A Bidirectional Attention-Based LSTM Model for Multi-Step Ahead Prediction in Dynamic Soft Sensors

In the field of dynamic soft sensor developments, multi-step ahead prediction of key quality variables is becoming increasingly important, which can provide a long-term data changing trends in the future time period. However, multi-step ahead prediction is currently facing various issues such as data uncertainty and unstable forecasts. In order to improve these issues, this paper proposes a bidirectional spatiotemporal attention-based LSTM network model (Bi-STA-LSTM), in which spatiotemporal attention extracts the relevant features from historical time series, and the bidirectional structure captures more historical and future trends. It can continue to provide useful information when the prediction step increases. By comparing different methods on the debutanizer dataset, the superiority of the Bi-STA-LSTM is verified.