STCast: Adaptive Boundary Alignment for Global and Regional Weather Forecasting

TL;DR

STCast tackles the challenge of accurate regional weather forecasting by learning adaptive global–regional boundaries and temporally specialized routing. It introduces Spatial-Aligned Attention (SAA) to initialize and refine global–regional distributions and a Temporal Mixture-of-Experts (TMoE) that routes monthly inputs to dedicated experts via a discrete Gaussian prior. The approach delivers state-of-the-art performance across low-resolution global, high-resolution regional, extreme-event, and ensemble forecasting tasks on ERA5 data, with strong long-horizon skill and robust typhoon-track predictions. By explicitly modeling spatial boundary evolution and temporal variability, STCast offers a scalable, Earth-aware framework for integrated weather forecasting with practical implications for accuracy and reliability.

Abstract

To gain finer regional forecasts, many works have explored the regional integration from the global atmosphere, e.g., by solving boundary equations in physics-based methods or cropping regions from global forecasts in data-driven methods. However, the effectiveness of these methods is often constrained by static and imprecise regional boundaries, resulting in poor generalization ability. To address this issue, we propose Spatial-Temporal Weather Forecasting (STCast), a novel AI-driven framework for adaptive regional boundary optimization and dynamic monthly forecast allocation. Specifically, our approach employs a Spatial-Aligned Attention (SAA) mechanism, which aligns global and regional spatial distributions to initialize boundaries and adaptively refines them based on attention-derived alignment patterns. Furthermore, we design a Temporal Mixture-of-Experts (TMoE) module, where atmospheric variables from distinct months are dynamically routed to specialized experts using a discrete Gaussian distribution, enhancing the model's ability to capture temporal patterns. Beyond global and regional forecasting, we evaluate our STCast on extreme event prediction and ensemble forecasting. Experimental results demonstrate consistent superiority over state-of-the-art methods across all four tasks.

Figures (53)

• Figure 1: (1) Illustration of 3 regional forecasting strategies: (a) Crop neighbor region from global forecasts and forecast with regional variables; (b) Directly training; (c) Forecast by densely connecting global-regional model with distribution. (2) Region forecasting comparison of 3 strategies.
• Figure 2: Illustration of our method. (a) The overall structure of low-resolution global weather forecasting, which includes input Atmospheric variables, an Encoder, a Processor, a Decoder, and output Atmospheric variables; (b) The high-resolution regional weather forecasting structure with Spatial-Aligned Attention (SAA) module; (c) The typhoon track prediction structure with predicted high-resolution MSL; and (d) The long-term weather forecasting and ensemble weather forecasting.
• Figure 3: Illustration of Spatial-Aligned Attention.
• Figure 4: Illustration of Temporal Mixture of Experts.
• Figure 5: Comparison of our method with 6 competitors on denormalized RMSE $\downarrow$ and ACC $\uparrow$ in Global Weather Forecasting.