WeatherBlend

Multi-model forecast blending for Sennen, Cornwall

Wind models

10 m wind speed — ERA5-truth `wind` + Dunkeswell-truth `wind_speed_lgb`; plus gust and direction (`wind_mvn`). MAE — m/s for speed/gust, ° for direction; lower better. Δ vs best single NWP — negative = blend wins.

10 m wind speed — Sennen Cove

Phase wind · v2026-10-04_161036

10 m wind-speed LightGBM blender, ERA5 WindSpeed10m truth. Per-NWP speed + direction (sin/cos) + ensemble spread + calendar. Trained 2026-10-04. Metric: Test MAE (m/s).

Lead Blend Best single Δ vs best
+24h 0.563 ecmwf_aifs025_single (0.590) -4.7%
+48h 0.749 ecmwf_aifs025_single (0.743) +0.7%
+72h 0.904 ecmwf_aifs025_single (0.922) -1.9%
Verify history (no runs yet)

No verify rows yet for phase wind. Next cycle: Mon/Thu 09:30 UTC, then 5d ERA5 latency.

Phase wind_speed_lgb · v2026-10-04_132929_wind_speed_lgb

10 m wind-speed LightGBM blender trained on Dunkeswell SYNOP obs (real-station truth) instead of ERA5. 29-feature lean+ORO+spread set. The honest real-wind predictor — see the cross-truth bake-off. Trained 2026-10-04. Metric: Test MAE (m/s).

Lead Blend Best single Δ vs best
+24h 1.372 ukmo_seamless (2.434) -43.6%
+48h 1.308 icon_seamless (2.417) -45.9%
+72h 1.423 icon_seamless (2.467) -42.3%
Verify history (no runs yet)

No verify rows yet for phase wind_speed_lgb. Next cycle: Mon/Thu 09:30 UTC, then 5d ERA5 latency.

Wind direction — Sennen Cove

Phase wind_mvn · v2026-10-04_132632_wind_mvn

PyTorch MLP predicting a bivariate normal over the (u, v) wind vector, trained on Dunkeswell SYNOP truth. Direction = atan2(−μ_u, −μ_v); also emits a calibrated direction confidence interval. Trained 2026-10-04. Metric: Test MAE (°).

Lead Blend Best single Δ vs best
+24h 10.906 ecmwf_ifs025 (12.115) -10.0%
+48h 14.011 ecmwf_ifs025 (15.240) -8.1%
+72h 18.310 ecmwf_ifs025 (19.495) -6.1%
Verify history (no runs yet)

No verify rows yet for phase wind_mvn. Next cycle: Mon/Thu 09:30 UTC, then 5d ERA5 latency.

10 m wind gust — Sennen Cove

Phase wind_gust_lgb · v2026-10-04_161132_wind_gust_lgb

10 m wind-gust LightGBM blender, ERA5 WindGusts10m truth. Per-NWP gust + gust/speed ratio + ratio spread over the gust-publishing NWPs. Trained 2026-10-04. Metric: Test MAE (m/s).

Lead Blend Best single Δ vs best
+24h 1.025 gem_seamless (1.594) -35.7%
+48h 1.195 gem_seamless (1.800) -33.6%
+72h 1.331 gem_seamless (1.635) -18.6%
Verify history (no runs yet)

No verify rows yet for phase wind_gust_lgb. Next cycle: Mon/Thu 09:30 UTC, then 5d ERA5 latency.