Due to compensation function of arc suppression coil

grounding current of distribution network is weak and protection is not easy to operate. In addition

traditional feeder detection method is not suitable for weak fault feature extraction. In order to improve accuracy and universality of fault feeder detection in distribution network

it established a mixed database containing three completely different topology models

and proposed a novel fault feeder detection method based on gramian angular field (GAF) and convolutional neural network (CNN). First

it used GAF to transform zero sequence current under different fault conditions

therefore

it obtained a large number of current characteristic images. Then

it trained CNN by mixed sample database including characteristic images. After determining CNN structure and parameters

CNN model is used to judge whether feeder is faulty or not. Finally

proposed method is verified by simulation data and field test data

and it has certain robustness to some disturbances

such as noise

data loss

asynchronous sampling

data proportion change

arc grounding fault occurrence.