Scholarlink

<p>The integration of digital technologies into smart grids enhances their efficiency but also introduces significant vulnerabilities, particularly in reconfigurable topologies where dynamic structural changes challenge conventional monitoring systems. This paper addresses the problem of detecting operational anomalies in such environments without relying on labeled data. We propose a label-free hybrid Deep Graph Representation Learning (DGRL) pipeline that combines unsupervised representation learning (GAT + VGAE) with pseudo-label-based semi-supervised classification (GCN), requiring no human annotation at any stage. A simulation laboratory was set up using MATPOWER and MATLAB to simulate IEEE 14, 69 and 300 bus systems under normal case, and network attack or reconfiguration. Raw structural and behavioral patterns were learned from unlabeled data, used reconstruction error to detect anomalies, and applied adaptive thresholding to pseudo-labels for GCN classification. Based on the generated data, this method achieved robust anomaly detection and stability to changes in the topology with 98% of accuracy. The economic evaluation revealed a Return on Security Investment of 47.3, which underlined the cost-effectiveness. This work provided a scalable and label-free solution for enhancing the operational monitoring of critical infrastructure, with the public release of the generated datasets to support further research.</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

Semi-supervised deep graph representation learning for anomaly detection in reconfigurable smart grids

  • 期刊论文

  • 发表日期：

    2026年7月23日

查看全文
  • Lahcen Hassine,
  • Saad Farah,
  • Hasna Chaibi,
  • Rachid Saadane,
  • Abdellah Chehri

摘要

The integration of digital technologies into smart grids enhances their efficiency but also introduces significant vulnerabilities, particularly in reconfigurable topologies where dynamic structural changes challenge conventional monitoring systems. This paper addresses the problem of detecting operational anomalies in such environments without relying on labeled data. We propose a label-free hybrid Deep Graph Representation Learning (DGRL) pipeline that combines unsupervised representation learning (GAT + VGAE) with pseudo-label-based semi-supervised classification (GCN), requiring no human annotation at any stage. A simulation laboratory was set up using MATPOWER and MATLAB to simulate IEEE 14, 69 and 300 bus systems under normal case, and network attack or reconfiguration. Raw structural and behavioral patterns were learned from unlabeled data, used reconstruction error to detect anomalies, and applied adaptive thresholding to pseudo-labels for GCN classification. Based on the generated data, this method achieved robust anomaly detection and stability to changes in the topology with 98% of accuracy. The economic evaluation revealed a Return on Security Investment of 47.3, which underlined the cost-effectiveness. This work provided a scalable and label-free solution for enhancing the operational monitoring of critical infrastructure, with the public release of the generated datasets to support further research.