Key points are not available for this paper at this time.
Abstract High‐resolution kilometer‐scale wind fields are essential for wind energy assessment and extreme weather analysis. However, current downscaling methods still have huge biases in complex terrain. In this study, we propose Terrain‐Aware Vision Transformer (TA‐ViT), a novel deep learning framework for wind field downscaling from 25 to 1 km, which integrates dual‐branch patch embedding, feature‐wise linear modulation (FiLM), type‐aware attention gating, and progressive upsampling to effectively capture terrain‐induced wind modifications. The model achieves superior performance with root mean square error of 1.15 m , Pearson correlation coefficient of 0.93, structural similarity index measure of 0.78, and peak signal‐to‐noise ratio (PSNR) of 20.48 dB. Notably, TA‐ViT demonstrates enhanced performance in steep terrain regions, outperforming traditional statistical downscaling method, interpolation method, and deep learning baselines. During typhoon passages, the model demonstrates robustness in capturing localized wind extremes under extreme and historical weather events. These findings suggest that TA‐ViT provides a promising physics‐aware solution for high‐resolution wind field downscaling.
Li et al. (Mon,) studied this question.