Quality prediction with neural network techniques for polypropylene production via the spheripol process

Abstract

In the polypropylene (PP) industry, melt index (MI) is the most important quality variable. Different grades of PP have their specific range of MI. Accurate prediction of MI is essential for efficient monitoring and off-grade reduction. Artificial Neural Network (ANN) models are proposed as the technique for MI estimation. It has powerful adaptive capabilities in response to nonlinear behaviour. In this research, ANN models for PP polymerization to predict the MI based on reactor parameters were developed. Three types of ANN models, the single hidden layer ANN (shallow ANN), stacked neural network (SNN) and deep learning are compared. The simulation results show that deep learning can perform better than shallow ANN and SNN by considering the accuracy of the prediction and detection of process fluctuation. All three model have proven that ANN are able to perform non-linear function approximation. Thus, ANN models are effective for supporting MI prediction such as for soft-sensors and process optimization in the polymer industry.