Closed-loop Identification for Model Predictive Control: Direct Method

Model predictive control (MPC) is a widely used control scheme that handles constraints directly. In practice, the initial performance of MPC is usually satisfactory after a careful setup stage. However, over time, physical changes in the plant may invalidate the predictive model used in MPC and control performance degrades. Thus at least a model update is needed to restore the plant performance. Since the initial commissioning of MPC can be long and costly, serious attention should be given to closed-loop identification for MPC. This paper shows that if the MPC exhibited complex enough behavior during normal operation then one can obtain good model update based solely on the informative operation data. If needed, one can design an experiment aimed at increasing the complexity of MPC in order to avoid undesirable actuator saturations. The explicit piecewise affine solution of MPC is used to analyze both questions.