A two-timescale approach to nonlinear model predictive control

Model predictive control (MPC) schemes generate controls by using a model to predict the plant's response to various control strategies. A problem arises when the underlying model is obtained by fitting a general nonlinear function, such as a neural network, to data: an exorbitant amount of data may be required to obtain accurate enough predictions. The authors describe a means of avoiding this problem that involves a simplified plant model which bases its predictions on averages of past control inputs. This model operates on a timescale slower than the rate at which the plant outputs and controls are sampled and updated. Not only does this technique give improved closed-loop performance from the same amount of open-loop data, but it requires far less on-line computation as well. The authors illustrate the usefulness of this two-timescale approach by applying it to a simulated exothermic continuously stirred tank reactor with jacket dynamics.