A Sequence-to-Sequence BiLSTM Model with Dual Attention Mechanism for Multi-step Prediction of Industrial Processes

Industrial processes are highly dynamical and complex. To more accurately capture the complex dynamic relationships for quality prediction, this paper proposes a sequence-to-sequence bidirectional long short-term memory network with spatiotemporal attention mechanism (Seq2Seq-STA-BiLSTM) for multi-step prediction of industrial processes. This model employs an encoder-decoder architecture using the Seq2Seq learning approach. The encoder applies a spatial attention mechanism to assign differentiated weights to auxiliary process variables, thereby enhancing the feature extraction precision. Meanwhile, the decoder emphasizes critical temporal dependencies through a temporal attention mechanism integrated with the BiLSTM architecture to effectively handle mutual dependencies in the sequence. These effective techniques employed in the proposed Seq2Seq-STA-BiLSTM model jointly improves the accuracy of multi-step prediction. An industrial CO<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</inf> absorber tower is studied to verify the effectiveness and reliability of the proposed soft sensor.