An Integrated Python Framework for Wind Turbine Predictive Maintenance: LSTM-Autoencoder Anomaly Detection Benchmarked on the CARE-to-Compare Dataset with Digital Twin Power-Curve Modelling and Monte Carlo Uncertainty Quantification

This paper presents an integrated Python-based framework for wind turbine predictive maintenance, combining data-driven anomaly detection, physics-informed digital twin (DT) modelling, probabilistic uncertainty quantification, and an operational decision layer into a unified pipeline. Six anomaly detection models were systematically benchmarked on the CARE to Compare dataset (Zenodo DOI: 10.5281/zenodo.10958774), comprising 98 SCADA time-series files from three wind farms spanning 89 turbine-years. Evaluation used the four-dimensional CARE metric capturing Coverage (C), Accuracy (A), Reliability (R), and Earliness (E). The LSTM Autoencoder achieved the highest CARE score of 0.670 (C=0.612, A=0.845, R=0.741, E=0.305), outperforming Isolation Forest + PCA (0.627) and Dense Autoencoder (0.543). A hybrid Digital Twin combining a polynomial Normal Behaviour Model with a Betz-law reference was calibrated on 95 sub-datasets, revealing a mean Performance Ratio of 0.814 for anomalous turbines in Wind Farm C versus 1.002 for healthy units. Monte Carlo uncertainty quantification (N=1,000) produced 90% confidence intervals on power predictions. A rule-based four-level decision framework achieved a True Positive Rate of 0.929 on the evaluated subset.