ABSTRACT
In this paper a new neural control architecture for chemical processes is proposed, which consists of a PI controller for initial stabilization and a radial basis function(RBF) network for linearization of nonlinear plants. The control input applied to the process is summation of outputs of the PI controller and the RBF network. First, a reference linear model, which defines relationship between the PI controller output and the process output, is determined from the analysis of the past operation data. Initially the PI controller operates the process alone. The RBF network is gradually trained to minimize the difference between outputs of the plant and the linear model. As the training of the RBF network goes on, the dynamics of the nonlinear plant added by the RBF network converges to that of the linear reference model. Then, the overall control problem becomes linear, and the linearized system can be easily controlled by the PI controller with closed-loop architecture, of which parameters can be determined by simple pole-zero assignments. Provided the initial PI controller stabilizes the plant, it remains stable throughout the whole training phase. Computer simulation shows that the proposed control architecture is very effective in controlling nonlinear chemical processes such as continuous stirred tank reactor and pH process.
