HRES Edges ML Models With 2.89 m/s RMSE in Northern Norway Wind Tests
A station-based comparison in Northern Norway finds HRES marginally ahead of two leading MLWP models for wind speed, with no post-training degradation observed. All models underestimate strong winds, highlighting remaining gaps in complex terrain. The work underscores that MLWP is now competitive but requires further refinement for operational local forecasting.
The arXiv study evaluated FourCastNet3, GraphCast, and ECMWF HRES against ground stations in fjord-dominated terrain where wind varies sharply over short distances. Researchers used several years of observations to test overall accuracy, post-training generalization, and high-wind events. HRES held a small edge in mean error, yet the two ML models showed no performance drop after their training cutoffs, indicating stable generalization to unseen years.
All three systems substantially underestimated peak winds, with FCN3 performing least poorly during the strongest events. This pattern aligns with earlier global benchmarks where MLWP models excel at typical conditions but lose resolution in orographic complexity. The findings matter for aviation, offshore operations, and avalanche forecasting that depend on accurate local gusts rather than broad averages.
Next steps include targeted fine-tuning on high-resolution regional reanalysis and ensemble post-processing to correct systematic underestimation of extremes. Without these adjustments, operational adoption in similar complex terrain will remain limited despite competitive average metrics.
Siyan Chen: MLWP high-wind RMSE will fall below 3.5 m/s in fjord terrain within 24 months after terrain-specific fine-tuning.
Sources (2)
- [1]Primary Source(https://arxiv.org/abs/2609.10564)
- [2]Supporting Source(https://www.science.org/doi/10.1126/science.adi2336)