MAI-Unet: A Multi-Scale Attention Interactive Network for Multivariate Long-Term Time Series Forecasting

In recent years, Transformer models based on multiscale fusion have made significant progress in the field of multivariate long-term time series forecasting(MLTSF). Compared to traditional Transformer models, cross-scale models can better model long-term dependencies and extract multi-level features. However, they focus more on time dependencies at different scales and rarely consider the interrelationships between time series variables. Based on this issue, we propose a Mixed Pooling Spatio-Temporal Attention (MPSTA) module, aimed at efficiently extracting time dependencies and variable correlations in sequences. Moreover, this module can more accurately calculate attention weights while maintaining the relative spatial and temporal positions of multiple variables. Additionally, we propose a novel multi-scale attention interaction Unet network (MAIUnet), which can establish long-term and short-term dependencies in time series through cross-scale learning, addressing the traditional multi-scale model’s inability to fully utilize adjacent scale information. Through extensive experiments on eight public datasets, compared to the best Transformer-based model, MAIUnet achieved 31 optimal results and 25 second-best results in all test cases. In particular, on the Weather dataset, we achieved the best performance, with MSE decreasing by $\mathbf{1 0. 0 6 \%}$ and MAE decreasing by $\mathbf{5. 6 6 \%}$.