The power curve is not a baseline

A power curve is a warranty instrument. It records what a specific turbine model produced under controlled test conditions, on a certified measurement campaign, with an instrument set nobody operates a fleet with. It is exactly the right document for accepting a machine at commissioning, and it is a poor reference for judging that same machine in its eleventh year on a site with complex terrain.

The discrepancy is not subtle. Air density is not the test condition's, turbulence intensity is not the test condition's, and the anemometer supplying the wind speed is mounted behind a rotor rather than on a met mast. Each of those shifts the measured curve independently, and none of them is a fault. So an operator comparing against the datasheet sees every machine on the site sitting below the line, concludes the offset is normal, and calibrates their expectation to it. That calibration is where a real four percent loss goes to hide.

It is worth being precise about what goes wrong, because the failure is not that the curve is inaccurate. The curve is accurate about the thing it describes. The failure is that the difference between a datasheet curve and an operating machine contains at least four terms - siting, degradation, instrumentation, and control configuration - and a single number cannot be attributed to any one of them. A deficit that could be caused by four things is not evidence for any of them.

The alternative is to build the reference from the machine's own history. What a specific turbine produced across the conditions it actually experienced is a description of that turbine including its site, its instrument bias, and its configuration, which means a departure from it is a change in the machine rather than a difference from a laboratory. Instrument bias in particular stops being a confounder and becomes part of the model, which is the single largest practical gain from fitting per turbine.

The objection to this is legitimate: a model fitted to a machine's own history will happily learn a fault that was present throughout that history, and then report the machine as healthy. This is real, and it is the reason a second reference is needed. Comparison against machines operating in the same conditions catches the case where a turbine has always been the weakest one, which a self-referential model structurally cannot. Neither reference is sufficient alone, and a finding should state which one is carrying it.

Nor does any of this abolish the power curve. Warranty and acceptance testing are contractual processes with defined methods, and a performance model does not replace them. What it replaces is the informal use of the curve as an everyday operating baseline, which was never what it was published for.

A question worth putting to any vendor selling performance analytics: what is a turbine measured against, and how does that reference handle a machine whose anemometer has read two percent low since it was replaced? If the answer is the manufacturer's curve, the product will find gross faults and miss the losses that actually accumulate. The channels that make the alternative possible are the ones a fleet already records, and how it works sets out what is done with them.