Application of feed forward neural network for fouling thickness estimation in low density polyethylene tubular reactor

Abstract

In this research, the Feed Forward Neural Network (FFNN) development using Levenberg-Marquardt training method for estimation of fouling thickness layer in Low density polyethylene tubular (LDPE) reactor is performed. The highly exothermic nature of the LDPE polymerization process and the heating-cooling prerequisite in the tubular reactor can create fouling problems. Thus, the FFNN modeling technique has been applied to predict the fouling thickness of LDPE in the cooling zone of the polymerization tubular reactor. The fouling formation on the inner wall in the cooling zone is rarely reported in previous related LDPE studies since it is difficult to measure. The fouling thickness layer might increase to an unsafe level if it is not monitored. In addition, the fouling layer has a low thermal conductivity which increases the resistance to heat transfer and reduces the effectiveness of heat exchangers. In order to develop the FFNN model, a set of fouling data is generated from the combination of LDPE tubular reactor model simulation using Aspen Dynamic and fouling build-up equation. In order to improve the FFNN model input selection, the Pearson correlation coefficient (PCC) method is implemented. Based on PCC analysis, polymer density, heat transfer coefficient, and reactor temperature in the respective cooling zone were considered as inputs for the developed FFNN model. The variation of the hidden node was evaluated to assess the accuracy of the model based on the coefficient of determination (R2) value. Based on the FFNN modeling results, the highest accuracy to predict fouling thickness was achieved using eight hidden nodes with 0.983 R2.