Wind turbine blade icing prediction based on CNN-BiGRU with Optimized training strategy

Wind turbine blade icing significantly impacts the safety of wind farms and the efficiency of power generation, making precise and timely prediction a critical challenge. This study proposes an innovative deep learning framework that integrates Convolutional Neural Networks (CNNs) and Bidirectional Gated Recurrent Units (BiGRU) to enhance icing prediction. CNNs extract spatial features, while BiGRU captures temporal dependencies, enabling the model to effectively distinguish icing occurrences. To improve model optimization, cosine annealing was employed for dynamic learning rate adjustment, while cross-entropy loss was used to address class imbalance. Experimental results demonstrate that a 2-layer CNN architecture trained over 50 epochs achieves a balance between accuracy and computational efficiency, with CNN_2Layer-BiGRU attaining 96.55% accuracy and a 96.51% F1-score, outperforming traditional models. This approach reduces dependency on manual feature engineering, improves prediction accuracy and computational efficiency, and provides a foundation for an intelligent diagnostic system for wind turbine blade icing prediction.