Icing detection and prediction for wind turbines using multivariate sensor data and machine learning

Adverse weather events can significantly compromise the availability and economics of a wind farm. This paper focuses on rotor icing detection, which constitutes a major challenge in wind farm operation. When ice accumulates on wind turbine blades, it causes substantial generation losses, operational disruptions, and safety hazards to the personnel, assets, and equipment in a wind farm. Alerts about early signs of rotor icing can assist operators in proactively initiating icing mitigation measures. To this end we propose a machine-learning-based framework that effectively learns the unique signatures of icing events. The framework effectively extracts salient features by condensing the multivariate turbine sensor data into a small-sized subset of information-rich descriptors. Those, along with power-curve-derived features, are used to train a deep-learning-based model that flags icing events and estimates icing probabilities. We also propose a new loss measure, called the icing power loss error (IPLE), that realistically quantifies the expected icing-related power losses. Our experiments show that the proposed framework achieves up to 96.4% accuracy in flagging icing events, while keeping the number of false alarms at minimum. When compared to prevalent data-driven benchmarks, up to 18.7% reduction in power loss estimation error is realized.