Spatiotemporal Multivariate Weather Prediction Network Based on CNN-Transformer

Weather prediction is of great significance for human daily production activities, global extreme climate prediction, and environmental protection of the Earth. However, the existing data-based weather prediction methods cannot adequately capture the spatial and temporal evolution characteristics of the target region, which makes it difficult for the existing methods to meet practical application requirements in terms of efficiency and accuracy. Changes in weather involve both strongly correlated spatial and temporal continuation relationships, and at the same time, the variables interact with each other, so capturing the dynamic correlations among space, time, and variables is particularly important for accurate weather prediction. Therefore, we designed a spatiotemporal coupled prediction network based on convolution and Transformer for weather prediction from the perspective of multivariate spatiotemporal fields. First, we designed a spatial attention encoder-decoder to comprehensively explore spatial representations for extracting and reconstructing spatial features. Then, we designed a multi-scale spatiotemporal evolution module to obtain the spatiotemporal evolution patterns of weather using inter- and intra-frame computations. After that, in order to ensure that the model has better prediction ability for global and local hotspot areas, we designed a composite loss function based on MSE and SSIM to focus on the global and structural distribution of weather to achieve more accurate multivariate weather prediction. Finally, we demonstrated the excellent effect of STWPM in multivariate spatiotemporal field weather prediction by comprehensively evaluating the proposed algorithm with classical algorithms on the ERA5 dataset in a global region.