Direction-Informed Deep Learning with Topographic Fusion for 10m Wind Downscaling

Abstract Near surface 10 m wind fields are shaped by multiscale interactions between large scale circulation and terrain induced processes, which poses challenges for maintaining spatial coherence during downscaling in large heterogeneous regions. To improve the representation of near‐surface winds in a physically informed manner, we propose a deep learning downscaling method that incorporates direction preserving normalization based on wind speed magnitude, a wind direction integrated loss function, and terrain elevation including spatial derivatives. This approach achieves downscaling of wind fields over the North China region from 0.25° to 0.1°. Using a U‐Net‐based framework, we compare three network architectures. High resolution terrain elevation data and its spatial derivatives were incorporated as additional inputs. The proposed design improves both wind speed and wind direction performance. Specifically, the mean wind speed RMSE decreased by 0.74 m/s, and the mean wind direction mean angular error decreased by 4.82°. Adding terrain elevation data further lowered the errors. The new architecture with terrain elevation data showed smaller wind speed errors in larger lakes areas. Therefore, combining terrain elevation data and using the new loss function effectively improves the physically consistent representation of downscaled near surface wind fields.