Model predictive control (MPC) uses a dynamic model to predict the future output over a horizon, then chooses a sequence of future moves that minimises an objective while respecting limits.
$$\min_{\Delta u}\ \sum_{k=1}^{P}\left\|y_{sp} - \hat y(k)\right\|^2_Q + \sum_{k=0}^{M-1}\left\|\Delta u(k)\right\|^2_R$$
subject to $u_{min} \le u \le u_{max}$ and $y_{min} \le \hat y \le y_{max}$.
Receding horizon
- Measure the current output and correct the prediction for model error.
- Solve the optimisation over the prediction horizon $P$ with $M$ future moves.
- Apply only the first move.
- At the next sample, repeat with new measurements.
Why industry uses it
- Handles constraints directly, so plants can run closer to limits.
- Handles many interacting inputs and outputs in one controller.
- Uses feedforward from measured disturbances naturally.
MPC is standard in refineries, petrochemical plants and power systems, usually as a layer above the PID loops: MPC sets the set-points and PID loops move the valves.