Scholarlink

Distribution network reconfiguration (DNR) is a key measure for enhancing the operational economy and reliability of power grids. However, traditional optimization methods rely on accurate models and are computationally time-consuming, making them difficult to apply in real-time scenarios. To address this, this paper proposes a DNR method based on Graph Convolutional Networks (GCN) and Deep Q-Learning. The GCN is utilized to extract grid topological information and construct node representations that integrate both global and local features. A reward function incorporating a nonlinear reward factor is designed to enhance the agent’s sensitivity to different power losses. Furthermore, an action masking mechanism based on graph theory is introduced to ensure that actions satisfy radial constraints, thereby improving exploration efficiency. Experiments on the IEEE 33-node system demonstrate that the proposed method remains effective in noisy environments, achieving an average power loss reduction of approximately 36.2%, and exhibits good robustness and real-time performance.

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

Graph Convolutional Network-Based Deep Reinforcement Learning for Dynamic Distribution System Reconfiguration

  • 会议论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年

查看全文
  • Ruilin Wang

摘要

Distribution network reconfiguration (DNR) is a key measure for enhancing the operational economy and reliability of power grids. However, traditional optimization methods rely on accurate models and are computationally time-consuming, making them difficult to apply in real-time scenarios. To address this, this paper proposes a DNR method based on Graph Convolutional Networks (GCN) and Deep Q-Learning. The GCN is utilized to extract grid topological information and construct node representations that integrate both global and local features. A reward function incorporating a nonlinear reward factor is designed to enhance the agent’s sensitivity to different power losses. Furthermore, an action masking mechanism based on graph theory is introduced to ensure that actions satisfy radial constraints, thereby improving exploration efficiency. Experiments on the IEEE 33-node system demonstrate that the proposed method remains effective in noisy environments, achieving an average power loss reduction of approximately 36.2%, and exhibits good robustness and real-time performance.