Temperature prediction of aluminum reduction cell based on integration of dual attention LSTM for non-stationary sub-sequence and ARMA for stationary sub-sequences

Abstract

In aluminum electrolysis process, cell temperature is an important index to reflect the current efficiency of electrolytic cell. Maintaining cell temperature in an appropriate range can improve current efficiency and economic benefits. Because aluminum electrolysis process is in a complex environment with high temperature, strong corrosivity, multivariable coupling and nonlinearity, cell temperature measuring instrument and equipment have the shortcomings of short service life and high cost. Therefore, this paper develops a cell temperature prediction model based on the integration of dual attention long short-term memory (DA-LSTM) and autoregressive moving average (ARMA). Firstly, the cell temperature series is decomposed into non-stationary sub-sequence and stationary sub-sequences by wavelet transform. The DA-LSTM is proposed to approximate non-stationary sub-sequence, which introduces a two-stage attention mechanism including feature attention mechanism and temporal attention mechanism. ARMA is employed to predict the stationary sub-sequences. Then, the integration model for the cell temperature prediction is reconstructed through multiple linear regression of DA-LSTM and ARMA. The integration model can overcome the problems that the prediction accuracy of a single model is low and the hysteresis phenomenon due to the mixed multi-information of aluminum electrolysis time series data. Experiments in a real-world aluminum electrolysis production process are conducted to demonstrate the effectiveness of the proposed model.