Temporal Conformal Prediction for Wind Turbine Anomaly Detection: A Robust Distribution-Free Framework with Conditional Clustering

This paper introduces a robust, distribution-free framework for early anomaly detection in wind turbine converters using adaptive conformal prediction. The method addresses the limitations of traditional monitoring systems, such as variable error magnitude, heteroskedasticity, and lack of finite-sample coverage guarantees, by generating statistically valid prediction intervals that adapt to evolving operating conditions. Our approach first clusters turbines based on operational pattern shapes to establish normal behavioral baselines. Within each cluster, we train regression models exclusively on healthy data and wrap them with a sequential conformal prediction procedure, providing 85% and 99% coverage intervals for warnings and critical alerts. Validated on 40+ turbines with 41 documented failures, the framework achieved a high recall rate (87.2%) and an average warning lead time of 72 days, significantly reducing false positives compared to baseline methods. The results demonstrate that conformal prediction enables reliable maintenance by quantifying uncertainty without restrictive distributional assumptions.