A Method for Identifying Fault Branches in Power Systems Based on Current Waveform Representation and MAC-ViT
TL;DR
A Mask-Aware Convolutional Vision Transformer that encodes active power, reactive power, and root-mean-square current as topology-informed power-flow images and demonstrates fault-branch identification under the simulated networks and disturbance conditions evaluated in this study.
Abstract
Accurate fault-branch identification requires system-wide measurements to represent both localized disturbance patterns and interactions among electrically distributed components. This study presents a Mask-Aware Convolutional Vision Transformer (MAC-ViT) that encodes active power, reactive power, and root-mean-square current as topology-informed power-flow images. The model combines channel–spatial feature reweighting, attention-guided token suppression, and convolutional query–key projection to integrate local disturbance features with global interactions. On 4,440 simulated samples from the IEEE 9-bus system covering six line-disconnection classes, MAC-ViT achieved 99.77% accuracy, exceeding ViT-Tiny (95.56%), Conv-ViT (95.05%), and a current-retention-ratio-based undercurrent criterion (97.64%). Its accuracy ranged from 93.92% to 78.94% across four predefined levels of simultaneous Gaussian noise, impulse noise, harmonic distortion, and linear drift, and from 90.70% to 70.22% when complete measurements from one to three nodes were unavailable. On an IEEE 39-bus dataset containing 15 faulted-branch classes under three-phase-to-ground faults, the model achieved 97.00% validation accuracy. MAC-ViT contains 12.43 million parameters, requires 3.18 GFLOPs, and achieved an average GPU inference latency of 4.96 ms per image on the tested hardware. These results demonstrate fault-branch identification under the simulated networks and disturbance conditions evaluated in this study.