A deep-learning-based fusion correction model for cold wave forecasting over the Tibetan Plateau

Over the past two decades, both numerical weather prediction (NWP) models and AI-based large meteorological models have significantly improved medium-range forecast accuracy. However, due to inherent model uncertainties and inadequate simulation of complex terrain, these models systematically underestimate extreme weather intensity in topographically challenging regions like the Tibetan Plateau. This study evaluates the performance of traditional NWP models, large meteorological models, and multi-model ensemble forecasts in predicting near-surface air temperature, using the 14th December 2023 cold wave event as a case study. Results indicate that while traditional and large meteorological models effectively capture spatial patterns of temperature anomalies, they consistently underestimate extreme cold intensity. Although the multi-model ensemble mean can improve the spatial correlation coefficient to some extent, its performance in predicting the scope and intensity of extreme low temperatures still needs improvement. To address these limitations, we propose a Swin Transformer fusion model (STF) incorporating positional encoding. This framework enables synergistic optimization of multi-model forecasts by systematically extracting and integrating the strengths of NWP and large meteorological models at specific spatiotemporal scales. During the cold wave's peak phase, STF reduced forecast root mean square error by up to 39.62%, with particularly notable improvements in error-sensitive regions. The model's dynamic preference-error hedging mechanism effectively combines the multi-model advantages, enhancing both forecast accuracy and operational robustness for extreme weather events. This work advances cold wave early warning systems for high-altitude regions, introduces novel methodologies for extreme weather prediction, and demonstrates promising practical applications.