Research On Predicion Method Of Wind Turbine Blade Icing Based On Residual Neural Network

Wind turbine blade icing faults directly affect the operational safety and power generation efficiency of wind turbines.To solve this problem,a method for wind turbine blade icing predicion based on XGBoost feature selection and one-dimensional residual neural network(1DRes-CNN)is proposed.First of all,in the data preprocessing stage,the data is labeled according to the normal operation of the wind turbine and the icing failure time period,and a part of the normal operation data is deleted through data visualization analysis.At the same time,the SCADA data is re-segmented using the time window method.After that,the data is standardized to eliminate the difference in the range of values between different features,and finally the SCADA data is data-balanced by the method of increasing the number of minorities by the SMOTE algorithm.Then,in the feature selection stage,the physical mechanism of wind turbine blade ice coating and the feature analysis method of SCADA data visualization are used to construct a series of invisible features to increase the feature information of the data.The XGBoost algorithm is used innovatively to automatically select the characterization of the wind turbine blade cover.The optimal feature subset of the ice state.Finally,in the selection of the predicion model,a one-dimensional residual neural network predicion model is constructed,and one-dimensional residual blocks are added to the deep neural network model.Through training and modeling of the data set,the detection of wind turbine blade ice coating is finally realized.Through experimental testing on the SCADA data set generated by two real wind turbines,the results show that the proposed model with selected features can increase the recall rate increased by 49.66%,comparing to the non-selected feature condition.In addition,compared with the support vector machine and the random forest model,the recall rate is increased by at least 10%,compared with the convolutional neural network,the recall rate is increased by about 3%,and comprehensive evaluation indicators such as accuracy,precision,and F1 value are used.Evaluate the diagnostic performance of the model.Experiments show that the model proposed in this paper can well identify the icing state of wind turbine blades.Finally,through experimental verification on the data of other numbered wind turbines in the same wind farm,it shows that the model also has a good performance.Diagnostic performance and generalization ability.The paper has 35 pictures,11 tables,and 60 references.