The Austrian Windatlas: An high-resolution Observation-Constrained, Ensemble Windspeed Analysis from Near-Surface to Turbine Hub Heights with near-realtime option

Reliable, high-resolution wind climatologies at turbine hub heights are fundamental to wind energy planning, yet existing products for Austria either lack observational constraint (ERA5, NEWA) or do not provide the continuous hourly time series needed to quantify variability and extremes. We present the Austrian Windatlas, a 25-year (1997–2021), hourly, 1 km observation-constrained wind analysis with option for near-realtime production from near-surface to 220 m above ground.

The atlas is produced via a two-stage pipeline. In Stage 1, near-surface (10 m) wind speed fields are reconstructed from the dense GeoSphere Austria TAWES network (~280 stations) using a shared EOF/rPCA decomposition framework that enables direct, fair comparison of six interpolation families: regression-kriging, Bayesian Additive Models for Location Scale and Shape (BAMLSS), random forest, and three deep-learning variants, each tested with and without ERA5 or CERRA reanalysis backgrounds. Validated against 54 permanently withheld stations, the enhanced deep-learning architecture with CERRA background achieves the best overall skill (RMSE = 1.46 m s⁻¹, r = 0.65 for 2020), outperforming the operational INCA+CERRA analysis by ~8% and demonstrating that the higher-resolution CERRA reanalysis (5.5 km) consistently provides better background constraint than ERA5 (31 km) across all method families. Notably, simple regression-kriging (RMSE = 1.51 m s⁻¹) remains highly competitive at a fraction of the computational cost.

In Stage 2, the gridded 10 m fields are extrapolated to hub heights (80–220 m AGL) using a further zoo of methods — from classical log-law and terrain-adaptive power-law formulations to a NEWA-trained, height-agnostic machine-learning model — with final validation against independent NEWA profile holdouts and operational SCADA data from Austrian wind turbines. The inter-method ensemble spread is retained throughout as a spatially explicit, per-timestep uncertainty estimate, propagated end-to-end from surface to hub height.

The resulting climatology captures long-term variability, seasonal and diurnal cycles, and orographic flow signatures systematically absent from reanalysis-only products. We discuss the complementarity of observation-constrained and NWP-calibration approaches to national wind atlasing, the added value of CERRA over ERA5 as a background field in Alpine terrain, and the atlas as a baseline for future statistical climate-change downscaling of wind resources under CMIP6 scenarios as well as for operational near-realtime analyses.