Digital Twin for Wind-Turbine Predictive Maintenance

This project presents a Python-based applied research framework for predictive maintenance of wind turbines using SCADA time-series data. It combines data-driven anomaly detection, physics-informed digital-twin modelling, uncertainty quantification, and an operational decision-support layer designed for maintenance-oriented actions.

 
• Benchmarked six anomaly-detection approaches on the CARE-to-Compare wind-turbine dataset.

 
• Used 98 SCADA time-series files from three wind farms, representing 89 turbine-years of operation.

 
• Evaluated the models using the CARE metric, covering Coverage, Accuracy, Reliability, and Earliness.

 
• The LSTM Autoencoder achieved the best reported CARE score in the study: 0.670, with C=0.612, A=0.845, R=0.741, and E=0.305.

 
• Integrated a physics-informed Digital Twin based on a polynomial Normal Behaviour Model and a Betz-law reference power model.

 
• Added Monte Carlo uncertainty quantification with 1,000 simulations to produce 90% confidence intervals for power predictions.

 
• Included a four-level rule-based operational decision layer that translates anomaly and uncertainty outputs into maintenance-oriented actions.

 
• The evaluated subset reached a reported true-positive rate of 0.929.

The project is relevant to applied research in digital twins for renewable-energy assets, SCADA-based condition monitoring, uncertainty-aware predictive maintenance, interpretable power-curve modelling, and operational AI for wind-farm reliability.

Zenodo record:


https://zenodo.org/records/19038126