Model Predictive Control Toolbox™ software provides code generation functionality for controllers designed in MATLAB® or Simulink®.
After designing an MPC controller in MATLAB, you can generate C code using MATLAB Coder™ and deploy it for real-time control.
To generate code for computing optimal MPC control moves for an implicit or explicit linear MPC controller:
Generate data structures from an MPC controller or explicit MPC controller
For an example, see Generate Code to Compute Optimal MPC Moves in MATLAB.
You can also generate code for nonlinear MPC controllers that use the default
fmincon solver with the SQP algorithm. To generate code for
computing optimal control moves for a nonlinear MPC controller:
Generate data structures from a nonlinear MPC controller using
To verify that your controller produces the expected closed-loop results,
simulate it using
nlmpcmoveCodeGeneration in place of
Generate code for
codegen (MATLAB Coder). This step requires
After designing a controller in Simulink using any of the MPC blocks, you can generate code and deploy it for real-time control. You can deploy controllers to all targets supported by the following products:
Simulink PLC Coder™
You can generate code for any of the Model Predictive Control Toolbox Simulink blocks.
|Types of Controllers||Blocks|
|Implicit MPC controllers|
|Explicit MPC controllers|
|Gain-scheduled MPC controllers|
|Adaptive MPC controllers|
|MPC controllers for automotive applications|
|Nonlinear MPC controllers that use |
For more information on generating code, see Simulation and Code Generation Using Simulink Coder and Simulation and Structured Text Generation Using Simulink PLC Coder.
The MPC Controller, Explicit MPC Controller, Adaptive MPC Controller, and Nonlinear MPC Controller blocks are implemented using the MATLAB Function (Simulink) block. To see the structure, right-click the block, and select Mask > Look Under Mask. Then, open the MPC subsystem underneath.
If your nonlinear MPC controller uses optional parameters, you must also generate code for the Bus Creator block connected to the params input port of the Nonlinear MPC Controller block. To do so, place the Nonlinear MPC Controller and Bus Creator blocks within a subsystem, and generate code for that subsystem.
You can generate CUDA® code for your MPC controller using GPU Coder™. For more information on supported GPUs, see GPU Support by Release (Parallel Computing Toolbox). For more information on installing and setting up the prerequisite product, see Installing Prerequisite Products (GPU Coder) and Setting Up the Prerequisite Products (GPU Coder).
To generate and use GPU code in MATLAB:
Design a linear controller using an
Generate the structures for the core, states, and online data from your
linear MPC controller using the
Optionally simulate your closed loop iteratively using the
mpcmoveCodeGeneration function and the data
structures created in the previous step.
Create a coder configuration options object using the
coder.gpuConfig function, and configure the code
Generate code for the
codegen function and the coder configuration
options object. Doing so generates a new function which uses code running on
Simulate your controller using the new generated function and the data structures.
For an example on using GPU code in MATLAB, see Use the GPU to Compute MPC Moves in MATLAB
To generate GPU code from a Simulink model containing any of these blocks, open the Configuration Parameters dialog box by clicking Model Settings. Then, in the Code Generation section, select Generate GPU code.
For details on how to configure your model for GPU code generation, see Code Generation from Simulink Models with GPU Coder (GPU Coder).
The sampling rate that a controller can achieve in a real-time environment is system-dependent. For example, for a typical small MIMO control application running on Simulink Real-Time, the sample time can be as long as 1–10 ms for linear MPC and 100–1000 ms for nonlinear MPC. To determine the sample time, first test a less-aggressive controller whose sampling rate produces acceptable performance on the target. Next, decrease the sample time and monitor the execution time of the controller. You can further decrease the sample time as long as the optimization safely completes within each sampling period under normal plant operating conditions. To reduce the sample time, you can also consider using:
Explicit MPC. While explicit MPC controllers have a faster execution time, they also have a larger memory footprint, since they store precomputed control laws. For more information, see Explicit MPC Design.
A suboptimal QP solution after a specified number of maximum solver iterations. For more information, see Suboptimal QP Solution.
A lower controller sample time does not necessarily provide better performance. In fact, you want to choose a sample time that is small enough to give you good performance but no smaller. For the same prediction time, smaller sample times result in larger prediction steps, which in turn produces a larger memory footprint and more complex optimization problem.
At each control interval, an implicit or adaptive MPC controller constructs a new QP problem, which is defined as:
subject to the linear inequality constraints
x is the solution vector.
H is the Hessian matrix.
A is a matrix of linear constraint coefficients.
f and b are vectors.
In generated C code, the following matrices are used to provide H, A, f, and b. Depending on the type and configuration of the MPC controller, these matrices are either constant or regenerated at each control interval.
|Constant Matrix||Size||Purpose||Implicit MPC||Implicit MPC with Online Weight Tuning||Adaptive MPC or LTV MPC|
|NM-by-NM||Inverse of the Hessian matrix, H||Constant||Regenerated||Regenerated|
|NM-by-NM||Inverse of the lower-triangular Cholesky decomposition of H|
|NC-by-NM||Linear constraint coefficients, A||Constant|
|Nxqp-by-(NM–1)||Used to generate f||Regenerated|
|NC-by-1||Used to generate b||Constant||Constant, except when there are custom constraints|
p is the prediction horizon.
Nmv is the number of manipulated variables.
Nmd is the number of measured disturbances.
Ny is the number of output variables.
NM is the number of optimization variables (m*Nmv+1, where m is the control horizon).
Nxqp is the number of states used for the QP problem; that is, the total number of the plant states and disturbance model states.
NC is the total number of constraints.
At each control interval, the generated C code computes f and b as:
xq is the vector of plant and disturbance model states estimated by the Kalman filter.
ml is the manipulated variable move from the previous control interval.
ut is the manipulated variable target.
vp is the sequence of measured disturbance signals across the prediction horizon.
rp is the sequence of reference signals across the prediction horizon.
When generating code in MATLAB, the
command generates these matrices and returns them in
You can generate code for linear MPC controllers that use a custom QP solver written in C or MATLAB code. The controller calls this solver in place of the built-in QP solver at each control interval.
For an example, see Simulate and Generate Code for MPC Controller with Custom QP Solver. For more information on custom QP solvers, see Custom QP Solver.
For information on using the FORCESPRO solver, see Implement MPC Controllers using Embotech FORCESPRO Solvers.