Predictive maintenance and power forecasting for wind farms
Hourly power forecasts from turbine SCADA data, detection of turbine-hours that deviate from the expected power curve, and the resulting loss measured in MWh.
01
What makes it hard in the field
Variable output
Fluctuating wind speeds directly affect production planning and imbalance costs.
Hidden underperformance
Turbines that raise no fault alarm but run below their power curve lose energy unnoticed.
Messy data
SCADA records contain missing, frozen or faulty sensor readings that must be cleaned before analysis.
02
Our approach
- AData cleaningDowntime and missing or frozen readings are separated out; a consistent time series is built for each turbine.
- BForecasts with intervalsTree-based models with weather forecast features extrapolated to hub height; every forecast comes with an interval targeting 90% coverage.
- CAnomalies and lossTurbine-hours deviating from the expected power curve are detected without labeled data; the loss is computed in MWh.
- DMaintenance decisionsAlarms are linked to physical fault classes and turned into work orders; technician findings are written back to the model.
03
Evidence
- Tested on open data: ENGIE La Haute Borne SCADA, 2014–2015
- Benchmarked against a persistence forecast
- Interval coverage reported
FAQ
- What data do you work with?
- Turbine SCADA records (typically 10-minute averages: wind speed, active power, rotor speed, temperatures) and, optionally, weather forecasts. No new sensors are needed to start.
- How can we see how accurate the forecasts are?
- We tested the method on an open wind farm dataset; error versus a persistence forecast and interval coverage are published on the case study page. During a pilot on your farm the same metrics are reported.
- Do we need labeled failure data?
- No. Anomaly detection is based on deviation from expected behavior; technician findings strengthen fault classification over time.
A pilot conversation for your wind farm
Let's discuss the structure of your SCADA data and the number of turbines, and agree how to set up a pilot on a selected group of turbines and what to measure.