Modelling And Optimization Of MI Prediction For Propylene Polymerization Process Based On Dynamic Fuzzy Neural Networks

Polypropylene (PP) is a thermoplastic polymer of propylene and one of the five most widely used plastics, which is applied in a number of applications including plastic containers, stationery and packaging. Melt index (MI) is an improtant index to determine the quality of PP, so the soft sensor prediction of MI is quite critical. In this paper, the soft sensor modeling of MI by dynamic fuzzy neural networks (D-FNN) is discussed. Then the parameters of D-FNN are optimized by artificial intelligent algorithms. Two improved intelligent algorithms are presented to optimize the parameters of D-FNN model. These models established work successfully on the actual data from the practical industrial plant, thus offering more options when dealing with the melt index prediction problems.The main work and contributions of this paper are as follows:(1) A D-FNN model for melt index prediction is proposed. The related variables of the propylene polymerization process are selected to establish the D-FNN model for MI prediction, and the input variables are prepocessed by principal component analysis (PCA) to simplify the model. The prediction results based on the practical data from a real industrial process prove the effectiveness of the proposed D-FNN model.(2) An AACO-D-FNN model for melt index prediction is proposed further. Considering the shortcoming of Ant Colony Optimization (ACO) algorithm which cannot find the optimal solution accurately, an adaptive ACO (AACO) is developed to adjust the step size and direction adaptively, thus improving the local search of ACO. The AACO-D-FNN model is validated with the practical data from a real industrial process.(3) Considering the advantages of Particle Swarm Optimization (PSO), and in order to avoid the drawback of standard PSO, whose particles are easy to be trapped into local optimization in the iterative process, a chaotic GA/PSO algorithm is proposed. The parameters of D-FNN are optimized with the chaotic GA/PSO algorithm, and the CHA-D-FNN model is presented. The prediction results based on the practical data from a real industrial process prove the effectiveness of the CHA-D-FNN model.