Extracting Spatio-Temporal Coupling Feature of Patches for Long-Term Multivariate Time Series Forecasting

The patching and variable channel-independent mechanism has been used in the long-term multivariate time series (MTS) forecasting models to capture local semantic information and learn different attention patterns. However, it cannot exploit the spatial dependencies between the patches and the correlation discrepancy of different temporal positions. To address this issue, a long-term MTS forecasting model based on spatio-temporal coupling feature of patches (LMTSF-SCF) is proposed in this study. Firstly, we design a temporal position channel-independent structure which uses squeeze-and-excitation (SE) network to extract spatial dependent feature of the patches and learn the correlation difference between the temporal positions. Then the variable channel-independent structure using Transformer encoder is leveraged to extract the temporal dependencies between the spatial patches features. Finally, the spatio-temporal coupling patches feature of the MTS is mapped into long-term forecasting values using a fully connected network. The experiment demonstrates that the proposed model achieves significant performance over that of the state-of-the-art baseline model DLinear, FEDformer, Autoformer, Informer, and PatchTST.