7 / 2025-03-11 23:07:14

Graph Attention and Siamese Networks for Wind Turbine Blade Icing Prediction

wind turbine blades, icing prediction, few-shot learning(FSL), graph attention network(GAT)

Icing on wind turbine blades in winter can lead to power generation reduction and income loss. Timely and accurate prediction of blade icing status is crucial for improving the safety and economic performance of wind farms. However, existing blade icing prediction methods fail to effectively address issues such as small sample sizes, non-Euclidean structure, and low prediction accuracy in multi-dimensional sensor data. To address these limitations, this study proposes a GAT-BiLSTM-SN-based blade icing prediction method. at first, the uses recursive feature elimination(RFE) is used to extract features strongly correlated with blade icing. Then, the focal loss function is used to assign more weight to the lower proportion of ice samples, addressing the significant class imbalance between ice and no ice categories. Finally, the GAT-BiLSTM-SN model is used to to establish a blade icing prediction model, in which 15 hours of historical data is input and the next 15 hours of icing status is output. Experimental results show that the GAT-BiLSTM-SN model achieves accuracy, precision, recall, and F1 scores of 0.9107, 0.9434, 0.8772, and 0.9091, respectively, in few-shot learning scenarios.