Optimization-based tuning strategies for linear and non-linear model predictive controllers.

摘要

The main goal of this research is the development and analysis of techniques for optimization-based tuning of Model Predictive Controllers (MPC). An off-line tuning procedure is developed in this dissertation and tested on two MPC formulations. The algorithm is carried out using CONSOLE, a multi-objective optimization software. The off-line tuning procedure was first tested on MPC, developed in this dissertation, that utilizes nonlinear models and Kalman filtering. Three illustrating examples were used and they revealed certain weak points that necessitated further understanding of the nature and properties of the tuning algorithm. For this reason, the first-order derivatives of the closed-loop response with respect to the MPC tuning parameters is developed. It is found that analytical expressions based on iterative equations can be obtained for the case of MPC algorithms that utilize linear models. The analytical expressions are formulated as recursive equations.From the point of view of process control practice, an on-line implementation of the MPC tuning procedure is more appealing for industrial use. Utilizing the sensitivity expressions and utilizing the first order (linear) approximation of the relationship between the closed-loop response and the tuning parameters, the on-line tuning procedure is cast as a constrained least squares problem. The effectiveness of the proposed algorithm is tested on complex nonlinear model of the Fluid Catalytic Cracking Unit (FCCU). The illustrative example demonstrated successful implementation despite the presence of model-plant mismatch.Part of this research is to ensure that the tuning technique not only forces the feedback response to satisfy certain requirements but also produces a nominally stable response. This issue is approached by requiring all the eigenvalues of the closed-loop transfer function of the nominal plant to lie inside the unit disk for any value of the tuning parameters. In the off-line tuning formulation the nominal stability condition is guaranteed by imposing additional hard constraint in CONSOLE. In the on-line tuning formulation the stability condition is used just as indicator for nominal stability of the adapted tuning parameters. (Abstract shortened by UMI.)