A wind farm's data can be wrong in quiet ways. An anemometer can read ten per cent low and the turbine still looks fine. An energy counter can run five per cent fast and nothing alarms, because nothing crosses a limit. Somewhere a spreadsheet then says the farm produced more than it did, or that a turbine is worse than it is.
Four wind farms have published their SCADA data under an open licence: Kelmarsh and Penmanshiel by their owner Cubico and Hill of Towie by RES, all three in Britain, and La Haute Borne in France by ENGIE. Two makers, three ways of exporting. On them we ran five tests. Each was written down, with its pass marks, before its data was opened, and none was tuned afterwards. Most failed their own marks. Here is what they showed.
Reading a maker's own names
A turbine's signals have names its maker chose: Temp. gearb. inlet 1s on a Senvion, wtc_GeOilTmp_mean on a Siemens. The international standard for wind power plant communication, IEC 61400-25, names the parts of a turbine, but in practice each maker ships its own list. Someone buying a farm, or running two makers side by side, has to work out what each signal is.
67% of a Siemens farm's documented signals readagainst 17% for a plain keyword list. Wrong on 4 of 32, more than the pass mark allowed.
We wrote the reader on Kelmarsh's Senvion names and tested it on the two farms it had not seen. On Hill of Towie's 48 documented Siemens signals it answered 67% and refused the rest with a reason. It got 44 more right per 100 than the keyword list (95% interval 23 to 60). But it was wrong on 4 of the 32 it answered, and the pass mark was one in ten. Verdict: FALSIFIED. Two of the wrong answers read the letters U and V in a generator winding's name as volts. Two read the side of a gearbox bearing as a different part. A single letter is not a word, and a guessed part is a guess.
Physics against alarm limits
An alarm list has fixed limits: too high, too low, stuck. Physics asks different questions. Does the power follow the turbine's curve at the wind it measured? Does its wind agree with the turbines around it? Does a day's energy match a day's power? We planted faults in copies of real days on Penmanshiel, a farm the checks were not set on, and counted what each method found.
99% of planted faults foundagainst 47% for alarm limits, at the same alarm budget
An anemometer reading ten per cent low for a week: found 53 times of 53, and never by the limits. A turbine producing ten per cent less: 49 of 52, never by the limits. An energy counter five per cent high: 57 of 57, never by the limits. None of these crosses a limit. That is what the physics is for.
But on real days, with nothing planted, the checks raised an alarm on 4.0% of turbine-days, and the pass mark was 3%. The limits did worse, at 7.2%. Verdict: FALSIFIED. Most of the extra alarms were turbines producing more than usual when the wind turned: the reference did not follow the wind direction.
Another maker, and a lesson that did not carry
We fixed both causes and tested again on Hill of Towie, a Siemens farm with another export. The checks alarmed on 31.8% of real days, nearly all from one check. On this farm the gear oil temperature holds one value for hours in normal running, so a check that called six identical readings frozen, which had worked on the Senvion farms, alarmed on 27% of turbine-days here. Verdict: FALSIFIED. The alarm limits, set with hindsight on that farm's own days, stayed at 1.8%.
Learning normal on the farm first
So the fourth test let every check learn what normal looks like on the farm itself, from its first two months, with no labels and nothing planted, as a product would on a new farm. The alarm limits learned the same way, from the same days. Then both were scored on the months after, on periods no test had read. Verdict: PARTIAL. The false alarms stayed under 3% on both farms, the first time in this series, and the checks found far more than the limits. What failed was one fault on one farm: an anemometer reading 10% low was found 19 times of 58 on Hill of Towie, where the turbines' wind wanders more than that from week to week.
| Farm | Auge alarms on real days | Limits' alarms | Auge found | Limits found |
|---|
| Hill of Towie wind farm | 1.9% | 7.9% | 83% | 34% |
| Kelmarsh wind farm | 2.8% | 5.1% | 99% | 54% |
A farm in France, with nothing changed
Then we took the method exactly as it was and ran it once on a farm none of the tests had read: La Haute Borne in north-eastern France, 4 Senvion turbines, published by ENGIE under the French government's open licence. ENGIE's portal is offline now, so we read a public copy and checked it against a second one. Verdict: PARTIAL.
1.0% of real days with an alarmagainst 2.8% for alarm limits learned from the same days
The false alarms held a third time. Finding faults did not. With 4 turbines and two winter months, the checks had only 170 turbine-days to learn from, and a limit from so few days is set by the most extreme of them. Here the limits came out wide: the wind check could only see an anemometer reading about 25% low, and the power check a turbine about 12% short of the farm's curve. So an anemometer reading 10% low was found 0 times of 60, and a turbine producing 10% less 9 times of 57. Learning normal on the farm keeps the false alarms down. It does not promise that a small fault can still be seen.
We also fixed the two misreadings the first test found, a lone letter read as volts and a gearbox bearing read as another part, before this farm's files were opened, and ran the naming test again on ENGIE's 30 documented signals. ENGIE writes short codes, like Gost for the gearbox oil sump temperature. The reader answered 5, right on 5 and wrong on 0, and refused the rest with a reason. The keyword list answered 0. Verdict for the naming: PARTIAL: not wrong once, but it answered too few to pass the mark for coverage.
The meter
One thing held wherever there was a meter to check against. From January to September 2019 Penmanshiel's turbines reported 1.13% more energy than its grid meter received, and every month lay between 0.97% and 1.42%: the size of loss that cables and a transformer between them would cause. At Hill of Towie the turbines' power over time came within 0.36% of the meter in 2019 and 0.60% in 2020. A month that steps out of that band is worth a phone call, and the console shows it.
What the console does with it
Turbines and the meter, in the console, shows each turbine's data, its power against the operator's curve beside the farm's other turbines, whether its two anemometers agree, and the farm's sum against the grid meter month by month. Its checks do what the fourth test did: they watch a farm's first 60 days, learn what is normal there, and flag only the days after. Next to the flags it says, for each check, the smallest fault it can still see on that farm, because the French farm showed that this can be large. Each flag gives the number that raised it and is a thing for a person to look at. It is not a verdict on the turbine.
Every result, with the criteria fixed before each run.