Nonlinear Model Predictive Control Based On Least Squares Support Vector Machine

Linear model predictive control (LMPC) has become one of the most successful advanced control algorithms in the chemical and petrochemical industry recently and the LMPC theory can be considered quite mature. Nevertheless, many manufacturing processes are inherently nonlinear, time-variant and uncertainty. Under these conditions, the control performance of LMPC is often not satisfactory. Thus, research on nonlinear predictive control has become an important issue in control field.In recent years, least squares support vector machine (LSSVM) based on statistical learning theory has been successfully applied in pattern recognition, system identification, etc. Therefore, nonlinear predictive control algorithms based on LSSVM for a class of complex plants with nonlinear behavior are studied in this thesis. The main research woks are as follows:1. A new nonlinear predictive control algorithm based on LSSVM and adaptive quantum particle swarm optimization (AQPSO) is proposed. LSSVM is used for modeling the nonlinear process, while AQPSO algorithm is used to derive the control law, which avoids calculating the complicated gradient information and the inverse matrix. The simulation results in a pH neutralization process demonstrate the feasibility and effectiveness of the proposed algorithm.2. A new nonlinear predictive control algorithm based on local LSSVM is proposed. Considering the generalization performance of the global learning based modeling methods is not satisfactory, the idea of local modeling is combined with LSSVM to construct a local LSSVM model. Further, this local LSSVM model is used as the predictive model of nonlinear predictive control. Considering the real-time problem, a method for reducing the complexity of modeling is proposed. The simulation results of a continuous stirred-tank reactor (CSTR) demonstrate the feasibility and effectiveness of this algorithm.3. The experiment of level control system of oil and water separator is done. The two control algorithms proposed by this thesis are applied to the equipment and the results of the application are satisfactory.