Spatio-temporal attention-based hybrid deep network for time series prediction of industrial process

Industrial time series involves a large amount of production process information, which effectively reflects the production status of the industrial process. To better understand characteristics and patterns of changes in production conditions, it is crucial to analyze and predict industrial time series data. Given the involvement of numerous parameters and complex physical-chemical reactions in industrial processes, attaining precise predictive performance utilizing a single model remains a formidable challenge. In this paper, we propose a novel hybrid deep learning prediction method based on spatio-temporal attention and temporal convolution network. The proposed method aims to handle the multivariate coupling characteristics and dynamic nonlinear features in industrial time series through different model structures for accurate prediction. In this method, historical data are first segmented into multiple consecutive inputs along the temporal dimension, which are then used as inputs to the subsequent attention mechanism module. To realize the mapping from points to series in the temporal dimension, the segmented input is processed using both the adaptive attention mechanism and one-dimensional convolution. Then the spatio-temporal coupling features are further explored through the spatio-temporal attention model. In addition, to extract dynamic nonlinear features from historical data, a parallel temporal convolutional network with temporal pattern attention is utilized. In order to evaluate the prediction performance of the proposed model, we use two different real-world industrial time series datasets for comprehensive evaluation. The experimental results demonstrate the effectiveness and accuracy of the proposed method. Code is available at https://github.com/TensorPulse/MACnet.