Research On Icing Detection Methods Of Wind Turbine Blades Based On Machine Learning

As a kind of clean energy,wind energy plays an important role in the production of human society.The wind turbine is an important device to capture wind energy.Most of wind turbines are located at a high altitude,and their blades are prone to be covered by ice when the temperature is low.Ice on the blades will reduce the utilization rate of wind energy,shorten the service life of equipment,and bring many safety risks.With the arrival of industrial big data era,it has become the trend of development to use massive data to mine equipment information to improve production efficiency.The purpose of this paper is to use the operation data of wind turbine and machine learning theories to develop effective methods for icing detection of wind turbine blades.The specific works are as fllows:Firstly,in order to solve the problem that the data-driven feature extraction method is difficult to effectively utilize the information of icing mechanism,this paper proposes a feature extraction method based on icing mechanism and data-driven strategy.Based on the data information,this method adds the information of icing mechanism,and further mines the deep physical features behind the blade icing.Under the same conditions,the F1 score corresponding to this feature extraction method is on average 7.72% higher than that corresponding to the data-driven feature extraction method.This method can effectively extract the features from the wind turbine operation data.Secondly,in order to solve the problem of poor generalization ability of the trained model caused by too unbalanced distribution of positive and negative samples in the wind turbine operation data,this paper proposes Easy-SMX algorithm,which is based on Synthetic Minority Oversampling Technique and integrated undersampling method.The algorithm can not only avoid overfitting caused by simply copying samples from the class with few samples,but also reduce the information loss of majority class.On the test set,the F1 score of this algorithm reachs 91.39%,which is 12.03% and 8.08% higher than that of Easy Ensemble algorithm and SMOTE,respectively.The algorithm can effectively detect the icing of wind turbine blades.Finally,in order to solve the problem that extracting features manually from multidimensional time series signals depends too much on professional knowledge and the extraction process is complex,this paper proposes 2D-CNN-LSTM network,which is used for icing detection of wind turbine blades.This network not only has CNN’s powerful ability of spatial feature extraction,but also has the advantages of LSTM to capture the long term dependence within and between sequences.On the test set,the F1 score of this network is as high as 95.67%,which is 9.22% and 6.25% higher than that of 2D-CNN and LSTM,respectively.This network can well complete the task of icing detection of wind turbine blades.