Every hour the correction models retrain and are scored against a held-out week of station observations across all served HRRR lead times (0–24 h). Each chart tracks that out-of-sample RMSE over time: the gap between the raw HRRR line and the corrected line is the skill the local model adds. Gust and direction are validated separately — corrections that cannot beat raw HRRR are disabled rather than served.
The wind meter charts show a local statistical correction layer built from HRRR and station observations. Models train on all served HRRR lead times (0–24 h) with the most recent week held out; a stage activates only when its held-out validation RMSE beats raw HRRR. Gust and direction corrections are validated and gated independently of wind speed.
| Station | Stage | Samples | HRRR RMSE | Corrected RMSE | Δ | Gust Δ |
|---|