Prediction of ice accretion on wind turbine blades using spatiotemporal features and siamese networks

Abstract:

The accurate prediction of wind turbine blade icing is essential for the safe and stable operation of wind power systems. To address the challenges of insufficient feature extraction, unclear distribution of multi-dimensional sensors, and class imbalance in few-shot learning scenarios, this paper proposes a prediction method based on a residual graph attention network-bidirectional LSTM-siamese network (ResGAT-BiLSTM-SN). First, the supervisory control and data acquisition (SCADA) data is processed through data cleaning, sliding-window sampling, and feature engineering, resulting in a blade-icing dataset suitable for various few-shot learning tasks. Second, based on 19 key variables, a non-fully connected undirected graph is built using mutual information (MI) and a weight matrix to capture the spatial distribution and correlations among sensor data. By integrating the graph Attention network (GAT) and the bidirectional long short-term memory network (BiLSTM) to extract spatiotemporal features, the ResGAT-BiLSTM-SN model is developed to perform 24-hour-ahead icing prediction on the constructed dataset. Simulation experiments are conducted using the data from turbines No.15 and No.21 provided by the 2017 Industrial Big Data Innovation Competition platform. The experimental results show that the ResGAT-BiLSTM-SN model achieves F1 scores above 0.9 across three few-shot learning scenarios, significantly outperforming other baseline models. Compared to the GAT-BiLSTM-SN model, the proposed model demonstrates clear improvements in predictive performance, validating its effectiveness and superiority.