Energy-Efficient Wind Turbine Operation via Multi-Class Blade Icing Recognition Using Mutual-Information Feature Selection and Ensemble Methods

Icing on wind turbine blades not only degrades aerodynamic efficiency but also poses a significant risk to structural integrity. However, conventional sensor-based or image-based detection systems remain costly and prone to weather-related interference. This study introduces a lightweight, multi-class diagnostic framework designed to classify turbine states as normal, icing, or other anomalies using native SCADA (Supervisory Control and Data Acquisition) data. After standardizing and balancing the dataset, key features were selected using three complementary techniques: Support Vector Machine (SVM) feature importance, Random Forest feature importance, and Mutual Information ranking. Mutual Information is particularly utilized to identify the non-linear dependencies between features and turbine states, enhancing the feature selection process. Five ensemble learning models—XGBoost, LightGBM, Gradient Boosting, Extra Trees, and AdaBoost—were benchmarked. Among these, the Extra Trees model provided the optimal trade-off between accuracy and computational efficiency, achieving an F<sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sub> score of 99.54% while maintaining minimal training and inference time. Due to its strong generalization ability and low inference latency, this framework can be efficiently deployed on edge computing platforms or directly within SCADA systems. It offers a scalable, real-time solution for wind farm icing detection and maintenance decision-making, paving the way for cost-effective, rapid-response systems for proactive wind turbine management.