Connect, model, act. No new hardware.

Curtailment is held out first. A curtailed hour looks exactly like severe underperformance in the raw record, and a model trained through one learns curtailment as normal behavior and then reports its absence as a gain. The same applies to forced outages, manual interventions, and icing periods on cold-climate sites.

Sensor faults are treated as findings rather than as noise. A drifting nacelle anemometer is one of the most common causes of an apparent loss that is not there, and it is also one of the most common ways a real loss stays hidden - the machine underproduces and reports less wind at the same time, so the pair looks consistent. Cross-checking a turbine's own wind measurement against its neighbors is part of the analysis, not a precondition for it.

The record is not cleaned before it is sent. Gaps, flat lines, and outliers identify a sensor fault or an unrecorded outage, and removing them upstream removes the evidence along with the mess. It is easier to reconstruct a messy export than to recover one that has already been tidied.

For each turbine, a normal behavior model learns what a channel should read given the conditions driving it: power from wind speed, direction, air density and turbulence intensity; gearbox oil temperature from load, ambient temperature and rotor speed. The model is trained on a period the machine was healthy, then run forward.

The residual - measured minus expected - is what carries the information, and the construction is what makes it interpretable. A raw temperature rising in July says nothing. A temperature five degrees above what the model expects for that load and that ambient says a great deal, and a residual that grows steadily across six weeks is a different object from a reading that is merely high on a warm afternoon.

This is why the comparison is against the turbine's own history rather than against a datasheet power curve. The published curve describes a new machine in standardized air on a test site; measured against it, almost every asset in a mature fleet appears to underperform, and a finding that applies to everything directs attention to nothing.

Each finding names the turbine, the loss mechanism, the period the deviation was observed over, the annual energy at stake, and the evidence behind the estimate. Where a price basis is available, the energy is also expressed in revenue, because a work list sequenced in megawatt-hours and a budget argued in dollars are otherwise two conversations.

Confidence is stated rather than implied. A production deficit is an inference from a model, not a counted quantity, and a figure presented without that distinction will not survive the meeting it was produced for. The ranking is inspectable for the same reason: an operations manager who disagrees with the weighting should be able to see the effect of a different one.