Estimator Based Inferential Control of an Ideal Quaternary Endothermic Reactive Distillation with Feed-Forward and Recurrent Neural Networks

Abstract

Abstract A feed-forward neural network (FNN) and a layered recurrent neural network (LRNN) based two composition estimators, respectively, are designed for the purpose of tight product purity control for an ideal, quaternary, hypothetical, kinetically controlled, reactive distillation (RD) column. The output variables of the considered control structure i.e. the compositions, are estimated using the chosen tray temperatures as inputs to the estimators. The performances of the estimators in the control of the column for the servo, regulatory, feed impurity disturbances and catalyst deactivation are studied. The estimator based control is found to be effective for the on-spec product purity control. One-to-one relation between the number of tray temperature measurements and their sensitivity to the accuracy of estimation is observed. Overall, the performance of LRNN is found to be superior over the FNN for the throughput manipulations tested for the more number  of inputs to estimator.