Prediction of wind turbine blade icing fault based on selective deep ensemble model

Abstract

In advance accurate prediction of wind turbine blade icing fault is of fundamental importance. Deep learning is the mainstream prediction technique while existing research about the prediction of wind turbine blade icing fault has primarily derived a single deep learning model. This study introduces the group method of data handling (GMDH) technique and proposes a GMDH-based selective deep ensemble (GSDE) model. First, the model combines the convolution neural network (CNN) with recurrent neural network (RNN), long short-term memory (LSTM), and gated recurrent unit (GRU) to construct CNN–RNN, CNN–LSTM, and CNN–GRU, respectively. Together with CNN, four cost-sensitive deep neural networks based on focal loss are formed and used as the basic prediction models. Second, a series of training sets are constructed by the Chi-square test. Four different basic prediction models are trained on each training set, and the prediction results of all base predictors are obtained. Third, the GMDH technique is applied to the cost-sensitive selective deep ensemble for final prediction results. Experiments are conducted to deeply verify the prediction performance of the GSDE model on two wind turbine datasets collected by the supervisory control and data acquisition (SCADA) system. Results show that the proposed model outperforms five existing ensemble models and five single deep learning models.