The anemometer is part of the machine

Every performance calculation on a wind farm rests on a wind speed, and on most turbines that wind speed comes from a cup or sonic anemometer mounted on the nacelle, downstream of the rotor. The rotor extracts energy from the flow and adds turbulence to it, so the instrument is measuring a disturbed field rather than the free stream, and the relationship between the two is corrected by a transfer function that was established for the turbine type rather than for that specific machine on that specific site.

The consequence is a per-machine bias that is stable, invisible, and directly in the path of every conclusion. A turbine whose anemometer reads two percent low will appear to produce more than expected at any given wind speed. One that reads two percent high will appear to underperform, permanently, with no fault anywhere in the machine. Since power goes roughly with the cube of wind speed at partial load, a two percent error in the input is a six percent error in the expectation.

This is the single strongest practical argument for fitting an expectation to each turbine's own history rather than to a fleet-wide or datasheet power curve. A model built from that machine's own record learns the bias as part of the machine's normal behavior, and consequently stops reporting it. What it then detects is a change from that normal, which is what an operator actually wants to be told about.

The complication is that the instrument itself changes. Anemometers are replaced, recalibrated, and occasionally iced, and each of those introduces a step in a series that a naive model reads as a change in the machine. A sensor replacement in March and a bearing degrading from April look similar in a residual plot, and the difference between them is not statistical - it is that somebody has a maintenance record of the first one. This is why maintenance history is worth supplying even though nothing formally requires it.

It follows that some findings are about instruments rather than about turbines, and those should be reported as such. A machine whose apparent performance steps by three percent on a single date, with no corresponding change in any mechanical channel, is telling you about its sensor. Reporting that as a performance loss would be technically defensible and operationally useless, because it sends a crew to inspect a drivetrain that is fine.

The honest limit is that no analysis of nacelle data can establish the true free-stream wind speed. Where that matters - a warranty claim, a formal power curve verification, an energy assessment for refinancing - the answer is a met mast or a lidar campaign, and any vendor implying otherwise is overselling what an operational dataset contains. What operational data can do is detect change and rank machines against each other, which covers most of what day-to-day operations needs.

A useful diagnostic: plot each turbine's apparent performance residual as a monthly series over two or three years and look for steps rather than trends. A step with a maintenance record beside it is an instrument change. A step without one is a question worth asking, and a slow trend is a different question entirely.