PREDICTION OF TECHNOLOGICAL SYSTEMS USING THE MECHANISM OF ATTENTION IN NEURAL NETWORKS

Abstract

A method is proposed for predicting the variables of a cyber-physical system that implements the technological process of phosphorus production; the variables are presented as a multidimensional time series. The method is based on the use of a deep neural recurrent network with an autoencoder structure, to which an attention mechanism is added. In it the information about the intermediate internal states of the encoder is available to the decoder and is used by it to form an output sequence of predictive values of process variables. The results of a model experiment in the MatLAB environment are presented, which showed a higher prediction accuracy of a neural network with the attention mechanism compared to a neural network without its use