| Gas Insulated Switchgear nowadays is being applied especially widely in the electric power system owe to it's practical advantage with small shape,assemble conveniently,artistic and performance excellently,etc.Partial discharge in GIS is the common phenomenon which token some of insulation with plenty of significant information of working in GIS.In case a long term partial discharge usually would cause fault even accident in GIS.Therefore,it is of great significance to realize classification correctly and effectively of partial discharge in GIS,to monitor and diagnose the fault of GIS equipment.At present,the algorithm of SVM and BP neural network algorithm are frequently applied for discriminating partial discharge defect in GIS,which can effectively discriminate to a certain extent.However,due to the limitation of their training mechanism,its recognition ability is limited and there is not much space for improvement.In recent years,the Ensemble Learning method and the Deep Learning method in the field of pattern recognition have overcome the shortcomings of the aforementioned pattern recognition algorithm to a certain extent,and have been widely used in all walks of life recognition problems,as the benefits are good.Firstly,this paper analyses the research significance of partial discharge pattern recognition in GIS from the realistic security and economic point of view,and the necessity of intellectualization of on-line monitoring and recognition in GIS.In order to improve the application status of traditional pattern recognition methods in partial discharge defect recognition in GIS,this paper mainly studies the implementation of partial discharge pattern recognition in GIS by Ensemble Learning XGBoost and Deep Learning.On the basis of this research,a joint training method is proposed to further realize the high accuracy and strong robustness of partial discharge pattern recognition model in GIS.The research contents and results include:(1)In the aspect of Ensemble Learning XGBoost to realize pattern recognition algorithm of partial discharge in GIS,the characteristics of partial discharge in GIS based on UHF method are studied and analyzed: free metal particle discharge,surface discharge,internaldischarge and corona discharge.It includes the symmetry of discharge period and phase,phase offset,atlas symmetry and discharge amplitude distribution of each of the four discharges.Based on these,74 statistical features are extracted,including mean of discharge,dual rate of discharge,discharge width ratio and phase deviation of discharge.The difference of accuracy between decision tree and Ensemble Learning XGBoost algorithm for partial discharge defect recognition in GIS is studied and analyzed.The high availability of XGBoost algorithm with 92.25% recognition accuracy is demonstrated.(2)In the aspect of Deep Learning to realize partial discharge pattern recognition algorithm in GIS,the shortcomings and improvement mechanism of Deep Learning neural network are studied.Aiming at the shortcomings of adding network layers,such as reducing the convergence speed of training and recognition accuracy,the optimization algorithms of SGD,AdaGrad,RMSProp and Adam are deeply studied.At the same time,the idea of Early Stopping,Regularization and Dropout are introduced to control network over-fitting.(3)Through the training of Deep Learning,feature maps with representational significance are obtained,and combined with the aforementioned Ensemble Learning XGBoost,an union training model,called BUCM(Boosting Union CNNs Model),can achieve higher defect recognition accuracy and lower false detection rate in partial discharge defect recognition of GIS.At the same time,in order to solve the problem of sample imbalance and model reliability,algorithm of SMOTE and Generative Adversarial Networks(GANs)are introduced to generate a small number of samples and optimize deeply the accuracy of BUCM pattern recognition. |
PDF Full Text Request