Leveraging Atmospheric Information Across Scales for Local Temperature Forecasting Using a Novel Multi‐Scale Perception Network

Regional weather forecasting systems typically depend on boundary conditions from global models to represent large‐scale atmospheric processes, while such coupling increases complexity and hinders end‐to‐end optimization for specific target locations. Here, we propose the Multi‐scale Perception Network (MPN), a unified deep learning model that explicitly processes atmospheric information across multiple spatial scales for air temperature ( T a ) forecasting. Using Hong Kong as a test case, MPN outperforms state‐of‐the‐art transformer‐based deep learning models and an operational physics‐based model, achieving up to 16% and 21% reductions in root mean squared error on 1–6 hr forecast horizons, respectively. Scale‐sensitivity experiments demonstrate that predictor domains of about 2,000 km generally provide sufficient atmospheric context for 1‐day forecasts, with improvements concentrated primarily in cold surge events, where broader‐scale information captures approaching frontal systems. Finally, our evaluation highlights distinct lead‐time‐dependent performance profiles and identifies forecast barriers linked to diurnal phase transitions, pointing to fundamental challenges and opportunities for improving short‐range regional forecasts.