Improving multi-variable weather forecasting with spatiotemporal feature interaction and large-kernel modeling

Accurate and timely weather forecasting is vital for sectors such as agriculture, aviation, and disaster response. While recent advances in deep learning have enabled data-driven forecasting models, most existing methods rely on autoregressive prediction of single variables, overlooking complex interdependencies among multiple meteorological factors. To address this limitation, we propose a novel multi-variable coupled forecasting model that captures both nonlinear interactions and spatiotemporal dynamics. Specifically, we introduce a Branch-Crossing Attention Fusion strategy within a Multi-factor Fusion Network to enhance feature integration. To better model long-range dependencies and differentiate regional weather patterns, we design a Novel Large Kernel Convolution module as a substitute for traditional convolutions in recurrent networks. Built upon these innovations, we develop NLKRNN, a spatiotemporal prediction network based on the NLK-LSTM unit. Experiments on a custom meteorological dataset covering temperature, dew point, and U/V wind components demonstrate that our approach achieves consistently improved performance compared with representative spatiotemporal forecasting baselines in multi-variable weather forecasting.