Predictive Maintenance for Wind Turbines

The problem

Wind turbine failures are expensive twice over: the repair itself, and the generation lost while a crew reaches a remote installation. Scheduled maintenance addresses this by servicing on a calendar, which means replacing components that were fine and still missing the ones that fail early.

How it works

Models built with scikit-learn and time series methods analyse vibration, temperature and rotational data streaming from turbine sensors over AWS IoT, learning what precedes a component failure rather than applying fixed thresholds. Degradation shows up as a change in pattern well before it crosses any single alarm limit, which is where the forecasting horizon comes from.

What shaped it

Lead time is the entire value. A prediction made a day out is nearly useless when a crew, vessel or crane has to be scheduled, so the models were built to forecast far enough ahead to be actionable rather than to maximise raw accuracy on imminent failures.

Sensor data from turbines is also messy — dropouts, drift and weather all appear as anomalies — so distinguishing genuine degradation from environmental noise was a substantial part of the work.

Outcome

Component failures were forecast before they occurred, cutting unplanned downtime and maintenance cost.