Hybrid Swin Transformer–U‐Net Improves Subseasonal Maritime Wind Forecasts in China's Coastal Waters

Abstract Reliable subseasonal maritime wind forecasts are essential for navigation safety and disaster prevention, yet current forecasting systems show noticeable biases over China's coastal waters. Here, we introduce TDSwin‐UnetSS (Time‐embedded Downscaling Swin‐Transformer U‐Net for SubSeasonal forecasting), an asymmetric hybrid model that integrates the Swin Transformer with a U‐Net architecture. The model incorporates: (a) Weibull distribution constraints to preserve the wind‐speed probability density, (b) spatial downscaling from 1.5 to 0.25 resolution, (c) time embedding for simultaneous 3–6 week lead‐time corrections, and (d) spatially adaptive learning focused on extreme wind‐prone regions. Applied to ECMWF S2S forecasts, TDSwin‐UnetSS reduces root mean squared error (RMSE) by 13.41%–16.67%, mean absolute error (MAE) by 10.69%–13.67%, and the absolute value of the bias by 80.00%–90.00%, while increasing anomaly correlation coefficient (ACC) by 18.18%–76.19% and structural similarity index measure (SSIM) by 5.71%–10.77% across different lead times, with RMSE reductions exceeding 19.76% during extreme wind events. This study demonstrates the application of the Swin Transformer to subseasonal wind bias correction, highlighting its potential for extended‐range forecasting through effective multi‐scale feature learning and physics‐informed hybrid modeling.