发电技术 ›› 2025, Vol. 46 ›› Issue (5): 977-985.DOI: 10.12096/j.2096-4528.pgt.24005

• 新型电力系统 • 上一篇    下一篇

基于Q-learning的集中-分布式形态配电网分层分区及网架规划方法

王炜1, 熊蕴1, 赵鹏臻2, 谢宁2, 靳翔1, 王正阳1   

  1. 1.国网江苏省电力有限公司连云港供电分公司,江苏省 连云港市 222000
    2.上海交通大学电子信息与电气工程学院,上海市 闵行区 200241
  • 收稿日期:2024-01-08 修回日期:2024-05-08 出版日期:2025-10-31 发布日期:2025-10-23
  • 作者简介:王炜(1977),男,高级工程师,主要从事电网设备运维、电网规划专业工作,15705130039@163.com
    赵鹏臻(1998),男,博士研究生,主要从事配电网形态和规划相关研究工作,809563642@sjtu.edu.cn
    谢宁(1973),女,博士,副教授,主要从事电力系统安全与稳定、电力系统经济运行、智能电网方向相关研究公众,本文通信作者,xiening@sjtu.edu.cn
  • 基金资助:
    国家重点研发计划项目(2020YFB2104500)

Q-Learning-Based Method for Distribution Layer Partitioning and Grid Planning in Hybrid Centralized-Distributed Networks

Wei WANG1, Yun XIONG1, Pengzhen ZHAO2, Ning XIE2, Xiang JIN1, Zhengyang WANG1   

  1. 1.Lianyungang Power Supply Company, State Grid Jiangsu Electric Power Co. , Ltd. , Lianyungang 222000, Jiangsu Province, China
    2.School of Electronic Information and Electrical Engineering, Shanghai Jiao Tong University, Minhang District, Shanghai 200241, China
  • Received:2024-01-08 Revised:2024-05-08 Published:2025-10-31 Online:2025-10-23
  • Supported by:
    National Key R&D Program of China(2020YFB2104500)

摘要:

目的 为完善集中-分布式配电网在分层分区和网架规划方面的研究,充分考虑集中-分布式配电网要求等多方面因素,提出了基于Q-learning的分布层分区和网架规划方法。 方法 针对训练过程中出现的有向和无向转换、网架连通性、考虑多因素的分类等问题,根据分层分区和网架规划2个阶段的不同需求,分别对Reward和Agent进行了改进,使模型考虑要素更为全面,并利用实际电网数据进行了仿真验证。 结果 相比于传统电网规划方法,使用所提集中-分布式规划方案在保证一定可靠性的同时有着更好的经济性。 结论 所提方法在高可再生能源发电渗透率配电网中具有较高应用价值,为集中-分布式形态配电网的规划设计提供了一种新的思路和技术手段。

关键词: 配电网, 集中-分布式形态, 分布层分区, 网架规划, 强化学习, 智能体, 机器学习

Abstract:

Objectives To improve the research on distribution layer partitioning and grid planning in hybrid centralized-distributed networks, this study fully considers various factors such as the requirements of hybrid centralized-distributed networks and proposes a method for distribution layer partitioning and network planning based on Q-learning. Methods The method addresses issues such as directed and undirected transitions, grid connectivity, and multi-factor classification during the training process. The Reward and Agent are improved based on the different requirements of the two stages: distribution layer partitioning and grid planning. This makes the model more comprehensive by incorporating more elements and is verified through simulations using actual power grid data. Results Compared to traditional grid planning methods, the hybrid centralized-distributed planning scheme using this method provides better economic efficiency while ensuring a certain level of reliability. Conclusions The proposed method demonstrates significant application value in distribution networks with high renewable energy penetration, providing a novel idea and technical approach for the planning and design of hybrid centralized-distributed networks.

Key words: distribution network, hybrid centralized-distributed form, distribution layer partitioning, grid planning, reinforcement learning, agent, machine learning

中图分类号: