A plant that produces less than the weather allowed has lost megawatt-hours, but the reason matters more than the number. If a grid operator or a price signal told the plant to throttle, that is curtailment: a constraint, not a fault. If the resource was available and the plant failed to convert it, that is underperformance. The two look identical at the meter and demand opposite responses.
Curtailment is a decision imposed on the plant; underperformance is a promise the plant broke.
Two kinds of missing megawatt-hours
A revenue meter records what was delivered, never what was possible. The shortfall against an ideal day has two distinct origins. Curtailment is a forced reduction the plant did not choose: a dispatch setpoint over the SCADA link, a frequency or voltage constraint, a thermal limit on the connection, or a negative price that makes generating uneconomic. Underperformance is everything that fails while the resource is there to convert: a tripped inverter, a derated string, a stuck pitch actuator, a fouled anemometer feeding the wrong wind speed to the turbine controller.
The tell is the setpoint. Curtailment leaves a record in the control system: an active-power limit asserted over Modbus or OPC UA, a market schedule, a transmission-operator instruction with a timestamp. Underperformance leaves no such order. Capturing the curtailment signal alongside the power reading, and storing both as aligned time series, is what lets you tell a throttled plant from a broken one after the fact.
Estimating available power, not guessing it
To quantify curtailment you need the counterfactual: what the plant would have produced with nothing holding it back. For wind, build available power per turbine from measured wind speed and the turbine power curve, over intervals where a curtailment setpoint was active. For PV, derive it from plane-of-array irradiance and a clear-running performance model, or from comparable unconstrained inverters on the same site. Lost energy is the integral of available minus actual over the curtailed period.
Two disciplines keep this honest. Attribute the gap to curtailment only when a constraint signal is genuinely present; otherwise you bury real faults in the curtailment bucket and stop investigating them. And validate the available-power model against unconstrained hours, because a counterfactual that overshoots in normal operation will overstate every compensation claim you file.
Why conflating them corrupts KPIs
Availability and performance ratio are meant to measure the plant's own behaviour. If curtailed hours count as downtime or as lost performance, every grid constraint drags those figures down and makes a healthy asset look faulty, masking the real faults you should be chasing. The standard fix is to treat documented curtailment as excused and report it separately from technical availability, so grid losses and equipment losses read as two distinct lines.
Forecasts fail the same way. A model trained on history that silently includes curtailed intervals learns that the plant underdelivers in conditions where it was simply told to stop, and bakes a downward bias into future output, straight into your intraday imbalance position. Flagging curtailed periods and either excluding or explicitly modelling them is the difference between a forecast that learns the physics and one that learns the grid operator's past instructions.
Separating curtailment from underperformance is not a reporting nicety; it decides which KPIs you trust, which faults you chase, and how much compensation you can defend. It rests on capturing the control signal next to the meter and modelling available power you can stand behind. That instrumentation and data layer is exactly what we build as part of our data & forecasting work.