Scholarlink

<p>To solve the problems of insufficient real-time performance, weak adaptability to coupling faults and poor reliability of traditional early warning methods for substation equipment, this study proposes an intelligent early warning model (GNN-DT) which integrates Graph Neural Network (GNN) and Digital Twin (DT). In this model, the substation system is abstracted as a weighted undirected graph, the physical model of key equipment is constructed through DT, and the parameters are adaptively modified by multi-source data fusion. a Graph Attention Mechanism (GAT) based on electrical characteristics is designed, the coupling relationship between devices and extract weak abnormal features is strengthened. With the help of Multi-Task Learning (MTL) framework with physical consistency constraints, the dynamic hierarchical early warning from device level to system level is completed. The experiment is verified by simulation data and real operation data. Experimental results show that the model achieves over 93% across four core metrics: accuracy, precision, recall, and F1-score. The false positive rate is as low as 3.5%, and the average early warning time is only 4.2&#xa0;min, which meets the engineering requirement of warning response time (&lt; 100 ms) for real-time monitoring systems in substations. Its overall performance is significantly superior to traditional machine learning, single deep learning, and single GNN/DT models. The above evaluation indicators comprehensively verify the superiority of the model from three dimensions: identification accuracy (accuracy, precision, recall, F1-score), reliability (false positive rate), and response timeliness (average early warning time). The model shows strong robustness in 10 typical fault scenarios, the recall rate of hidden faults is over 89%, and it still has excellent generalization ability in data scarce scenarios. The single sample inference time of 4.0ms can meet the requirements of engineering real-time deployment. This study realizes the organic integration of data-driven and mechanism-driven, provides reliable technical solutions for intelligent operation and maintenance of substations. The study is of great significance to improving the safe and efficient operation level of power grids.</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

Intelligent early warning of substation equipment based on GNN network and digital twin

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月13日

查看全文
  • Luyao Pei,
  • Weican Yuan,
  • Cheng Xu,
  • Chentao Wang

摘要

To solve the problems of insufficient real-time performance, weak adaptability to coupling faults and poor reliability of traditional early warning methods for substation equipment, this study proposes an intelligent early warning model (GNN-DT) which integrates Graph Neural Network (GNN) and Digital Twin (DT). In this model, the substation system is abstracted as a weighted undirected graph, the physical model of key equipment is constructed through DT, and the parameters are adaptively modified by multi-source data fusion. a Graph Attention Mechanism (GAT) based on electrical characteristics is designed, the coupling relationship between devices and extract weak abnormal features is strengthened. With the help of Multi-Task Learning (MTL) framework with physical consistency constraints, the dynamic hierarchical early warning from device level to system level is completed. The experiment is verified by simulation data and real operation data. Experimental results show that the model achieves over 93% across four core metrics: accuracy, precision, recall, and F1-score. The false positive rate is as low as 3.5%, and the average early warning time is only 4.2 min, which meets the engineering requirement of warning response time (< 100 ms) for real-time monitoring systems in substations. Its overall performance is significantly superior to traditional machine learning, single deep learning, and single GNN/DT models. The above evaluation indicators comprehensively verify the superiority of the model from three dimensions: identification accuracy (accuracy, precision, recall, F1-score), reliability (false positive rate), and response timeliness (average early warning time). The model shows strong robustness in 10 typical fault scenarios, the recall rate of hidden faults is over 89%, and it still has excellent generalization ability in data scarce scenarios. The single sample inference time of 4.0ms can meet the requirements of engineering real-time deployment. This study realizes the organic integration of data-driven and mechanism-driven, provides reliable technical solutions for intelligent operation and maintenance of substations. The study is of great significance to improving the safe and efficient operation level of power grids.