Super local wind prediction Valkenburgse meer [under development]

What is the super local forecast? This dashboard combines two local machine learning models that take large-scale wind-model predictions as input and are trained on historical forecast values with matching measured wind values at this location. The local models are calibrated to local data to improve prediction performance by learning systematic local deviations from the large-scale model. One model is dedicated to the remaining part of the current day and gives strong weight to the most recent measured wind updates. A second model is dedicated to next-day (day-ahead) prediction. Models are retrained daily, next-day/current-day prediction lines are refreshed hourly during daytime, and measured-wind updates on the current-day plot are refreshed every 6 minutes.

Current-day prediction

Measured wind speed up to now, plus the latest Harmonie and super-local prediction for the remaining hours of today.

Click a plotted point to view exact values.

If interactivity is unavailable, the static plot below remains visible.

Next-day prediction

Day-ahead forecast for tomorrow: Harmonie baseline versus the super-local model for wind speed and direction.

Click a plotted point to view exact values.

If interactivity is unavailable, the static plot below remains visible.

How much better are the super local forecasts?

Current-day performance by wind direction

The spider diagram compares realised current-day forecast error by forecast wind direction; lower values are better. Super local improves most for E, SE, NE, S winds; the largest gain is 1.9 kts for E. For that sector the super-local model overestimates by 0.3 kts on average. The advantage is smallest, or negative, for NW, N winds.

Next-day performance by wind direction

The spider diagram compares mean absolute error by forecast wind direction; lower values are better. The next-day champion improves most for SE, E, NE, S winds. The advantage is smallest for SW, W winds.

Model-gate evaluation history

Top panel shows the holdout wind-speed comparison used by the model gate, including the aligned Harmonie baseline from the same holdout forecast inputs. Bottom panel shows the corresponding MAE comparison for Harmonie, challenger, and champion.

What is the super local forecast? This dashboard combines two local machine learning models that take large-scale wind-model predictions as input and are trained on historical forecast values with matching measured wind values at this location. The local models are calibrated to local data to improve prediction performance by learning systematic local deviations from the large-scale model. One model is dedicated to the remaining part of the current day and gives strong weight to the most recent measured wind updates. A second model is dedicated to next-day (day-ahead) prediction. Models are retrained daily, next-day/current-day prediction lines are refreshed hourly during daytime, and measured-wind updates on the current-day plot are refreshed every 6 minutes.