Research On Single-phase Grounding Fault Diagnosis Method Of Medium Voltage Distribution Network

The 10 k V medium-voltage distribution network plays an important role in the entire power system facing the user terminals.However,with the development of cities and the deepening of the rural revitalization strategy,the scale of the distribution network has been expanding year by year in recent years.The number and length of its lines are gradually increasing.Among them,the single-phase grounding fault accounts for the highest proportion,because it has the characteristics of running with faults,which is easy to cause the expansion of faults.Therefore,when a fault occurs in the medium-voltage distribution network,it is necessary to identify the fault as soon as possible and take corresponding measures to deal with it quickly and safely in order to avoid further expansion of the fault.After introducing the research status of grounding faults in distribution networks at home and abroad,the development trend of identification methods for single-phase grounding faults in distribution networks is summarized in terms of the extraction method of fault feature and the selection of intelligent classifiers.The advantages of Hilbert-Huang Transform(HHT)and deep learning algorithms for feature extraction and fault classification and identification of nonlinear and non-stationary fault signals.Through the analysis of the transient and steady state characteristics of the single-phase grounding fault in the distribution network,combined with the transient and steady-state fault characteristics of the distribution network,the deep learning algorithm is used to identify the single-phase grounding fault.After analyzing the steady-state and transient principles of single-phase grounding faults,this paper builds a 10 k V medium-voltage distribution network simulation model based on the matlab/simulink power system simulation module,aiming at metallic grounding,low-resistance grounding,high-resistance grounding,and stable arc grounding and intermittent arc grounding fault types,extract the bus three-phase voltage,three-phase current and zero-sequence voltage on the line in the first two cycles of the five types of faults and eight cycles after the fault.In this paper,the HHT time-frequency analysis method is proposed to process the data,construct block time-frequency picture samples,and give an example of the time-frequency decomposition of metallic grounding faults,which shows the good decomposition effect of HHT on grounding faults in distribution networks.First,for the identification of five types of single-phase-to-ground faults in distribution networks,this paper proposes a supervised learning Convolutional Neural Networks(CNN)identification method,and constructs a 7-layer convolutional neural network model.After training the CNN model through the training set,use the test set to test its recognition accuracy and compare the accuracy with the Support Vector Machine(SVM).The results show that CNN has a higher recognition accuracy than SVM.Then,aiming at the problem that the stable arc and the intermittent arc are not easy to distinguish in the transition stage when the neutral point is input through the arc suppression coil,this paper proposes an unsupervised learning Convolutional Deep Belief Networks(CDBN)identification method,and constructs a 7-layer convolutional deep belief network model.The comparison shows that under the same network structure,the average recognition accuracy of CDBN is not only higher than CNN,but also CDBN is better than CNN in arc recognition,and its adaptability is better than CNN.At last,this paper designs a MATLAB-GUI diagnostic platform,which integrates the medium-voltage distribution network simulation model,the HHT time-frequency decomposition algorithm,and the CDBN model for visual man-machine interface operations.By setting different fault parameters,different types of single-phase grounding faults can be diagnosed.The interface is intuitive,the operation is simple,and it has practical engineering application significance.Finally,the experimental data is obtained through the physical experiment platform to fill and test the sample capacity.The experimental results show that both CNN and CDBN have improved the recognition rate,which proves that the recognition ability of deep learning can be enhanced by enriching the diversity of fault samples.