Multivariate Long Sequence Time Series Forecasting Based on Robust Spatiotemporal Attention

Multivariate long sequence time series forecasting is critical for various applications like power consumption planning and weather forecasting. Effective forecasting in this context requires capturing long-term coupling correlation between sequences across different time windows. Existing methods focus solely on temporal feature dependencies, neglecting the combined temporal and spatial features. Furthermore, they overlook outlier-driven performance degradation. In this study, we propose MSAL, a robust prediction method for long sequence time series that breaks the spatiotemporal boundary of long sequence and eliminates the influence of outlier bias in the data on forecasting accuracy. First, a convolutional mapping module in MSAL is used to capture the temporal feature dependencies of sequences to efficiently handle extremely long input sequences. Second, MSAL utilizes a spatial convolutional attention pooling structure to capture the spatial feature dependencies between variables. Finally, by introducing robust continuous parameters for adaptive loss calculation, the outlier data is excluded. Extensive experiments conducted based on large-scale datasets demonstrate that the MSAL method achieves better accuracy and robustness than state-of-the-art baseline methods on long sequence time series forecasting problems with outliers.