Abstract:
To present a neural network forward model to predict a concentration and temperature profiles in a batch reactor for a MMA production, and a neural network inverse model to predict a jacket temperature set point of a heating/cooling system. The neural network forward and inverse models have been developed based on the Lenvenberg-Marquardt training algorithm. An obtained optimal neural network structure for the forward model has been employed to predict state variables over a predictive horizon within a model predictive control (MPC) algorithm for searching optimal control actions via successive quadratic programming (SQP). To control the temperature in the batch reactor, the neural network based control approaches studied in this work consisting of a neural network direct inverse control (NNDIC) and a neural network based model predictive control (NNMPC) have been formulated. In addition, a dynamic optimization approach has been applied to find out an optimal operating temperature to achieve maximizing the MMA product at specified final time. An obtained optimal temperature is then applied as a set point for the controller design. Robustness tests of the proposed controllers have been studied with respect to the changes in operating parameters. Simulation results have indicated that the NNMPC controller is more robust than the PID and NNDIC controllers and the NNDIC controller is more robust than the PID controller. Therefore, the NNMPC controller gives the best control results among the PID and NNDIC controllers in the nominal and plant/model mismatch cases.