A Temporal-Spatial Attention Encoder-Decoder Network for Multivariate Time Series Forecasting and Its Application in Hot Blast Stove Production

Deep learning has been widely used in multivariate time series prediction, which can automatically learn hidden information in historical process data. Due to the change of external conditions or the influence of unobservable variables, the sequence pattern of time series may change with time. There are also differences between local sequential patterns and global sequential patterns. In this paper, a temporal-spatial self-attention encoding and decoding (TSSA-ED) network is proposed. Firstly, the original data is divided according to different scale factors, and the sequence information in different scale data is extracted by TSSA-ED network. TSSA-ED network consists of temporal attention module and spatial attention module, which are used to capture temporal dependence and spatial dependence respectively. The input forms of the two modules are different, one is all time steps of single variable, and the other is all variables of the same time step. Finally, the model is applied to predict the dome temperature and exhaust gas temperature in the production process of hot blast stove, which proves the effectiveness of the proposed method. Compared with other traditional models, it is proved that the method proposed in this paper has higher prediction accuracy.