| 引用本文: | 陈俊斌,余涛,潘振宁.面向主动配电网实时优化调度的图强化学习方法[J].控制理论与应用,2024,41(6):999~1008.[点击复制] |
| CHEN Jun-bin,YU Tao,PAN Zhen-ning.Graph reinforcement learning for real-time optimal dispatch of active distribution network[J].Control Theory and Technology,2024,41(6):999~1008.[点击复制] |
|
| 面向主动配电网实时优化调度的图强化学习方法 |
| Graph reinforcement learning for real-time optimal dispatch of active distribution network |
| 摘要点击 671 全文点击 169 投稿时间：2023-02-27 修订日期：2023-07-15 |
| 查看全文 查看/发表评论 下载PDF阅读器 |
| DOI编号 DOI: 10.7641/CTA.2023.30091 |
| 2024,41(6):999-1008 |
| 中文关键词 主动配电网 实时优化调度 图表示学习 图强化学习 图神经网络 |
| 英文关键词 activate distribution network real-time optimal dispatch graph representation graph reinforcement learning graph neural network |
| 基金项目 国家自然科学基金委员会–国家电网公司智能电网联合基金项目(U2066212), 国家自然科学基金项目(52207105), 中国博士后科学基金项目(2022 M721184)资助. |
|
| 中文摘要 |
| 主动配电网的新能源、储能等能源形式可以有效提高运行的灵活性和可靠性, 同时新能源和负荷也给配电网带来了双重不确定性, 致使主动配电网的实时优化调度决策维度大、建模精度差. 针对这一问题, 本文提出结合图神经网络和强化学习的图强化学习方法, 避免对复杂系统的精准建模. 首先, 将实时优化调度问题表述为马尔可夫决策过程, 并将其表述为动态序贯决策问题. 其次, 提出了基于物理连接关系的图表示方法, 用以表达状态量的隐含相关性. 随后, 提出图强化学习来学习将系统状态图映射到决策输出的最优策略. 最后, 将图强化学习推广到分布式图强化学习. 算例结果表明, 图强化学习在最优性和效率方面都取得了更好的效果. |
| 英文摘要 |
| The renewable energy system, energy storage system and other energy resources of active distribution network can effectively improve the flexibility and reliability of operation. Meanwhile, renewable energy and load also bring
uncertainty to the distribution network, resulting in large dimensions of real-time optimal dispatch and poor modeling accuracy of active distribution network. To solve this problem, a graph reinforcement learning method combining graph neural network and reinforcement learning is proposed to avoid accurate modeling of complex systems. Firstly, the real-time optimal dispatch is described as Markov decision process and dynamic sequential decision problem. Secondly, a graph representation method based on the physical connection is proposed to express the implied correlation of state variables. Then a graph reinforcement learning is proposed to learn the optimal strategy for mapping system state graph to decision output. Finally, the graph reinforcement learning is developed to distributed graph reinforcement learning. The simulations show that graph reinforcement learning achieves better results in optimality and efficiency. |
|
|
|
|
|