High-Resolution Offshore Wind Climatology and Energy Potential in the Gulf of Tonkin Downscaled by an Enhanced Swin Transformer Model

High-resolution offshore wind information is crucial for wind-energy development in regions with complex coastlines (such as the Gulf of Tonkin, Vietnam) and monsoonal dynamics. However, coarse-resolution reanalysis products like ERA5 (31 km) fail to capture the small-scale variability required for reliable wind-resource and climatological assessments. Here, we propose a validated deep-learning downscaling framework based on the ConvSwin2SR (Convolutional Swin Transformer for Super-Resolution) architecture, termed ConvSwin2SR_BI2, which integrates Residual-in-Residual Dense Blocks and Squeeze-and-Excitation modules to enhance spatial feature extraction and wind-field reconstruction. The model is trained with four years (2020–2023) of 3-km WRF-ARW simulations initialized by ERA5 and then applied to generate a 30-yr (1994–2023) 6-hourly offshore wind dataset at 3-km resolution over the Gulf of Tonkin. Compared with bilinear interpolation, CNN, U-Net, and earlier ConvSwin2SR variants, ConvSwin2SR_BI2 achieves markedly better performance (MAE < 1 m s⁻¹; CC ≈ 0.90) for both U₁₀ and V₁₀ components. Independent validation against CCMP and ASCAT satellite winds (2019–2023) confirms improved spatial fidelity, with RMSE reductions of up to 45% relative to ERA5 across coastal and offshore zones. Seasonal and sub-regional analyses demonstrate robust gains during both winter and summer monsoon regimes. The resulting 30-yr 3-km wind power density and capacity-factor climatologies indicate 10%–15% higher energy potential than ERA5-based estimates in key coastal development zones. This study provides the first Transformer-based deep-learning framework validated for offshore wind downscaling in Vietnam, delivering a high-resolution Gulf of Tonkin wind atlas useful for wind-energy planning, marine safety, and regional climate applications.