Insights / Field note · WindBy Robert Sills CEng MIMechE9 September 2026

Anemometers are calibrated running, not stopped

The wind speed a turbine reports is a corrected reading, and the correction is measured while the rotor turns. Parked, the same sensor reads around 21% low against a neighbour that kept running.

← All insights

The wind speed in a turbine’s SCADA record is not a raw measurement. The anemometer sits behind the rotor, so the controller applies a nacelle transfer function to convert what the sensor sees into an estimate of the free stream. An earlier post looked at how good a nacelle transfer function is against a lidar measuring the wind ahead of the rotor.

That correction is measured on an operating turbine. IEC 61400-12-2 and its companion IEC 61400-12-6 build the transfer function from concurrent records where the machine is online, and the data rejection rules remove anything that is not normal operation. Nothing in the derivation describes a stopped rotor, which matters because a stopped rotor is exactly the condition in which the reading is used to value lost energy.

Scatter of one turbine's nacelle wind speed against a neighbouring turbine that was generating throughout. Teal points recorded while the subject turbine was also operating lie along a line of slope 1.03. Amber points recorded while it was parked sit clearly below, on a fitted line of slope 0.81. — One turbine’s nacelle anemometer against a neighbouring turbine that was generating in every interval shown, 10 minute records from January 2023 to April 2026. Fits are least squares through the origin on the full samples. The operating cloud is drawn from a random one in sixteen sample so the parked records stay visible.

The relationship changes when the rotor stops

Comparing a turbine against a neighbour separates the sensor from the weather, because both machines see the same passing weather system.

Across 126,952 intervals where both turbines were generating, the fitted slope is 1.03. The pair track each other closely, which is what a working sensor and a valid correction look like. Across the 1,348 intervals where the subject turbine was parked with the rotor stopped and the neighbour kept generating, the fitted slope falls to 0.81. The same sensor, against the same neighbour, reads around 21% low once the rotor stops.

The parked records span 15 separate outages and reach 25 m/s, so this is not one long event or one unusual week.

The error is largest in the middle of the range

Pooling all 21 turbines on the site shows the same effect and reveals that its size depends on the wind speed.

Dividing each turbine’s own reading by the mean of the turbines generating in the same interval gives a ratio that should sit at 1.00 when the sensor and the correction are working. Across 1.6 million operating records it does exactly that, flat at 1.00 from 4 to 17 m/s, which is the control that makes the rest of the chart trustworthy. The parked line runs well below it, bottoming out near 0.86 around 6 to 7 m/s and climbing to about 0.95 by 15 m/s.

What the data shows

A parked nacelle anemometer reads low against turbines that are still generating, on every turbine tested.
The shortfall is roughly 14% at 6 to 7 m/s and narrows to around 5% by 15 m/s.
The same comparison on operating records sits at 1.00 across the whole range, so the method is not manufacturing the result.
The size varies by turbine and by year, from around 10% to around 30%, so a single fleet correction factor would be the wrong fix.

Method notes. Operating means rotor speed above 5 rpm, export above 20 kW and no curtailment, bad data or startup and shutdown flag. Parked means rotor speed below 0.5 rpm with no export. The pooled chart uses 34,047 parked and 1.6 million operating 10 minute records from 2022, 2025 and 2026 to date. Bins above 17 m/s are not plotted because fewer than 50 parked records fall in them. The measured direction is the opposite of what rotor induction alone would predict, so the transfer function argument explains why the correction should not be trusted at standstill without explaining the sign of the error.

Binned median ratio of each turbine's own wind speed to the mean of the turbines generating in the same interval, against wind speed. The operating line is flat at 1.00 across the range. The parked line sits near 0.86 at 6 to 7 metres per second and rises to about 0.95 by 15 metres per second, with a wide interquartile band. — All 21 turbines pooled for 2022, 2025 and 2026 to date. Each turbine’s own wind speed divided by the mean of the turbines generating in the same 10 minute interval, binned in 1 m/s steps. The band is the parked interquartile range.

What it costs depends on the reference used

A wind speed that reads low understates the energy a stopped turbine would have produced, but not by a fixed amount.

Where the potential is read straight off a power curve at the turbine’s own wind speed, the understatement flows through directly. Recalculating one site’s downtime for a single year against peer derived wind instead raised the reported loss by 8.6%.

The practical conclusion is narrower than the headline. Estimate the wind a stopped turbine would have seen from the turbines that were still running, and keep its own anemometer as the last resort rather than the default. PowerVeritas builds downtime losses this way as part of its Performance Reviews.

Charts and analysis by PowerVeritas. Where open datasets are used, sources are credited on the attributions page.