Model Predictive Control

Transcription

1 Model Predictive Control Linear time-varying and nonlinear MPC Alberto Bemporad A. Bemporad - "Model Predictive Control" 1/30

2 Course structure Linear model predictive control (MPC) Linear time-varying and nonlinear MPC MPC computations: quadratic programming (QP), explicit MPC Hybrid MPC Stochastic MPC Data-driven MPC MATLAB Toolboxes: MPC Toolbox (linear/explicit/parameter-varying MPC) Hybrid Toolbox (explicit MPC, hybrid systems) Course page: A. Bemporad - "Model Predictive Control" 2/30

3 Linear Time-Varying Model Predictive Control

4 LPV models Linear Parameter-Varying (LPV) model { x k+1 = A(p(t))x k + B(p(t))u k + B v (p(t))v k y k = C(p(t))x k + D v (p(t))v k that depends on a vector p(t) of parameters The weights in the quadratic performance index can also be LPV The resulting optimization problem is still a QP ] 1 min z 2 z H(p(t))z + F (p(t)) z s.t. [ x(t) r(t) u(t 1) G(p(t))z W (p(t)) + S(p(t)) [ x(t) ] r(t) u(t 1) The QP matrices must be constructed online, contrarily to the LTI case 2018 A. Bemporad - "Model Predictive Control" 3/30

5 Linearizing a nonlinear model: LPV case An LPV model can be obtained by linearizing the nonlinear model { ẋ c (t) = f(x c (t), u c (t), p c (t)) y c (t) = g(x c (t), p c (t)) p c R np = a vector of exogenous signals (e.g., ambient conditions) At time t, consider nominal values x c, ū c, p c and linearize ẋ c (t) f (x c (t) x c ) + f (u c (t) ū c ) x xc,ū }{{ c, p } c u xc,ū }{{ c, p } c A c B c [ ] f [ + f( x c, ū c, p c ) pc(t) p ] c p 1 xc,ū c, p c }{{} B vc Convert (A c, [B c B vc ]) to discrete-time and get prediction model (A, [B B v ]) Same thing for the output equation, to get matrices C and D v 2018 A. Bemporad - "Model Predictive Control" 4/30

6 LTV models Linear Time-Varying (LTV) model { x k+1 = A k (t)x k + B k (t)u k y k = C k (t)x k At each time t the model can also change over the prediction horizon k The measured disturbance is embedded in the model The resulting optimization problem is still a QP min z 1 2 z H(t)z + s.t. [ x(t) r(t) u(t 1) G(t)z W (t) + S(t) ] F (t) z [ x(t) ] r(t) u(t 1) As for LPV-MPC, the QP matrices must be constructed online 2018 A. Bemporad - "Model Predictive Control" 5/30

7 Linearizing a nonlinear model: LTV case LPV/LTV models can be obtained by linearizing nonlinear models { ẋ c (t) = f(x c (t), u c (t), p c (t)) y c (t) = g(x c (t), p c (t)) At time t, consider nominal trajectories U = {ū c (t), ū c (t + T s ),..., ū c (t + (N 1)T s )} (example: U = shifted previous optimal sequence or input ref. trajectory) P = { p c (t), p c (t + T s ),..., p c (t + (N 1)T s )} (no preview: p c(t + k) p c(t)) Integrate the model and get nominal state/output trajectories X = { x c (t), x c (t + T s ),..., x c (t + (N 1)T s )} Y = {ȳ c (t), ȳ c (t + T s ),..., ȳ c (t + (N 1)T s )} Examples: x c (t) = current x c (t); x c (t) = equilibrium; x c (t) = reference 2018 A. Bemporad - "Model Predictive Control" 6/30

8 Linearizing a nonlinear model: LTV case While integrating, also compute the sensitivities A k (t) = x c(t + (k + 1)T s ) x c (t + kt s ) B k (t) = x c(t + (k + 1)T s ) ū c (t + kt s ) C k (t) = ȳ c(t + kt s ) x c (t + kt s ) Approximate the NL model as the LTV model x k+1 {}}{ x c (k + 1) x c (k + 1) = A k (t) x k { }} { { }} { (x c (k) x c (k)) +B k (t) (u c (k) ū c (k)) y k = C k (t) (x c (k) x c (k)) }{{} y c (k) ȳ c (k) }{{} x k u k 2018 A. Bemporad - "Model Predictive Control" 7/30

9 LTV-MPC example Process model is LTV d 3 y dt 3 + y 3d2 dt 2 + 2dy + (6 + sin(5t))y = 5du dt dt + ( ( )) cos 2 t u LTI-MPC cannot track the setpoint, LPV-MPC tries to catch-up with time-varying model, LTV-MPC has preview on future model values (See demo TimeVaryingMPCControlOfATimeVaryingLinearSystemExample in MPC Toolbox) 2018 A. Bemporad - "Model Predictive Control" 8/30

10 LTV-MPC example Define LTV model Models = tf; ct = 1; for t = 0:0.1:10 Models(:,:,ct) = tf([5 5+2*cos(2.5*t)],[ sin(5*t)]); ct = ct + 1; end Ts = 0.1; % sampling time Models = ss(c2d(models,ts)); Design MPC controller sys = ss(c2d(tf([5 5],[ ]),Ts)); % average model time p = 3; % prediction horizon m = 3; % control horizon mpcobj = mpc(sys,ts,p,m); mpcobj.mv = struct('min',-2,'max',2); % input constraints mpcobj.weights = struct('mv',0,'mvrate',0.01,'output',1); 2018 A. Bemporad - "Model Predictive Control" 9/30

11 LTV-MPC example Simulate LTV system with LTI-MPC controller for ct = 1:(Tstop/Ts+1) real_plant = Models(:,:,ct); % Get the current plant y = real_plant.c*x; u = mpcmove(mpcobj,xmpc,y,1); % Apply LTI MPC x = real_plant.a*x + real_plant.b*u; end Simulate LTV system with LPV-MPC controller for ct = 1:(Tstop/Ts+1) real_plant = Models(:,:,ct); % Get the current plant y = real_plant.c*x; u = mpcmoveadaptive(mpcobj,xmpc,real_plant,nominal,y,1); x = real_plant.a*x + real_plant.b*u; end 2018 A. Bemporad - "Model Predictive Control" 10/30

12 LTV-MPC example t Simulate LTV system with LTV-MPC controller for ct = 1:(Tstop/Ts+1) real_plant = Models(:,:,ct); % Get the current plant y = real_plant.c*x; u = mpcmoveadaptive(mpcobj,xmpc,models(:,:,ct:ct+p),nominals,y,1); x = real_plant.a*x + real_plant.b*u; end Simulate in Simulink A B C Clock Zero-Order Hold t D U Y X DX Time Varying Predictive Model model mo ref Adaptive MPC mv y u Time-Varying Plant ysim To Workspace1 mpc_timevarying.mdl 1 Reference Adaptive MPC Controller usim To Workspace 2018 A. Bemporad - "Model Predictive Control" 11/30

13 LTV-MPC example Simulink block need to provide 3D array of future models mpc_timevarying.mdl 2018 A. Bemporad - "Model Predictive Control" 12/30

14 Example: LPV-MPC of a nonlinear CSTR system MPC control of a diabatic continuous stirred tank reactor (CSTR) Process model is nonlinear dc A dt dt dt = F V (C Af C A ) C A k 0 e E RT = F V (T f T ) + UA ρc (T j T ) H C A k 0 e E RT pv ρc p T : temperature inside the reactor [K] (state) C A : concentration of the reagent in the reactor [kgmol/m 3 ] (state) T j : jacket temperature [K] (input) T f : feedstream temperature [K] (measured disturbance) F Tf T A B C Af : feedstream concentration [kgmol/m 3 ] (measured disturbance) Objective: manipulate T j to regulate C A on desired setpoint CAf Tj CA >> edit ampccstr_linearization (MPC Toolbox) 2018 A. Bemporad - "Model Predictive Control" 13/30

15 Example: LPV-MPC of a nonlinear CSTR system Process model >> mpc_cstr_plant 1 CAi 2 Ti 3 Tc Feed Concentration Reactor Temperature Feed Temperature Reactor Concentration Coolant Temperature CSTR 1 T 2 CA % Create operating point specification. plant_mdl = 'mpc_cstr_plant'; op = operspec(plant_mdl); op.inputs(1).u = 10; % Feed concentration condition op.inputs(1).known = true; op.inputs(2).u = ; % Feed concentration condition op.inputs(2).known = true; op.inputs(3).u = ; % Coolant temperature condition op.inputs(3).known = true; x0 = [op_report.states(1).x;op_report.states(2).x]; y0 = [op_report.outputs(1).y;op_report.outputs(2).y]; u0 = [op_report.inputs(1).u;op_report.inputs(2).u;op_report.inputs(3).u]; % Obtain linear plant model at the initial condition. sys = linearize(plant_mdl, op_point); sys = sys(:,2:3); % First plant input CAi dropped because not used by MPC 2018 A. Bemporad - "Model Predictive Control" 14/30

16 Example: LPV-MPC of a nonlinear CSTR system MPC design % Discretize the plant model Ts = 0.5; % hours plant = c2d(sys,ts); % Design MPC Controller % Specify signal types used in MPC plant.inputgroup.measureddisturbances = 1; plant.inputgroup.manipulatedvariables = 2; plant.outputgroup.measured = 1; plant.outputgroup.unmeasured = 2; plant.inputname = 'Ti','Tc'; plant.outputname = 'T','CA'; % Create MPC controller with default prediction and control horizons mpcobj = mpc(plant); % Set nominal values in the controller mpcobj.model.nominal = struct('x', x0, 'U', u0(2:3), 'Y', y0, 'DX', [0 0]); 2018 A. Bemporad - "Model Predictive Control" 15/30

17 Example: LPV-MPC of a nonlinear CSTR system MPC design (cont d) % Set scale factors because plant input and output signals have different % orders of magnitude Uscale = [30 50]; Yscale = [50 10]; mpcobj.dv(1).scalefactor = Uscale(1); mpcobj.mv(1).scalefactor = Uscale(2); mpcobj.ov(1).scalefactor = Yscale(1); mpcobj.ov(2).scalefactor = Yscale(2); % Let reactor temperature T float (i.e. with no setpoint tracking error % penalty), because the objective is to control reactor concentration CA % and only one manipulated variable (coolant temperature Tc) is available. mpcobj.weights.ov = [0 1]; % Due to the physical constraint of coolant jacket, Tc rate of change is % bounded by degrees per minute. mpcobj.mv.ratemin = -2; mpcobj.mv.ratemax = 2; 2018 A. Bemporad - "Model Predictive Control" 16/30

18 Example: LPV-MPC of a nonlinear CSTR system Simulink diagram 3 Measurement Noise u0(1) u0(2) Disturbance 3 CAi Ti Tc plant model CSTR T CA Reactor Feed 3 3 Temperature u0(1) 2 [2x2] A T A [2x2] [2x2] B B [2x2] [2x2] C CA C [2x2] [2x2] D D [2x2] 8{24} [2x1] CAi U U [2x1] [2x1] Y Y [2x1] Ti [2x1] X X [2x1] [2x1] DX DX [2x1] Tc [2x1] poles Poles [2x1] Successive Linearizer Pole 3 Concentration 3 Estimated True Selector 3 Coolant 3 mv est.state Adaptive MPC model mo ref md Adaptive MPC Controller 8{24} CA Reference linearize + discretize Reference Copyright The MathWorks, Inc. LPV-MPC >> edit ampc_cstr_linearization 2018 A. Bemporad - "Model Predictive Control" 17/30

19 Example: LPV-MPC of a nonlinear CSTR system Closed-loop results 2018 A. Bemporad - "Model Predictive Control" 18/30

20 Example: LTI-MPC of a nonlinear CSTR system Closed-loop results with LTI-MPC, same tuning Noise 3 3 u0(1) CAi CAi T Reactor u0(2) Ti Ti Tc CA Feed 3 3 Temperature Disturbance CSTR Coolant mo 0 LTI-MPC mv MPC 2 ref md 2 CA Reference MPC Controller Reference 2 2 Concentration True Copyright The MathWorks, Inc A. Bemporad - "Model Predictive Control" 19/30

21 Example: LTI-MPC of a nonlinear CSTR system Closed-loop results 2018 A. Bemporad - "Model Predictive Control" 20/30

22 LTV Kalman filter Process model = LTV model with noise x(k + 1) = A(k)x(k) + B(k)u(k)+G(k)ξ(k) y(k) = C(k)x(k) + ζ(k) ξ(k) R q = zero-mean white process noise with covariance Q(k) 0 ζ(k) R p = zero-mean white measurement noise with covariance R(k) 0 measurement update: time update: M(k) = P (k k 1)C(k) [C(k)P (k k 1)C(k) + R(k)] 1 ˆx(k k) = ˆx(k k 1) + M(k) (y(k) C(k)ˆx(k k 1)) P (k k) = (I M(k)C(k))P (k k 1) ˆx(k + 1 k) = Aˆx(k k) + Bu(k) P (k + 1 k) = A(k)P (k k)a(k) + G(k)Q(k)G(k) 2018 A. Bemporad - "Model Predictive Control" 21/30

23 Extended Kalman filter Process model = nonlinear model with noise measurement update: time update: x(k + 1) = f(x(k), u(k), ξ(k)) y(k) = g(x(k), u(k)) + ζ(k) C(k) = g x (ˆx k k 1, u(k)) M(k) = P (k k 1)C(k) [C(k)P (k k 1)C(k) + R(k)] 1 ˆx(k k) = ˆx(k k 1) + M(k) (y(k) g(ˆx(k k 1), u(k))) P (k k) = (I M(k)C(k))P (k k 1) ˆx(k + 1 k) = f(ˆx(k k), u(k)) A(k) = f x (ˆx k k, u(k), E[ξ(k)]), G(k) = f ξ (ˆx k k, u(k), E[ξ(k)]) P (k + 1 k) = A(k)P (k k)a(k) + G(k)Q(k)G(k) 2018 A. Bemporad - "Model Predictive Control" 22/30

24 Nonlinear Model Predictive Control

25 Nonlinear MPC (Mayne, Rawlings, Diehl, 2017) Nonlinear prediction model {xk+1 = f(xk, uk) Nonlinear constraints h(x k, u k ) 0 y k = g(x k, u k ) N 1 Nonlinear performance index min l N (x N ) + l(x k, u k ) k=0 Optimization problem: nonlinear programming problem (NLP) min z F (z, x(t)) s.t. G(z, x(t)) 0 H(z, x(t)) = 0 z = u 0. u N 1 x 1. x N 2018 A. Bemporad - "Model Predictive Control" 23/30

26 Nonlinear optimization (Nocedal, Wright, 2006) (Nonconvex) NLP is harder to solve than QP Convergence to a global optimum may not be guaranteed Several NLP solvers exist (such as Sequential Quadratic Programming (SQP)) NL-MPC is not used in practice so often, except for dealing with strong dynamical nonlinearities and slow processes (such as chemical processes) 2018 A. Bemporad - "Model Predictive Control" 24/30

27 Fast nonlinear MPC (Lopez-Negrete, D Amato, Biegler, Kumar, 2013) Fast MPC: exploit sensitivity analysis to compensate for the computational delay caused by solving the NLP Key idea: pre-solve the NLP between time t 1 and t based on the predicted state x (t) = f(x(t 1), u(t 1)) in background Get u (t) and the sensitivities u x within sample interval [(t 1)T s, tt s ) At time t, get x(t) and compute u(t) = u (t) + u x (x(t) x (t)) Note that still one NLP must be solved within the sample interval 2018 A. Bemporad - "Model Predictive Control" 25/30

28 From LTV-MPC to NL-MPC Key idea: Solve a sequence of LTV-MPC problems at the same time t Given the current state x(t) and reference {r(t + k), u r (t + k)}, initial guess U 0 = {u 0 0,..., u 0 N 1 } and corresponding state trajectory X 0 = {x 0 0,..., x 0 N } A good initial guess U 0, X 0 is the previous (shifted) optimal solution At a generic iteration i, linearize the NL model around U i, X i : { xk+1 = f(x k, u k ) A i = f(xi 0, u i 0) x, B i = f(xi 0, u i 0) u y k = g(x k, u k ), C i = g(xi 0, u i 0), D i = g(xi 0, u i 0) x u 2018 A. Bemporad - "Model Predictive Control" 26/30

29 Nonlinear MPC For h = 0 to h max 1 do: 1. Simulate from x(t) with inputs U h, parameter sequence P and get state trajectory X h and output trajectory Y h 2. Linearize around (X h, U h, P ) and discretize in time with sample time T s 3. Let U h+1 be the solution of the QP problem corresponding to LTV-MPC 4. Find optimal step size α h (0, 1]; 5. Set U h+1 = (1 α h )U h + α h U h+1; Return solution U hmax The method above is a Gauss-Newton method to solve the NL-MPC problem Special case: just solve one iteration with α = 1 (a.k.a. Real-Time Iteration) (Diehl, Bock, Schloder, Findeisen, Nagy, Allgower, 2002) 2018 A. Bemporad - "Model Predictive Control" 27/30

30 Nonlinear MPC (Gros, Zanon, Quirynen, Bemporad, Diehl, 2016) Example x Linear MPC RTI Fully converged Linear MPC RTI Fully converged x x x u Linear MPC RTI Converged time 0 Fig. 7. Illustration of the RTI solution vs. the linear MPC solutions at the discrete time instant i = 2, with state noise A. Bemporad - "Model Predictive Control" 28/30 u1-0.4

31 Learning nonlinear models for MPC (Masti, Bemporad, CDC 2018) Idea: use autoencoders and artificial neural networks to learn a nonlinear state-space model of desired order from input/output data output map dead-beat observer state map O k = [y k... y k m] (Hinton, Salakhutdinov, 2006) I k = [y k... y k n a+1 u k... u k n b +1] 2018 A. Bemporad - "Model Predictive Control" 29/30

32 Learning nonlinear models for MPC - An example (Masti, Bemporad, CDC 2018) System generating the data = nonlinear 2-tank benchmark x 1(k + 1) = x 1(k) k 1 x1(k) + k 2(u(k) + w(k)) x 2(k + 1) = x 2(k) + k 3 x1(k) k 4 x2(k) y(k) = x 2(k) + v(k) LTV MPC Model is totally unknown to learning algorithm The performance achieved with the derivative-based controller suggests that an LTV-MPC formulation might also works well. We also assess its robustness using a model achieving 61% BFR in open loop Artificial neural network (ANN): 3 hidden layers 60 exponential linear unit (ELU) neurons For given number of model parameters, autoencoder approach is superior to NNARX Jacobians directly obtained from ANN structure for Kalman filtering & MPC problem construction Computation time per step: ~40ms LTV-MPC results ODYS CONFIDEN 2018 A. Bemporad - "Model Predictive Control" 30/30