Pendulum/Cart Assembly

The plant for this example is the following pendulum/cart assembly, wherezis the cart position andthetais the pendulum angle.

The manipulated variable for this system is a variable forceFacting on the cart. The range of the force is between -100 and 100 (MKS units are assumed). The controller needs to keep the pendulum upright while moving the cart to a new position or when the pendulum is nudged forward by an impulse disturbancedFapplied at the upper end of the inverted pendulum.

Control Objectives

Assume the following initial conditions for the pendulum/cart assembly:

The control objectives are:

The upright position is an unstable equilibrium for the inverted pendulum, which makes the control task more challenging.

The Choice of Time-Varying MPC

InControl of an Inverted Pendulum on a Cart, a single MPC controller is able to move the cart to a new position between-10and10. However, if you increase the step setpoint change to20, the pendulum fails to recover its upright position during the transition.

To reach the longer distance within the same rise time, the controller applies more force to the cart at the beginning. As a result, the pendulum is displaced from its upright position by a larger angle, such as60degrees. At such angles, the plant dynamics differ significantly from the LTI predictive model obtained attheta=0. As a result, errors in the prediction of plant behavior exceed what the built-in MPC robustness can handle, and the controller fails to perform properly.

To avoid the pendulum falling, a simple workaround is to restrict pendulum displacement by adding soft output constraints tothetaand reducing the ECR weight (from the default value of 1e5 to 100) to soften the constraints.

mpcobj.OV(2).Min = -pi/2; mpcobj.OV(2).Max = pi/2; mpcobj.Weights.ECR = 100;

However, with these new controller settings it is no longer possible to reach the longer distance within the required rise time. In other words, controller performance is sacrificed to avoid violation of the soft output constraints.

To move the cart to a new position between-20and20while maintaining the same rise time, the controller needs to have more accurate models at different angles so that the controller can use them for better prediction. Adaptive MPC allows you to solve a nonlinear control problem by updating linear time-varying plant models at run time.

Control Structure

For this example, use a single LTV MPC controller with:

mdlMPC ='mpc_pendcartLTVMPC'; open_system(mdlMPC);

Because all the plant states are measurable, they are directly used as custom estimated states in the Adaptive MPC block.

While the cart position setpoint varies (step input), the pendulum angle setpoint is constant (0= upright position).

Linear Time-Varying Plant Models

At each control interval, LTV MPC requires a linear plant model for each prediction step, from current timekto timek+p, wherepis the prediction horizon.

In this example, the cart and pendulum dynamic system is described by a first principle model. This model consists of a set of differential and algebraic equations (DAEs), defined in thependulumCTfunction. For more details, seependulumCT.m.

TheSuccessive Linearizerblock in the Simulink model generates the LTV models at run time. At each prediction step, the block obtains state-space matricesA,B,C, andDusing a Jacobian in continuous-time, and then converts them into discrete-time values. The initial plant states x(k) are directly measured from the plant. The plant input sequence contains the optimal moves generated by the MPC controller in the previous control interval.

Adaptive MPC Design

The MPC controller is designed at its nominal equilibrium operating point.

x0 = zeros(4,1); u0 = zeros(1,1);

Analytically obtain a linear plant model using the ODEs.

[~, ~, A, B, C, D] = pendulumCT (x0,情况);植物= ss (A, B,C([1 3],:),D([1 3],:));% position and angle

To control an unstable plant, the controller sample time cannot be too large (poor disturbance rejection) or too small (excessive computation load). Similarly, the prediction horizon cannot be too long (the plant unstable mode would dominate) or too short (constraint violations would be unforeseen). Use the following parameters for this example:

Ts = 0.01; PredictionHorizon = 60; ControlHorizon = 3;

Create the MPC controller.

mpcobj = mpc(c2d(plant,Ts),Ts,PredictionHorizon,ControlHorizon);

-->The "Weights.ManipulatedVariables" property is empty. Assuming default 0.00000. -->The "Weights.ManipulatedVariablesRate" property is empty. Assuming default 0.10000. -->The "Weights.OutputVariables" property is empty. Assuming default 1.00000. for output(s) y1 and zero weight for output(s) y2

There is a limitation on how much force can be applied to the cart, which is specified using hard constraints on the manipulated variableF.

mpcobj.MV.Min = -100; mpcobj.MV.Max = 100;

It is good practice to scale plant inputs and outputs before designing weights. In this case, since the range of the manipulated variable is greater than the range of the plant outputs by two orders of magnitude, scale the MV input by100.

mpcobj.MV.ScaleFactor = 100;

To improve controller robustness, increase the weight on the MV rate of change from0.1to1.

mpcobj.Weights.MVRate = 1;

To achieve balanced performance, adjust the weights on the plant outputs. The first weight is associated with cart positionz, and the second weight is associated with angletheta.

mpcobj.Weights.OV = [0.6 1.2];

Use a gain as the output disturbance model for the pendulum angle. This represents rapid short-term variability.

setoutdist(mpcobj,'model',[0;tf(1)]);

使用自定义状态估计,因为所有的植物ates are measurable.

setEstimator(mpcobj,'custom');

Closed-Loop Simulation

Validate the MPC design with a closed-loop simulation in Simulink.

open_system([mdlMPC'/Scope']); sim(mdlMPC)

-->The "Model.Noise" property is empty. Assuming white noise on each measured output.

In the nonlinear simulation, all the control objectives are successfully achieved.

bdclose(mdlMPC);