Application of Convolutional Neural Network in Fault Line Selection of Distribution Network

Abstract

Aiming at the problem that the effect of the existing fault line selection methods is mainly determined by the fault features constructed by manual extraction, and the fault feature extraction process is complex and time-consuming, a new method based on convolutional neural network (CNN) is proposed. Firstly, the fault voltage and current signal data are collected, and the time-frequency energy matrix of the fault signal is constructed by the HHT band-pass filtering method, which is regarded as the two-dimensional matrix form of input data of CNN. Then the time-frequency energy matrix is input into the CNN, and the fault features are extracted autonomously through the convolution layer and pooling layer of the network, which is used to train the network to realize fault line selection and fault phase judgment. Finally, the results show that the method not only has a high accuracy of fault line selection but also can complete the fault line selection and fault phase judgment at the same time without adjusting any parameters, which realize the shared weights of the two non-independent problems. Meanwhile, under the influence of noise interference, compensation degree, network structure change, and other factors, the proposed method has good robustness. However, compared with the FCM, SVM, DNN, and DBN, CNN can still identify the faulty line and keep the optimal accuracy in the case of two-point grounding fault, and other methods have errors inline selection or low accuracy. Therefore, the experimental results also provide a new idea for fault line selection of distribution networks.

Keywords

• fault line selection
• time-frequency energy matrix
• convolutional neural network
• shared weights