Machine Learning Prediction of Wind Turbine Blade Icing Considering Class Imbalance

Wind turbine blade icing is one of the most significant operational problems for wind farms in cold-climate regions, leading to substantial losses in electricity generation, increased mechanical loads, and hazardous situations. Accurate icing prediction is critically important for optimizing wind farm operation and applying preventive measures in a timely manner. This study aims to apply and comparatively analyze sampling methods for predicting wind turbine blade icing under conditions of class imbalance. Six approaches are examined: class weighting, SMOTE, ADASYN, Tomek links, SMOTE-Tomek, and the proposed TaA method. Unlike a simple sequential combination of Tomek links and ADASYN, the TaA method incorporates three novel components: an adaptive Tomek links cleaning threshold that accounts for the local density of the minority class; weighted synthetic sample generation (based on ADASYN) that considers the distance to the nearest majority-class instance; and an iterative cleaning and generation procedure. Feature selection based on the Gini criterion and hyperparameter optimization are employed. Experiments are conducted using real meteorological data from the Ulyanovsk wind farm and its surrounding area. The study addresses both the detection of current icing and forecasting with a horizon of up to nine hours. Three machine learning algorithms—Random Forest, Support Vector Machines (SVM), and Gradient Boosting—are evaluated, alongside a Long Short-Term Memory (LSTM) neural network. Temporal data splitting is used to ensure objective evaluation. Performance is assessed using a set of metrics robust to class imbalance: Precision, Recall, F1-score, Specificity, Balanced Accuracy, MCC, AUC-ROC, and PR-AUC. The results demonstrate that hybrid sampling methods outperform class weighting as well as the individual SMOTE, ADASYN, and Tomek links methods. For the Random Forest model, the best performance is achieved using the TaA method: compared to the baseline model without balancing, the F1-score increases 2.7-fold, and the PR-AUC improves by 42%. The TaA method statistically significantly outperforms SMOTE-Tomek in terms of F1-score for Random Forest and gradient boosting, whereas SMOTE-Tomek proves preferable for the LSTM network. The practical significance of this work lies in the potential to use the obtained results to improve the quality of icing forecasting at wind farms in cold regions.