Modelling And Optimization Of MI Prodiction For Propylene Polymerization Process Based On Artificial Intelligent Optimization

Polypropylene (PP) is playing a more and more important role in industry, military, social life and so on, which makes the quality control of the propylene polymerization process quite crucial, especially the prediction of the melt index of polypropylene product. In this article, the soft sensor prediction of melt index in the propylene polymerization process is discussed. Radial basis function (RBF) neural network is employed to develop the soft sensor prediction model, and then the artificial intelligent algorithms are used to optimize the structure of the RBF model. Several improved algorithms are proposed based on the traditional artificial intelligent algorithms and they improve the model's structure and the performance. These optimized models work quite well on the practical data from real industrial process, thus they offer several options for the melt index prediction application.Major work of this article is as following:(1) The propylene polymerization process is discussed and several variables are chosen to develop the RBF model for melt index prediction, at the same time principal component analysis (PCA) is implied to simplify the input variables of model. The result proves that the PCA improved the model and the model predicts the melt index well.(2) Based on the particle swarm optimization (PSO) and simulated annealing (SA), a novel MPSO_SA algorithm is proposed and used to optimize the structure of the RBF model for melt index prediction mentioned above. The MPSO_SA algorithm combines PSO's strong global search ability and SA's strength in local search, as a result of which it obtains a wonderful optimization capability. Research based on the practical data shows that the combination of PSO and SA is quite effective and the MPSO_SA-RBF model owns a very good prediction performance.(3) An adaptive ant colony optimization (ACO) algorithm is proposed and used to optimize the structure of RBF model for melt index prediction, after which the performance of the A-ACO-RBF model is quite good. Then the randomness of the artificial intelligent algorithms and the over-fitting problem of single RBF neural network are taken into account. Based on the ACO-RBF model, aggregated ACO-RBF model is developed, with two aggregating strategies, one of which is average-strategy and the other one is adaptive-weight-strategy. Since the aggregated ACO-RBF model has overcome the two problems (randomness and over-fitting) to some extent, it obtains a significantly great improvement in the prediction performance, especially the adaptive weighted aggregated ACO-RBF model. These models all work quite well in the melt index prediction and they can serve in the practical applications.