基于双注意力深度强化学习的有源配电网电压管理算法  

Voltage Management Method for Active Distribution Networks Based on Double Attention Deep Reinforcement Learning Algorithm

作　　者：蔡雪丹 赵文会 颜俊芳 CAI Xuedan;ZHAO Wenhui;YAN Junfang(Shanghai University of Electric Power,Shanghai 200090,China;Huzhou Power Supply Company,State Grid Zhejiang Electric Power Co.,Ltd.,Huzhou,Zhejiang 313000,China;Ningbo Yongyao Power Supply Service Co.,Ltd.,Ningbo,Zhejiang 315000,China)

机构地区：[1]上海电力大学,上海200090 [2]国网浙江省电力有限公司湖州供电公司,浙江湖州313000 [3]宁波永耀供电服务有限公司,浙江宁波315000

出　　处：《上海电力大学学报》2025年第1期35-42,49,共9页Journal of Shanghai University of Electric Power

基　　金：国家重点研发计划(2022YFE0207700)。

摘　　要：随着配电网中分布式电源渗透率的增加,配电网电压越限和网损增加的问题愈发严重。由于传统基于模型的无功电压管理算法十分依赖电网的精准建模,但电网实际的精确参数难以获取,且新能源机组等无功功率优化设备与现有的配电网电压管理体系不匹配,因此提出了一种基于深度强化学习的有源配电网电压管理算法。首先,从数据层、调控层及设备层建立电压管理架构。然后,针对传统设备与无功功率优化设备的协调管理问题,构建了双时间尺度电压控制框架,对快、慢时间尺度分别建立基于双注意力自适应熵的多智能体柔性行动器-判别器算法和竞争深度Q网络算法求解模型。最后,通过IEEE33节点配电网系统模型进行仿真分析,证明所提算法具有更好的电压管理效果和更快的求解速度。With the increasing penetration rate of distributed renewable energy generation in the distribution network,the problems of voltage exceeding limits and increased network losses in distribution networks have become increasingly serious.The traditional model-based reactive power and voltage optimization methods heavily rely on precise modeling of the power grid,but it is difficult to obtain accurate parameters of the actual power grid.The voltage management strategy with the participation of renewable energy is not yet perfect.This paper proposes a data-driven dual time scale voltage control strategy based on deep reinforcement learning methods.Firstly,the voltage management architecture is proposed from different levels such as data layer,regulation layer and equipment layer.Then,for the coordinated management problem of renewable energy and traditional reactive power compensation equipment,a dual time-scale voltage control framework is constructed to coordinate the reactive power optimization equipment with different response characteristics,and the model is solved by the multi-agent deep deterministic strategy gradient algorithm with the dual attention adaptive entropy and the dual deep Q network algorithm for the fast and slow time scales,respectively.Finally,the IEEE 33 node distribution system is used to verify the advantages of the data-driven solution in achieving speed and effectiveness in optimizing reactive power and voltage.

关 键 词：深度强化学习 双时间尺度 无功电压管理 双注意力自适应熵 

分 类 号：TM732[电气工程—电力系统及自动化]