The average that hides the loss
A site reporting a healthy capacity factor can be carrying three turbines that are each four percent down. The arithmetic explains why nobody notices.
Start with the arithmetic, because it settles the question faster than any argument about dashboards. Take a forty-turbine site. One machine is producing four percent below what it should. Spread across the site, that turbine contributes one fortieth of the total, so the site-level deficit is one tenth of one percent. Three such machines make it three tenths. Against a monthly capacity factor that moves five to ten percent with the weather, a three-tenths deficit is not a small signal. It is no signal at all.
This is the structural reason underperformance survives conventional reporting. It is not that operators are inattentive; it is that the metric they are given cannot resolve the thing they would need to see. A site average is an instrument with a resolution, and the resolution is worse than the effect by an order of magnitude. Adding a chart does not change what the number can distinguish.
The obvious response is to compare each turbine to itself over time, and it fails for a specific reason. The wind a machine saw last March is not the wind it saw this March, so a year-on-year comparison confounds the machine with the weather. Improve the comparison by normalizing against the manufacturer's power curve and the problem changes rather than resolving: that curve describes a new turbine under test conditions, so every machine on a mature site sits below it, and a genuine four percent loss disappears into the offset everything else already has.
What does resolve it is comparing a turbine against machines operating in materially the same conditions at the same time. The weather is then common to both sides of the comparison and cancels, which converts a question about absolute output into a question about a difference, and a difference of four percent is well outside the noise. This is why the unit of analysis has to be the turbine rather than the site, and why the reference has to be the fleet rather than a datasheet.
The steelman for site averages is real and worth stating. They exist because settlement, contractual availability, and shareholder reporting all operate at site level, and those are the numbers a business actually runs on. Nobody should stop producing them. The point is narrower: a metric built for commercial reporting was never designed to detect a four percent deviation on one asset, and using it that way is asking an instrument to do something outside its range.
The limit of the comparative approach is worth naming too. It needs comparable machines. A single-turbine site, a machine in genuinely unique terrain, or a unit whose nearest siblings sit deep in a wake are all cases where the reference is weaker, and a finding on those should carry lower confidence rather than the same confidence with a caveat. Where we cannot build a defensible expectation, the right answer is to say the turbine is out of scope.
There is a test worth running on data an operator already holds. Take twelve months of ten-minute SCADA data, bin it by wind speed and direction, and rank the turbines by their median output within each bin rather than by their annual totals. Machines that are consistently in the bottom of their own bins, across conditions, are not unlucky. That ranking takes an afternoon, it uses nothing that has to be bought, and it usually surprises people, which is the most useful property a diagnostic can have.