Accurate fault identification in distribution networks is fundamental for fault detection

location

and line restoration

which are essential for ensuring the reliability of power supply to users. To address issues such as the susceptibility to local measurement point influence and the poor generalization ability of existing methods

this paper proposes a fault identification method for distribution networks based on spatio-temporal data cubes. It analyzes the differences in propagation characteristics of different types of faults and disturbance signals

and demonstrates the superiority of using panoramic fault data from the entire network. The approach decomposes fault signals into time-frequency spectra and stacks spectra from different measurement points to construct spatio-temporal data cubes. Finally

a convolutional neural network (CNN) model based on 3D convolutional kernels is proposed. The spatio-temporal data cube serves as the basic data unit input to the network model

and feature extraction and error identification are achieved through convolutional operations of 3D kernels. Through extensive simulations and field tests

the method proposed in the paper can accurately distinguish different types of faults and disturbances

has a high accuracy in identifying high-resistance grounding faults

and demonstrates strong noise resistance.