Integrated identification and model predictive control using iterative refinement

Abstract

An iterative refinement approach is presented for designing a high performance control system for an imprecisely known plant. The design procedure involves the integration of identification and model predictive control (MPC) using repeated closed-loop identification tests and successively improving the model (and ultimately, closed-loop performance) with each successive iteration. The method is appealing to industrial practice because real-time closed-loop data can be used directly to enhance the performance of a predictive controller without the need to deactivate the control loop during identification testing. The iterative refinement strategy is "plant-friendly" in that it tries to keep the identification test as short as possible while keeping the plant within operating limits and contraints. Constraint enforcement (on controlled, associated and manipulated variables) is naturally implemented through the use of MPC. The application of the method is demonstrated on a highly nonlinear diabatic (i.e., nonadiabatic) continuous stirred tank reactor problem.