LONG-TERM PREDICTIONS USING RECURRENT NEU RAL NETWORKS FOR STATE CHANGES IN POLYMERIZATION REACTORS

Long-term predicting methods using neural networks (NN) are discussed for state changes in polymerization reactors. The temperature at the o utlet of a continuous bulk polystyrene polymerization reactor is the p resent target of predictions using a layered neural network and some r ecurrent neural networks (RNN). Some structural problems in a general RNN are indicated, and two improvements (H -RNN with additional proces sing in hidden layer units, M-RNN with an additional calculating modul e of a hidden layer in a layered NN) are proposed on data processing a nd arranging by hidden layer units in RNN. As to the above networks, e ach predictive performance can be comparatively evaluated using mean s quare error. The predictive performance of H-RNN and M-RNN is superior to that of a layered NN in the initial stage of predictions, or in th e state change with maximum or minimum points. In particular, long -te rm predictive performance is widely satisfied by M-RNN where the combi ned structure of RNN with hidden units of layered NN is built to impro ve accuracy in initial stage of predictions.