Abstract:In order to consider the correlation between process variables and quality variables, and solve the problem that the timing and dynamic characteristics of batch process lead to low prediction accuracy, a model based on Improved Spatial and Temporal Attention Long short-term Memory (ISTA-LSTM) is proposed to predict intermittent process quality. First, three-dimensional batch process data are unfolded into two-dimensional matrix according to batch-variable method. The two-dimensional data are normalized by min-max method. Secondly, partial least squares (PLS) method is used to reduce the dimension of the original data and extract the features of the data. Based on temporal attention and spatial attention, a double-layer LSTM network with encoder–decoder structure is constructed. In order to retain all the necessary input information, the neural network parameters are learned adaptively by the attention mechanism. Pay attention to the importance of each process variable to the quality variable and assign the corresponding attention value. Then, grid search algorithm with cross verification is used to optimize the prediction model by hyperparameters, and the prediction model is established. Finally, the proposed algorithm model is verified on the penicillin fermentation simulation experimental platform. The experimental results show that the model proposed in this paper is feasible and effective for batch process quality prediction.