Direction-aware multiscale terrain fusion for 10-m wind forecast downscaling over Xinjiang, China

Abstract High-resolution near-surface winds are critical for wind-resource assessment, atmospheric transport modelling, and weather-risk management. However, global numerical weather prediction models do not adequately resolve fine-scale wind variability over complex terrain. We propose a Direction-Aware Multi-Scale Terrain Fusion U-Net (DMTF-UNet) to downscale 0.25° Global Forecast System 10-m wind forecasts to the 0.1° ERA5-Land grid over Xinjiang, China. A single model predicts the horizontal wind components at seven forecast lead times from 24 to 168 h. DMTF-UNet applies direction-preserving wind normalization and persistently fuses elevation and terrain-gradient information across multiple network scales. A vector-aware loss jointly constrains wind components, speed, and direction. Evaluation used a temporally independent 2024–2025 test set, with bilinear interpolation followed by linear regression and a conventional U-Net as benchmarks. All methods used common wind predictors and the same evaluation protocol. Across three independent training runs, DMTF-UNet achieved a mean wind-speed root-mean-square error of 1.013 m s⁻¹ and wind-direction mean angular error of 26.41°. This represented a 10.5% reduction in wind-speed error relative to U-Net. DMTF-UNet outperformed both benchmarks at all seven leads, with statistically significant reductions in wind-speed and direction errors relative to U-Net. Spatial error maps showed marked wind-speed improvements in the Tarim Basin–southern Tianshan region, with widespread but spatially heterogeneous and lead-dependent directional gains. These findings support terrain-informed, vector-aware learning for deterministic near-surface wind downscaling over complex terrain.