| 引用本文: | 朱振山,陈炜龙,张新炳.基于图深度强化学习的配电网电压控制方法[J].电力自动化设备,2026,46(5):137-145 |
| Zhu Zhenshan,Chen Weilong,Zhang Xinbing.Voltage control method of distribution network based on graph deep reinforcement learning[J].Electric Power Automation Equipment,2026,46(5):137-145 |
|
| 摘要: |
| 配电网传统电压控制方法计算时间长且高度依赖源荷预测数据和配电网精确模型参数，难以实现实时动态调节，且在应对拓扑变化时存在显著局限。为此，将配电网电压控制问题建模为马尔可夫决策过程，通过高斯过程回归潮流替代模型降低对配电网精确参数的依赖，并将图卷积网络引入柔性演员-评论家算法，提出基于图深度强化学习的电压控制方法。该方法借助图卷积网络增强智能体对配电网拓扑特征的学习能力，显著提高电压控制性能。为解决拓扑结构变动时需重新学习控制策略的难题，提出迁移图强化学习方法，加速算法收敛进程。仿真验证表明，该方法在数据含噪、部分缺失的复杂工况下仍保持稳定调控效果，且在电压控制精度与拓扑适应性上表现良好。 |
| 关键词: 电压控制 深度强化学习 迁移学习 图卷积网络 马尔可夫决策过程 配电网 |
| DOI：10.16081/j.epae.202601012 |
| 分类号:TM73 |
| 基金项目:福建省科技创新战略联合研究项目资助(2023R0153) |
|
| Voltage control method of distribution network based on graph deep reinforcement learning |
|
Zhu Zhenshan1,2, Chen Weilong1, Zhang Xinbing3
|
|
1.School of Electrical Engineering and Automation, Fuzhou University, Fuzhou 350108, China;2.Fujian Province University Engineering Research Center of Smart Distribution Grid Equipment, Fuzhou 350108, China;3.Minhou County Power Supply Company of State Grid Fujian Electric Power Co.,Ltd.,Minhou 350100, China
|
| Abstract: |
| The traditional voltage control method of distribution network has long calculation time and is highly dependent on the source-load prediction data and accurate model parameters of distribution network. It is difficult to achieve real-time dynamic adjustment, and has significant limitation in dealing with topology change. Therefore, the voltage control problem of distribution network is modeled as a Markov decision process. The Gaussian process regression power flow substitution model is used to reduce the dependence on accurate parameters of distribution network. The graph convolution network is introduced into the flexible actor-critic algorithm, and a voltage control method based on graph deep reinforcement learning is proposed, which uses the graph convolution network to enhance the agent’s learning ability of distribution network topological characteristics, and significantly improves the voltage control performance. In order to solve the problem of relearning the control strategy when the topological structure changes, a transfer graph reinforcement learning method is proposed to accelerate the convergence process of the algorithm. The simulation verification shows that the method still maintains stable control effect under the complex working conditions of data containing noisy and partially missed, and performs well in voltage control accuracy and topology adaptability. |
| Key words: voltage control deep reinforcement learning transfer learning graph convolutional network Markov decision process distribution network |
