| Model predictive control (MPC) is a widely used advanced process control technique in the process industry. According to the internal model principle, the internal model of MPC has to include both exact plant and disturbance models to be optimal. However, in practice, the MPC usually assumes a step-like disturbance or a fixed disturbance model. As a result, the MPC will be suboptimal when disturbance changes slowly. Moreover, it lacks a tool assessing the optimality of control performance in terms of the MPC model.;In this dissertation, a new MPC disturbance adaptation method is presented. Starting from a single-input-single-output (SISO) semiconductor manufacturing process, we replaced the conventional run-to-run controller by an adaptive EWMA controller. It is shown that the plant model mismatch can be compensated by adapting the disturbance model. Analysis has been done to show that the adaptive controller is stable and converges to the optimal controller.;The proposed method is then extended to multi-input-multi-output (MIMO) systems. For the ease of practical applications, the integrated moving average (IMA) model with order (1,1) is recommended. The equivalence between the IMA(1,1) parameter and the prediction error filter constant in commercial MPC has been established. Implementation of disturbance adaptation is explained.;Another disturbance modeling tool is presented. It focuses on the closed-loop identification of a nonparametric disturbance model. The method incorporates the plant model information during the conversion from observer Markov parameters to system Markov parameters.;A new control performance assessment method evaluating MPC model quality is then presented. Feedback invariant principle is introduced, based on which a method estimating disturbance innovations is given. A model quality index is developed as the performance benchmark, which compares prediction errors with disturbance innovations. It is shown that the model quality index related to the MPC performance index.;Most industrial processes are optimized by a linear programming (LP) problem on top of the MPC. A new control performance monitoring method for cascaded LP-MPC system is developed. The block-lower-triangular interactor matrix is introduced to form a new method that is able to determine the performance benchmark based on controlled variable (CV) priorities coming from the LP results. |
