Joint near-surface wind downscaling and hub-height estimation for wind power projections

Climate projections are widely used to estimate future wind power production and assess the impacts of climate change on wind energy resources. However, these models typically operate at coarse spatial resolutions ( \(\sim\) 100 km) and provide wind components only at 10 m above the surface, rather than at typical wind turbine hub heights (100 m). This study presents a deep learning framework to simultaneously downscale near-surface wind components and estimate hub-height wind fields, avoiding the use of the conventional wind power law for vertical extrapolation. A ResUNet-based architecture is trained on coarsened ERA5 reanalysis data over the area of Greece, together with static features such as elevation and land use, to reconstruct high-resolution (0.25 \(^{\circ }\) ) wind fields at both 10 m and 100 m. The proposed method shows improved performance compared to the power-law approach, reducing RMSE (24.38%) and MAE (42.78%) in hub-height wind speed estimation. By jointly learning spatial and vertical dependencies, the model helps reduce biases associated with empirical extrapolation.