Model predictive control uses a model to forecast a system, picks the best control moves over a future horizon, applies the first, then repeats.
Model predictive control (MPC) is a control method that uses a model of the system to plan ahead. At each time step it forecasts how the system will respond over a future window, called the horizon, chooses the sequence of control actions that best meets its goals, applies only the first action, and then repeats the whole process at the next step with fresh measurements. This is the receding horizon idea: the plan is always redone, so new information is used as soon as it arrives. The goals are written as a cost to minimise, for example energy cost plus a penalty for overflow volume or for water levels away from target. Constraints are written in directly: pump limits, gate ranges, minimum tank levels, maximum pressures. This makes MPC well suited to systems with many linked parts and hard limits. A water example is pump scheduling in a supply network. The controller uses a demand forecast and a network model to plan pump runs over the next 24 hours, so that tanks stay within safe levels while more pumping moves to cheaper tariff hours. Every 15 minutes or every hour it replans using the latest tank levels. In sewers, MPC can use rainfall forecasts to empty storage ahead of a storm. In irrigation canals it is used to hold water levels at gates. MPC needs a model fast enough to run many times per step, so simplified or surrogate models often replace full hydraulic models. It also needs forecasts of disturbances (demand, inflow, rain) and reliable live measurements. Its limits follow from this. If the model or forecast is wrong, the plan is wrong, though replanning at each step corrects part of the error. Choosing the cost weights takes judgement, and large nonlinear problems can be slow to solve within the time step.