AI-enhanced diagnostics for manufacturer-independent validation of wind turbine performance: Results from the WindKI project

Accurate, manufacturer-independent assessment of wind turbine performance remains a key challenge with direct implications for energy yield, as conventional power-curve analyses are often biased by uncertainties in nacelle-mounted wind measurements. This paper presents an anomaly detection framework developed within the WindKI research initiative, which aims to improve wind turbine performance diagnostics using data-driven methods. The proposed framework combines multiple unsupervised anomaly detection models with a rank-based ensemble strategy to identify statistically unusual operating behavior without requiring labeled training data or turbine-specific tuning. The approach is evaluated using the CARE dataset, comprising predefined normal and anomalous events from three wind farms. Results show consistent prioritization of anomalous events across sites, with ROC–AUC values of approximately 0.8, indicating robust performance despite the unsupervised setting. The framework provides a scalable foundation for diagnosing underperformance and investigating potential failure modes in wind energy assets.