This is machine translation

Translated by Microsoft
Mouseover text to see original. Click the button below to return to the English version of the page.

Note: This page has been translated by MathWorks. Click here to see
To view all translated materials including this page, select Country from the country navigator on the bottom of this page.

Designing Model Predictive Controller Equivalent to Infinite-Horizon LQR

This example shows how to design an infinite-horizon model predictive controller by setting the weights on the terminal predicted states.

Define Plant Model

The linear open-loop dynamic model is defined below. DC gain is also computed.

Ts = 0.1;                                   % Sampling time
A = [0.8 Ts;0 0.9];
B = [0;Ts];
C = [1 0];
sysd = ss(A,B,C,0,Ts);                      % Discrete-time plant model
dcg = dcgain(sysd);                         % DC-gain of prediction model

Design an Infinite-Horizon LQR Controller

Compute the Riccati matrix associated with the LQR problem with output weight Qy and input weight Qu:

Qy = 10;                                    % Output weight: y'*Qy*y
Qu = 0.1;                                   % Input weight:  u'*Qu*y
[K,P] = lqry(sysd,Qy,Qu);                   % LQR gain and Riccati matrix

Weight the terminal state x'(t+p)*P*x(t+p), where p is the prediction horizon of the MPC controller. Compute the Cholesky factor chol(P) of the Riccati matrix P, so that the terminal penalty becomes:

x'(t+p)*P*x(t+p) = [chol(P)*x(t+p)]'*[chol(P)*x(t+p)] = yc'(t+p)*yc(t+p)

where the new output yc(t+p) = chol(P)*x(t+p) and the state x(t) is assumed to be fully measurable.

cholP = chol(P);

Change and augment the output vector to include the full state x and yc: Output = state vector x + output yc such that yc'*yc = x'*P*x

sysd.C = [eye(2);cholP];
sysd.D = zeros(4,1);

Convert the LQR Controller Into a Finite-Horizon MPC Controller

Label the new additional output signal yc(t) as unmeasured:

sysd = setmpcsignals(sysd,'UO',[3 4]); % Cholesky factor is not measured
-->Assuming unspecified output signals are measured outputs.

Create the controller object with sampling period, prediction and control horizons:

p = 3;                   % Prediction horizon (for any p>=1, unconstrained MPC = LQR)
m = p;                   % Control horizon = prediction horizon
mpcobj = mpc(sysd,Ts,p,m); % MPC object
-->The "Weights.ManipulatedVariables" property of "mpc" object is empty. Assuming default 0.00000.
-->The "Weights.ManipulatedVariablesRate" property of "mpc" object is empty. Assuming default 0.10000.
-->The "Weights.OutputVariables" property of "mpc" object is empty. Assuming default 1.00000.
   for output(s) y1 and zero weight for output(s) y2 y3 y4 

Specify actuator saturation limits as MV constraints.

mpcobj.MV = struct('Min',-3,'Max',3);

Define input and output weights at each step of the prediction horizon (terminal weights are set later). Very small weights on input increments are included to make the QP problem associated with the MPC controller positive definite:

mpcobj.Weights.OV = [sqrt(Qy) 0 0 0];       % Output weights (only on original output)
mpcobj.Weights.MV = sqrt(Qu);               % Input weight
mpcobj.Weights.MVRate = 1e-5;               % Very small weight on command input increments

Define set-points for the output and input signals:

ry = 1;                                     % Output set point
mpcobj.MV.Target = ry/dcg;                  % Set-point for manipulated variable

Impose the terminal penalty x'(t+p)*P*x(t+p) by specifying a unit weight on yc(t+p) = chol(P)*x(t+p). The terminal weight on u(t+p-1) remains the same, that is sqrt(Qu):

Y = struct('Weight',[0 0 1 1]);             % Weight on y(t+p)
U = struct('Weight',sqrt(Qu));              % Weight on u(t+p-1)
setterminal(mpcobj,Y,U);                    % Set terminal weight y'*y = x'*P*x

Since the measured output is the entire state, remove any additional output disturbance integrator inserted by the MPC controller:


Remove the state estimator by defining the following measurement update equation:

x[n|n] = x[n|n-1] + I * (x[n]-x[n|n-1]) = x[n]

Because setterminal function resets state estimator to default value, we use setEstimator function to change state estimator after setterminal is called.

setEstimator(mpcobj,[],eye(2));             % State estimate = state measurement

Compare MPC and LQR Controllers

Compute the gain of the MPC controller when constraints are inactive, and compare it to the LQR gain:

mpcgain = dcgain(ss(mpcobj));
fprintf('\n(unconstrained) MPC: u(k)=[%8.8g,%8.8g]*x(k)',mpcgain(1),mpcgain(2));
fprintf('\n                LQR: u(k)=[%8.8g,%8.8g]*x(k)\n\n',-K(1),-K(2));
-->The "Model.Noise" property of the "mpc" object is empty. Assuming white noise on each measured output channel.

(unconstrained) MPC: u(k)=[-2.8363891,-2.1454028]*x(k)
                LQR: u(k)=[-2.8363891,-2.1454028]*x(k)

The state feedback gains are exactly the same.

Simulate Using Simulink®

To run this example, Simulink® is required.

if ~mpcchecktoolboxinstalled('simulink')
    disp('Simulink(R) is required to run this example.')

Define a set-point for the new extended output vector containing x and chol(P)*x:

rx = (eye(2)-A)\B/dcg;
r = [rx;cholP*rx];                          % Set point for extended prediction model

Simulate closed-loop control of the linear plant model in Simulink. Controller "mpcobj" is specified in the block dialog.

mdl = 'mpc_infinite';
open_system(mdl);                           % Open Simulink(R) Model
sim(mdl);                                   % Start Simulation

The closed-loop response shows good setpoint tracking performance.


See Also

Related Topics