From surface observations to hub-height wind fields: A two-stage framework combining ML-based interpolation and terrain-aware vertical extrapolation for wind energy applications in the Austrian Alps

Accurate spatiotemporal wind fields at wind turbine hub heights (80–200 m) are essential for renewable energy resource assessment and grid integration studies, yet observational constraints typically limit measurements to sparse station networks at ideally 10-meter reference height. In complex Alpine terrain and semi-complex northern and eastern Austrian terrain, both horizontal interpolation and vertical extrapolation pose exceptional challenges due to orographic flow acceleration, valley channelling, and stability-driven wind shear variations.

We present a comprehensive two-stage framework that transforms sparse concise and quality controlled surface observations into high-resolution, multi-level wind fields and wind power potential suitable for wind energy applications over Austria. The framework processes 25 years (1996–2020), 30 possible if considering all available data, of hourly wind observations from approximately 280 stations and produces gridded wind fields at 1 km horizontal resolution across multiple hub heights.

For horizontal interpolation, an Empirical Orthogonal Function (EOF) decomposition reduces computational complexity by factor ~250× while retaining >95% of spatiotemporal variance. We compare six interpolation approaches: Inverse Distance Weighting, Kriging with External Drift (KED), Random Forest, Bayesian Additive Models for Location, Scale and Shape (BAMLSS), and a Deep Neural Network. Eight terrain-aware covariates capture orographic effects, including topographic position indices, wind exposure indices, surface roughness from CORINE land cover, and ERA5 reanalysis as large-scale atmospheric forcing. Terrain covariates prove essential, with largest gains in complex topography.

For the vertical extrapolation, seven complementary methods extrapolate 10-meter wind fields, for every interpolation method separately, to hub heights of 80, 100, 120, 140, 160, 180, 200, 220, and 250 m: enhanced logarithmic and power law profiles with spatially-variable roughness lengths, stability-dependent extrapolation using surface wind speed as atmospheric stability proxy, directional-terrain correction accounting for orographic sheltering and acceleration, roughness-adaptive method selection, a deep learning model, and a multi-method ensemble providing uncertainty quantification through ensemble spread. GPU acceleration enables efficient processing of massive datasets (~200 million grid points per monthly file).

Preliminary validation against the New European Wind Atlas demonstrates that both the ensemble approach and the machine learning approach captures diurnal wind shear variations and reproduces known orographic patterns, with largest improvements over traditional single-method extrapolation in areas of complex topography.

The resulting multi-decadal, hourly wind speed dataset at multiple hub heights provides a novel resource for Austrian wind energy resource assessment, capacity factor estimation, and renewable integration studies. The modular framework design supports both retrospective climate analysis and operational nowcasting applications.