基于条件生成对抗网络与多智能体强化学习的配电网可靠性评估方法  

A Reliability Assessment Method for Distribution Networks Based on Conditional Generative Adversarial Network and Multi-agent Reinforcement Learning

作　　者：徐慧慧 田云飞 赵宇洋 彭婧 石庆鑫 成锐 XU Huihui;TIAN Yunfei;ZHAO Yuyang;PENG Jing;SHI Qingxin;CHENG Rui(Economic and Technological Research Institute of State Grid Gansu Electric Power Co.,Ltd.,Lanzhou 730030,China;State Key Laboratory of New Energy Power Systems(North China Electric Power University),Beijing 102206,China)

机构地区：[1]国网甘肃省电力公司经济技术研究院,甘肃兰州730030 [2]新能源电力系统国家重点实验室(华北电力大学),北京102206

出　　处：《中国电力》2025年第4期230-236,共7页Electric Power

基　　金：国家自然科学基金资助项目(52307094);国网甘肃省电力公司咨询项目(W24FZ2730050)。

摘　　要：在大规模分布式光伏接入场景下,为提升配电网可靠性评估的计算效率和精度,提出一种基于条件生成对抗网络与多智能体强化学习的评估方法。首先,采用序贯蒙特卡洛模拟生成系统的时序状态序列,并结合条件生成对抗网络与多分辨率气象因素,刻画源荷场景的多元特性,包括时序性、波动性、随机性及源荷相关性;其次,建立多智能体强化学习模型,提出融合模仿学习与探索学习的训练算法,使智能体通过与专家经验模型的交互学习获取最优策略。最后,基于IEEE RBTS BUS-2系统进行仿真验证。仿真结果表明:该方法在学习曲线、稳定性方面均优于传统方法,显著提高了配电网可靠性评估的精度和计算效率,具有优越的实用价值。To enhance computational efficiency and accuracy in reliability assessment of distribution networks with large-scale distributed photovoltaic integration,a novel assessment method is proposed based on conditional generative adversarial network and multi-agent reinforcement learning.Firstly,the sequential state sequences of the system are generated using Sequential Monte Carlo simulation,and a conditional generative adversarial network(CGAN)is combined with multi-resolution meteorological factors to characterize the multivariate characteristics of source-load scenarios,including temporal dependency,volatility,randomness,and source-load correlation.Secondly,a multi-agent reinforcement learning(MARL)model is established,and a training algorithm integrating imitation learning and exploratory learning is proposed,enabling the agents to acquire optimal policies through interactive learning with an expert experience model.Finally,the simulationg is verified based on the IEEE RBTS BUS-2 system.Simulation results demonstrate that the proposed method outperforms traditional methods in terms of learning curve and stability,significantly improving both the accuracy and computational efficiency in distribution network reliability assessment,possessing superior practical values.

关 键 词：配电网 可靠性评估 生成对抗网络 

分 类 号：TM732[电气工程—电力系统及自动化]