Insights / Worked example · WindBy Robert Sills CEng MIMechE16 September 2026

Can a nacelle anemometer measure turbulence?

Four months of nacelle lidar data on one Hill of Towie turbine put the SCADA anemometer to the test. Mean wind speed agrees within 1%, turbulence agrees on average but not record by record and SCADA data alone closes half the gap.

← All insights

Turbulence intensity taken from a nacelle anemometer is treated with suspicion, and rightly so. The sensor sits behind the rotor in flow the rotor and nacelle have already disturbed, and the controller corrects only the mean wind speed. An earlier post looked at nacelle transfer function accuracy against a lidar, which is the same question asked of the mean rather than of the variability around it.

Version 2 of the Hill of Towie open dataset makes the turbulence question testable. It adds a nacelle mounted lidar on turbine T07 measuring 208 m ahead of the rotor, which is 2.5 rotor diameters, alongside the site’s 10 minute SCADA. Four months of the two together give 4,460 concurrent records.

Two scatter plots side by side. On the left, nacelle anemometer mean wind speed against lidar wind speed, with the points lying tightly along the 1:1 line and an R squared of 0.89. On the right, anemometer turbulence intensity against lidar turbulence intensity over the same records, a far wider cloud either side of the line with an R squared of 0.51. — Turbine T07 primary anemometer against the nacelle lidar at the 208 m range gate, 4,460 ten minute records from January to April 2026. The same sensor, the same records and the same reference on both panels. Fits are least squares through the origin.

Mean wind speed is not the problem

The anemometer tracks the lidar closely once the nacelle transfer function has done its work.

Over the four months the anemometer averaged 7.95 m/s against 7.92 m/s from the lidar, a bias of 0.4%, with an R² of 0.89 record by record. The site’s other lidar, a ground based unit near turbine T11, gives the same answer for that turbine to within 2%. Two turbines and two independent instruments agree, so mean wind speed is settled.

The scatter that does exist is not evenly spread. Two direction sectors, 90 to 120 degrees and 150 to 180 degrees, carry most of it, with the anemometer reading roughly 1 m/s low in both. A single site wide transfer function cannot correct something that depends on direction.

Turbulence intensity is right on average and wrong record by record

The four month mean is close, and almost nothing below that timescale is.

The anemometer averaged a turbulence intensity of 0.165 against 0.160 from the lidar, a bias of 3%. The R² is 0.51. Only 53% of 10 minute records fall within 20% of the lidar value and one record in ten is out by more than 50%. A number that is reliable over four months and unreliable over ten minutes is a different kind of measurement from the wind speed beside it.

The error also has a shape. Binned by the lidar value, the anemometer reads about 55% high in the calmest bin, crosses zero near a turbulence intensity of 0.20 and reads 22% low in the roughest. The likely explanation is that the rotor and nacelle add turbulence of their own, which dominates the reading when the inflow is calm. The sensor then smooths the extremes when the inflow is rough. Whatever the cause, the sensor compresses the range towards the middle and the two ends largely cancel in the four month mean.

Mean error of the nacelle anemometer against the lidar, binned by lidar turbulence intensity. The anemometer line starts 55 per cent high in the calmest bin, falls steadily through zero near a turbulence intensity of 0.20 and reaches 22 per cent low in the roughest bin. The SCADA model line follows the same shape but much flatter, staying within about 13 per cent across the range. — Mean error against the lidar in each turbulence intensity bin, for the raw anemometer and for the SCADA model described below. Bins hold between 343 and 646 records. A flat line at zero would be a sensor with no level dependent bias.

A model built from SCADA alone closes half of the gap

Most of what the anemometer misses is recoverable from tags the turbine already records.

A gradient boosted model using five SCADA quantities, wind direction, both anemometer turbulence intensities, the wind vane standard deviation and wind speed, predicts the lidar value with an R² of 0.78 against 0.51 for the anemometer alone. Records within 20% of the lidar rise from 53% to 79%, and the level dependent bias in the chart above halves. Adding every other available SCADA tag changes nothing.

Wind direction is the strongest single input, which fits the mechanism, because both the nacelle flow distortion and the terrain upstream depend on where the wind comes from. Rotor speed, pitch and power add nothing measurable, so the controller signals that are near deterministic functions of wind speed carry no turbulence information once averaged to 10 minutes.

Scatter of SCADA model turbulence intensity against lidar turbulence intensity over the same 4,460 records. The cloud is substantially tighter around the 1:1 line than the raw anemometer scatter, with an R squared of 0.78 and a fitted slope of 0.97. — The same records as the first chart, with the SCADA model in place of the raw anemometer. Every value plotted is an out of fold prediction, from folds grouped by calendar day so that no record shares a day with its training data.

What the data shows

Mean wind speed from the nacelle anemometer agrees with the lidar to within 1% over four months, with an R² of 0.89.
Turbulence intensity agrees to within 3% on the four month mean but only reaches an R² of 0.51 record by record.
The turbulence error depends on the turbulence level, around 55% high in calm inflow and 22% low in rough inflow, so the small average hides two large errors that cancel.
Five SCADA quantities lift the record by record R² to 0.78 and halve the level dependent bias.

Method notes. Records are filtered to normal operation, at least 80% lidar data availability in the interval, no rain flagged by the lidar and lidar wind speed between 4 and 16 m/s. All wind directions are used, so wake affected sectors sit in both the anemometer and the lidar. The lidar reference is the ZX TM spectral average turbulence at the 208 m range gate and the 59 m rotor segment height, and it carries its own uncertainty, so these results are a comparison against a lidar estimate rather than against a mast. No filter is applied to the lidar fit residual, because selecting on it keeps only the least complex inflow and flatters the anemometer. SCADA timestamps are end of period and the lidar timestamps are start of period, an offset worth correcting before any comparison of this kind.

What it means for a turbine by turbine estimate

Averaged over enough data, the turbine’s own anemometer is a better starting point for turbine by turbine turbulence than a single mast scaled to each turbine’s mean wind speed.

That is most true in complex terrain, where one mast represents the inflow at every other turbine poorly. The nacelle anemometer is the only sensor that was on every turbine for every 10 minute period of the operating history, so a corrected version of it is the closest thing to a measured turbulence history a fleet has. PowerVeritas reviews the measurement data a site already holds and sets out a plan to collect or correct the best available data for use as part of its life extension strategy support.

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