Wind turbine blades icing failure prognosis based on balanced data and improved entropy

To improve the accuracy of icing failure prediction, which is often limited due to unbalanced condition data, a novel balancing algorithm based on boundary division synthetic minority oversampling technology (BD-SMOTE) and a method for predicting the icing failure of wind turbine blades in the short term based on multiple neural network combination are presented. First, the original data set obtained by sensors is balanced by BD-SMOTE. Then, the key features are extracted by multivariate and multiscale entropy based on a continuous smooth coarse (CSMMSE) algorithm, and the values of three kinds of features in the near future are predicted by the Elman neural network (ENN). Finally, a back-propagation (BP) neural network is adopted to predict the icing failure of wind turbine blades. Compared with the results of other methods, the prediction deviation of the ENN is smaller; the prediction results demonstrated the effectiveness and superiority of the proposed method.