Data-Driven Regional Weather Forecasting Guided by Global Context

Weather forecasting plays a critical role in many aspects of modern society. Compared to global weather forecasting, regional weather forecasting (RWF) can achieve higher resolution and more detailed analyses for the target region. Since global weather forms an interconnected system, incorporating the global weather state is crucial for accurate and reliable RWF. However, existing RWF methods suffer from two issues: (1) They tend to overlook the influence of global weather patterns on regional weather by limiting the modeling scope to a confined geographic area. (2) They usually require extra regional boundary processing to impose lateral boundary forcing for yielding more reasonable and accurate regional forecasts. To address above issues, we propose a global context guided data-driven RWF method that can realize highly effective RWF. First, to our best knowledge, we propose to introduce entire global weather state into RWF modeling for the first time, so as to fully consider the influence of external weather. Second, we design a dual-branch architecture that directly accepts both regional and global weather data as inputs for learning, which eliminates the additional processing on regional boundaries. Finally, we develop a multi-stage regional-global attention mechanism, which fuses regional and global weather features to yield global context, and use it to guide RWF. Experimental results show that the proposed method can achieve state-of-the-art performance and demonstrate the significant improvement in regional forecasting by considering global weather state.