Researchers introduce KiloGen, a new method that sharpens coarse operational wind forecasts into kilometre-scale wind maps over complex terrain. KiloGen combines a high-resolution wind “prior” learned from Weather Research and Forecasting (WRF) model simulations with 25 km forecasts from the European Centre for Medium-Range Weather Forecasts (ECMWF) during generation. Applied to Shanxi, China, KiloGen reconstructs terrain-controlled wind features and reduces forecast error, especially at elevated and topographically complex sites.

The team trained a diffusion generative model on high-resolution WRF output for the two horizontal wind components at 10 m height (zonal U10 and meridional V10). A diffusion prior is a statistical model that can produce many realistic fine-scale wind fields that look like the WRF simulations. At inference time KiloGen runs the model backward (a reverse diffusion step) while repeatedly mapping the generated field to the coarser ECMWF grid and nudging it toward the operational forecast. That combination keeps the large-scale forecast pattern while adding plausible local details. The authors call this “zero shot” because they do not train on matched pairs of ECMWF and WRF fields.

The trained prior closely matches WRF statistics. Generated samples reproduce WRF spatial patterns and percentile fields with high pattern correlations (0.96–0.99). Spectral tests show that KiloGen restores high‑wavenumber variance that is missing in the coarser ECMWF fields. In station comparisons KiloGen has the lowest overall wind speed root mean square error (RMSE) among the products tested. The operational ECMWF 0.25° forecast had a domain mean RMSE of about 1.75 m s⁻¹ and larger errors at higher elevation. KiloGen reduces RMSE most strongly during strong winds, by roughly 10% for observed winds above 20 m s⁻¹, and it improved on the 0.25° ECMWF forecast in all 13 strong wind events examined.