Adaptive Explicit Model Predictive Controller For coupled Continuous Stirred Tank Reactor Using Least Squares Support Vector Machines

Abstract

Classification based adaptive control of nonlinear processes is a practical solution as compared to nonlinear controller design. In this paper, a highly nonlinear two tank coupled continuous stirred tank reactor (C-CSTR) controlled based on a support vector machine based classification (SVM-C) is discussed. The input- output transient dynamics of this C-CSTR is modeled into empirical first order plus delay time (FOPDT) models in different regions. These local dynamics are used in design of Linear Quadratic Regulator (LQR). The performance of such SVM classified LQRs’ is compared against a more elaborate neural network (NN) based nonlinear MPC (NN-MPC). Data to train both SVM-C and NN-MPC are obtained from the first principle model of C-CSTR implemented in MATLAB. Closed loop control of C-CSTR by both controllers is analyzed based on visualization as well as quantitative results. It is observed that the LQR with SVM-C performs close enough to the NN-MPC with a much lower footprint in terms of computation.