Improving medium‐range temperature forecast over the Tibetan plateau through spatially adaptive fusion

Tibetan Plateau exerts profound impacts on the global weather system, whereas its medium-range temperature forecast is challenging due to the complex topography. This study introduces Swin Transformer Fusion (STF), a spatially adaptive ensemble strategy that integrates forecasts from numerical weather prediction models (EC, GFS) and large meteorological models (Pangu, Fengwu). STF reduces 2-m temperature (T2m) forecast bias by 29.14%–38.45% across 1–10 days lead, outperforming conventional multi-model ensemble mean especially in high-bias regions. Attribution analysis reveals that Fengwu contributes most to STF's output despite Pangu's superior accuracy, challenging the assumption that higher-performing models should be weighed more heavily. These findings demonstrate that STF's spatially adaptive fusion enables region-specific integration of model strengths, offering an effective ensemble forecast strategy for topographically complex regions.