Research On Fault Line Selection Of Distribution Network Based On Transfer Convolutional Neural Network

Most of the 6～35kV medium voltage distribution network in China are small current grounding system,and the probability of single-phase grounding fault is as high as 80%,the fault should be cut out in time because of the high demand of users for power supply quality,which makes fault line selection of distribution network undoubtedly become a hot research issue.However,due to the extensive access of cable lines,the compensating effect of arc-suppression coils and the variable operation mode of the system,it not only increases the difficulty of fault line selection,but also makes it difficult to guarantee the reliability and accuracy of fault line selection by the traditional methods relying on manually extracted fault features.Therefore,it is necessary to further explore the problem of fault line selection,which is of great significance to the safe and stable operation of distribution network.A transfer convolutional neural network(T-CNN)method driven by small sample data is proposed to realize fault line selection based on the lack of tagged fault data in actual distribution network in this thesis.Firstly,the characteristics of the grounding fault in the distribution network are analyzed,and it is determined that transient signal components is the main signal of representing the fault line characteristics.Secondly,a small current grounding system model are built and the fault conditions are set,the fault data set required for network training is obtained by simulation,and aiming at the two-dimensional matrix form of the input data of the network model,the time-frequency energy matrix of the fault signals is constructed by Hilbert time-frequency band-pass filtering method.Then,due to the limited learning ability of convolutional neural network(CNN)in small data set,the transfer learning method is used to fine-tuning training CNN to obtain the T-CNN model,and the model is suitable for the learning of small data set.Finally,combined with relevant theories and parameters,a 10 kV simulation model of distribution network is built and the complex environment of the actual distribution network is simulated,which analyze the advantages of transfer learning method in small sample data set and the performance of line selection of T-CNN model through simulation.The simulation results show that the proposed method can solve the time-consuming and complex problems of manual feature extraction with the help of the adaptive feature learning ability of network,and the T-CNN model is not constrained by the fewer data and the complex network structure and can still correctly judge the faulty line,and it has higher accuracy and faster speed than the CNN model without transfer learning and is less affected by transition resistance,noise and other factors,and the deep belief network(DBN)model and traditional classifier(support vector machine)are used for classification comparison to further verify the advantages of this model in line selection under a small number of training samples.In addition,the introduction of transfer learning under small sample data set also provides a new idea for the study of other aspects of distribution network.