Improved Performance of Robust MPC with Feedback Model Uncertainty

Robust model-predictive controllers use an estimate of model uncertainty in the on-line controller calculation and can be overly conservative for some uncertainty descriptions. This paper discusses the various causes of conservative control with particular emphasis given to the concept of 'closed-loop' probabilistic predictions. A multi-input-multi-output MPC is proposed in which an off-line, non-convex calculation is used to characterize the closed-loop uncertainty a priori. This uncertainty information is incorporated into a convex, quadratic program resulting in a MPC formulation that can be efficiently solved on-line. A distillation column case study demonstrates the benefits of the proposed robust MPC.