Quiet is not the same as calm

Three situations look almost identical in a production series. The machine is degrading. The machine is curtailed under a grid, noise, or environmental constraint. The machine is in a period of genuinely low wind resource. All three produce a run of low ten-minute values, and any analysis that cannot separate them will confidently report the second and third as faults.

Curtailment is the one that catches people, because it is systematic. A noise regime that limits output on summer evenings when the wind is from the south produces a group of machines that underperform together, at a predictable time, for a defensible reason. To a model with no knowledge of the regime, that is a strong, repeatable, entirely spurious finding, and it will be a strong finding every summer.

The status and alarm log is what resolves it, which is why it belongs in the SCADA record rather than among the optional extras. Operating state codes distinguish a machine that was asked to produce less from a machine that failed to, and a manual stop from a technical one. Without them, every hour of every constrained period is scored as a deficit, and the resulting list is dominated by decisions the operator made on purpose.

The residual case is the interesting one: low production, normal state codes, no curtailment regime, and neighbours that are producing. That combination is what a genuine finding looks like, and it is only visible once the first two explanations have been removed. This is the practical reason SCADA data has to arrive unfiltered - an export with curtailed periods stripped out looks cleaner and has had the evidence removed that would have exonerated a machine.

It is worth conceding how far this can be pushed. Not every constraint is logged. Undocumented regimes exist, particularly where a noise or shadow-flicker limit was implemented years ago in a controller setting that nobody recorded centrally, and no amount of modeling recovers a fact that was never written down. What a model can do is notice that a group of machines behaves identically under a specific set of conditions and flag the pattern as systematic rather than as five separate faults, which is a prompt to go and find the regime.

That prompt is worth taking seriously in both directions. On several occasions the honest interpretation of a repeating group deficit is not a fault and not a documented constraint, but a constraint that is being applied more aggressively than the operator believes. That is a finding about the control configuration rather than about the machines, and it is not something an alarm log will ever raise on its own.

The concrete step: before analyzing anything, assemble the curtailment regimes for each site - grid, noise, shadow flicker, environmental - with their trigger conditions and the dates they were in force. It is unglamorous, it usually requires talking to two different people, and it removes the largest single source of false findings in wind performance analysis.