Wind turbine blade icing diagnosis using B-SMOTE-Bi-GRU and RFE combined with icing mechanism

Low temperature and high humidity in winter can easily lead to blade icing, which severely impacts the actual power output and safe operation of wind turbines. However, existing blade icing diagnostic methods suffer from inadequacies in feature extraction and the accuracy. Therefore, we propose an efficient and accurate blade icing diagnosis method. Initially, we develop hybrid features that include both short-time and long-time features. RFE is combined with icing mechanism to extract short-time features, and the sliding window algorithm is used to extract long-time features. Secondly, the B-SMOTE method is used to synthesize minority icing samples. Finally, the B-SMOTE-Bi-GRU model is applied to validate the superiority of the proposed method. Compared to the features extracted based on the icing mechanism and the sliding window algorithm, the average improvements in recall, TNR, and accuracy for each wind turbine is between 0.26% and 4.87 %, 0.23%–4.60 %, and 0.05%–2.73 %, respectively. In comparison to the other models, the B-SMOTE-Bi-GRU method proposed in this paper enhances the recall and TNR of each wind turbine by between 0.06%-13.10 % and 0.07%–12.01 %, respectively. The results indicate that the hybrid features and B-SMOTE-Bi-GRU model proposed in this paper offer a more effective and accurate approach to blade icing diagnosis.