Scholarlink

A lightweight deep learning model is proposed for detecting DC series arc faults in residential PV energy storage systems (PESSs). It employs a dual-branch architecture consisting of a multi-scale attention-based convolutional neural network and a two-layer gated recurrent unit. The model takes multimodal input, consisting of raw current sequences and gramian angular summation field (GASF) representations. And it utilized a cross-attention module to fuse temporal and GASF representations. The performance of the model is tested on a PC with Jetson Nano-level computational capability. The accuracy and inference delay are 99.64% and 44.5 ms per sample, respectively. The model size is only 0.30 MB. These results demonstrate high detection accuracy and strong potential for edge deployment. It can be used to develop the arc fault circuit interrupters.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Lightweight CNN–GRU Model of Arc Fault Detection Based on Multimodal Feature Fusion Algorithm

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Zhiyong Wang,
  • Li Tang,
  • Fengyi Guo,
  • Xingquan Lv

摘要

A lightweight deep learning model is proposed for detecting DC series arc faults in residential PV energy storage systems (PESSs). It employs a dual-branch architecture consisting of a multi-scale attention-based convolutional neural network and a two-layer gated recurrent unit. The model takes multimodal input, consisting of raw current sequences and gramian angular summation field (GASF) representations. And it utilized a cross-attention module to fuse temporal and GASF representations. The performance of the model is tested on a PC with Jetson Nano-level computational capability. The accuracy and inference delay are 99.64% and 44.5 ms per sample, respectively. The model size is only 0.30 MB. These results demonstrate high detection accuracy and strong potential for edge deployment. It can be used to develop the arc fault circuit interrupters.