Scholarlink

<p>Accurate prediction of cascading faults in power grids can ensure system stability and prevent large-scale outages. Existing methods typically require learning effective statistical information from datasets with sufficient labels; however, in real-world scenarios, fault samples are scarce, leading to inaccurate fault identification. To address this, this paper presents a Hybrid Graph-Temporal Transformer (HGTT) to predict cascading faults in AC/DC hybrid power grids. The HGTT model integrates Graph Attention Networks (GAT) and Temporal Transformers, effectively capturing both spatial and temporal dependencies through an attention mechanism that accounts for electrical distances between nodes, as well as a causal attention-based temporal feature extraction module. Additionally, two self-supervised tasks are introduced to reduce reliance on labeled data. Experimental results show that HGTT achieves up to 14.8% higher accuracy than SVM, and reduces labeled data requirements by 50% under the self-supervised learning setting compared to fully supervised training.</p>

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

The research on cascading failure prediction in AC-DC hybrid power grids based on deep learning

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月30日

查看全文
  • Wenchao Qin,
  • Xufei Liu,
  • Yiran Tao

摘要

Accurate prediction of cascading faults in power grids can ensure system stability and prevent large-scale outages. Existing methods typically require learning effective statistical information from datasets with sufficient labels; however, in real-world scenarios, fault samples are scarce, leading to inaccurate fault identification. To address this, this paper presents a Hybrid Graph-Temporal Transformer (HGTT) to predict cascading faults in AC/DC hybrid power grids. The HGTT model integrates Graph Attention Networks (GAT) and Temporal Transformers, effectively capturing both spatial and temporal dependencies through an attention mechanism that accounts for electrical distances between nodes, as well as a causal attention-based temporal feature extraction module. Additionally, two self-supervised tasks are introduced to reduce reliance on labeled data. Experimental results show that HGTT achieves up to 14.8% higher accuracy than SVM, and reduces labeled data requirements by 50% under the self-supervised learning setting compared to fully supervised training.