Wind farms · what the rule would have done, and whether the radar agreed

The daily wind-farm forecast suggests stopping only when the passage is very high and stopping is cheap. Before anyone relies on it, it has to answer two questions on the past: what would it have cost, and were the nights it picked really heavy? Both are answered here on the ERA5 reanalysis (Copernicus) at each zone's point from 2021, the archive the zone thresholds are cut on.

1. What stopping would have cost

The model is run over every past night of every zone. The level comes from the zone's own thresholds (percentiles over the nights around the same date) and the advice from the same table as the forecast. Nights are counted per season and zone; output lost is the share of the zone's night output (generic power curve, ERA5 wind at 100 m) that falls on the nights the rule picks; migration covered is the share of the predicted density that passes on them.

Stopping on every very high night would cover a little more migration at about four times the cost: the extra nights are windy ones. Stopping only on the calm ones keeps most of the protection for a small share of the output, because heavy passage tends to come on nights with little wind.

2. Were those nights heavy? The nearest radar

No radar measures at a wind farm, but the nearest one says whether the night was busy in that region. Each zone within 75 km of a radar of the European network is paired with it, and the radar's measured night is placed in that radar's own record for the same season and the same years. A busy night is one in the top quarter of that record, as in the city verification, so by chance about a quarter of the nights are busy (23.9% here).

The levels sort the nights the right way: a low night is busy about half as often as chance and a very high one more than twice as often. The nights the rule says to stop are the busiest of all. Per zone, the median rank correlation between predicted and measured density is +0.63.

What this does and does not show

• Reanalysis, not forecast. The weather is ERA5, the best estimate of what happened, not the forecast that would have been published days before. It measures the model and the rule; the error of the forecast itself is measured by the city verification since publication started.
• The radar is not at the farm. Pairs are 8 to 75 km apart and the radar sees the whole column from about 200 m to 3 km, not the rotor layer. A busy night at the radar says the region was busy.
• Left out. Zones farther than 75 km from any radar, nights with less than 60% radar coverage, radar-nights with fewer than 100 reference nights, and the radar record before the hindcast years: several radars measured on another scale before 2020 (the median night of 2016 is five times today's), which would make every recent night look quiet.
• Not independent of the model. The model was trained on these radars. It was validated leaving whole radars out, but a zone near a training radar is not a fully new place.

Weather: ERA5 hourly reanalysis, Copernicus Climate Change Service. Radar profiles: ENRAM / Aloft (CC0). Turbines © OpenStreetMap contributors (ODbL).