Mathematical model of the controller Consider the open-loop system x = Ax +B u u +B w w z = C z x +D zu u y = C y x +D yw w Ansatz for controller w(t) u(t) System Controller y(t) z(t) x K = A K x K +B K y u = C K x K The interconnection (i.e. the closed-loop system) for h T = (x,x K ) can be written as h = A h + Bw, z = C h. 24/24