Detailed vertical wind speed profiles are essential for wind energy resource assessment, but direct measurement campaigns are expensive and spatially limited. Global reanalysis datasets such as ERA5 provide broad coverage but at coarse vertical resolution. Bridging this gap — estimating site-specific, high-resolution profiles from widely available reanalysis variables — is both practically valuable and methodologically non-trivial.
This study introduces a methodology using TabNet, an attention-based deep learning model designed for tabular data. Rather than predicting wind speeds at individual heights directly, the approach first compresses profiles into five Chebyshev polynomial coefficients representing 12 heights from 10–500 m, then trains TabNet to predict these coefficients from 34 ERA5 meteorological features. Training and testing splits mimic the measure-correlate-predict framework commonly used in industry.
TabNet achieved R²=0.93 for mean wind speed and R²=0.65 for wind shear, performing consistently across diverse wind regimes including high-shear, well-mixed, low-level jet, and high-wind conditions. Feature importance analysis highlighted atmospheric stability metrics — sensible heat flux, temperature gradient, boundary layer height — alongside wind speed variables as the most influential predictors, providing physically interpretable insight into the model’s behavior.