Wind-Regime and Spatial-Transfer Performance of a Station-Trained 5 km Wind-Speed Correction over Hainan Island

Regional 10 m wind-speed products extend coverage beyond sparse station networks, but their errors vary with wind regime and location. This study evaluates a station-trained LightGBM correction of a 5 km WRF/HNR wind-speed product over Hainan Island. Raw and corrected fields were collocated with stations using a four-neighbour inverse-distance operator. A spatially separated 2022 holdout retained 67 stations after five coordinate-overlap exclusions. Across 582,610 common hourly pairs, corrected RMSE decreased from 3.198 to 1.211 m s−1 and Pearson R increased from 0.368 to 0.651. The paired RMSE reduction was 1.986 m s−1 [95% CI 1.839, 2.122]. Improvement was largest below 3.4 m s−1 (2.203 m s−1) and remained positive for 3.4–7.9 m s−1 (0.566 m s−1). In contrast, corrected RMSE increased by 0.990 m s−1 for 8.0–10.7 m s−1 and by 1.656 m s−1 for ≥10.8 m s−1. A time-only adjustment reduced pooled RMSE to 1.870 m s−1 but also degraded the two upper strata. The 2016–2022 maps show a lower corrected product climatology, with a mean corrected-minus-raw increment of −2.273 m s−1 across the Hainan buffer. Temporal partitions and AWS-selected station groups show variation across season, time of day, and setting. The correction reduces station-sampled errors for light-to-moderate wind speeds; it changes speed magnitude while retaining the raw-product wind direction.