Skip to content
Back to search results

Model Predictive Control

Introduced by
Jacques Richalet
Year
1978
Type
algorithm
Family
Optimal Control
Font size

What does it do?

Model Predictive Control (MPC) is an advanced control strategy that uses an explicit process model to predict future system behavior over a finite horizon and solves an optimization problem at each control step.

First formalized by Richalet et al. in 1978, MPC has become the dominant approach in process control industries, from chemical plants to autonomous vehicles, because it naturally handles constraints and can optimize multiple objectives simultaneously.

The intuition behind the method

The core idea of MPC is to use a mathematical model of the system to simulate multiple possible future scenarios and compute the best control actions that will keep the system close to desired setpoints while respecting operational constraints. At each sampling time, the algorithm predicts the system output over a future horizon, solves an optimization problem to find the control input sequence that minimizes a cost function, applies only the first control action, and then repeats the process with new measurements. This feedback mechanism—prediction, optimization, action, measurement, repeat—makes MPC robust to disturbances and model uncertainty while explicitly accounting for physical and operational limits.

History

Richalet introduced the concept in 1978 as Heuristic Model Predictive Control (HMPC) for industrial processes. Garcia et al. provided a comprehensive theoretical framework in 1989, establishing connections to dynamic programming and stability theory. Mayne et al. (2000) formalized constrained MPC and stability conditions. The field has since expanded to nonlinear MPC (NMPC), explicit MPC (pre-computed control laws), and fast optimization algorithms that enable real-time implementation in embedded systems.

When to use it

Use MPC when you need to handle constraints explicitly (e.g., valve limits, safety bounds), manage multiple correlated variables (multivariable control), or when disturbances and non-linearities are significant. It is most effective for systems with slow to moderate dynamics where computation time allows for real-time optimization. Typical applications include chemical process control, HVAC systems, power plants, and autonomous driving. Avoid MPC if your system demands microsecond control rates or if a high-fidelity process model is unavailable.

Limitations and considerations

  • Computationally expensive; solving the optimization problem at each step may be infeasible for very fast systems.
  • Requires an accurate process model; poor model accuracy degrades performance significantly.
  • Nonlinear MPC is non-convex and may get stuck in local minima.
  • Stability and feasibility guarantees are non-trivial, especially for constrained problems.

Source conditions do not replace independent scientific review or validation by an analysis engine.

Strengths

  • Naturally handles input and state constraints through explicit optimization.
  • Provides multivariable control, coordinating multiple inputs and outputs automatically.
  • Robust to disturbances and model mismatch because of the receding horizon feedback loop.
  • Allows incorporation of future reference trajectory or set-point changes.
  • Easy to tune (cost function weights) compared to traditional PID tuning.

Applications

  • Chemical process control: Optimizing temperature, pressure, and flow rates in reactors and distillation columns while respecting safety constraints.
  • Autonomous vehicles: Managing acceleration, steering, and deceleration to follow a desired trajectory while avoiding obstacles and respecting road constraints.
  • Power grid management: Balancing supply and demand, controlling generation scheduling, and voltage stability over a rolling forecast horizon.
  • Building HVAC systems: Coordinating heating, cooling, and ventilation to maintain comfort setpoints while minimizing energy consumption.
  • Robotics path planning: Computing smooth, feasible trajectories that respect actuator limits and collision constraints.

How it works

How it works

1

Use the system model and current state to predict future outputs over a prediction horizon (typically 10-100 steps ahead).

2

Formulate and solve a constrained optimization problem to find the optimal input sequence that minimizes tracking error and control effort while respecting bounds on inputs and states.

3

Implement only the first control input from the optimized sequence to the process.

4

Measure the new system state, update the prediction, and repeat the entire cycle at the next sampling time.

Common pitfalls

  • Setting the prediction horizon too short leads to myopic, suboptimal control; too long increases computation.
  • Failing to tune the cost function (e.g., weighting between error and control effort) leads to sluggish or aggressive responses.
  • Using an oversimplified or inaccurate model that does not capture system dynamics results in poor real-world performance.

Frequently asked questions

What is the receding horizon principle and why is it important?

The receding horizon principle means solving an optimization problem over a finite future window, applying only the first control action, and then shifting the window forward. This mechanism creates a feedback loop that corrects for errors and disturbances at each step, making MPC robust even when the model is not perfect.

How do I choose the prediction horizon length?

A rule of thumb is to set the prediction horizon to 3-5 times the settling time of the system's dominant time constant. Too short leads to myopic behavior; too long increases computation without additional benefit and may cause infeasibility. Empirical tuning or open-loop response analysis helps.

Can MPC handle nonlinear systems?

Yes, Nonlinear MPC (NMPC) can handle nonlinear system models. However, the optimization becomes non-convex, requiring iterative solvers that are slower and may converge to local minima. Linearization around a nominal trajectory or approximation methods are often used to reduce computation time.

What software tools are available for implementing MPC?

Popular tools include MATLAB Model Predictive Control Toolbox, Simulink, open-source packages like CasADi (for optimization), and specialized commercial platforms like EMSO and gProms. For embedded systems, libraries like ACADO and fast quadratic programming solvers (e.g., qpOASES) enable real-time implementation.

Related methods

Sources

  1. Richalet, J., Rault, A., Testud, J., & Papon, J. (1978). Model predictive heuristic control. Automatica, 14(5), 413-428.

    Open DOI reference
    Copied.
  2. Garcia, C. E., Prett, D. M., & Morari, M. (1989). Model predictive control: Theory and practice. Automatica, 25(3), 335-348.

    Open DOI reference
    Copied.
  3. Mayne, D. Q., Rawlings, J. B., Rao, C. V., & Scokaert, P. O. (2000). Constrained model predictive control: Stability and optimality. Automatica, 36(6), 789-814.

    Open DOI reference
    Copied.