军事知识服务平台
基于扩散模型的配电网无功优化强化学习方法
A Reinforcement Learning Approach for Reactive Power Optimization in Distribution Networks Based on Diffusion Models
摘要 针对数据驱动的配电网无功优化方法面临训练样本获取成本高、传统物理模型方法适应性不足的双重挑战,提出一种基于条件扩散模型增强数据的配电网无功优化强化学习方法.首先,构建了考虑量测误差的马尔可夫博弈模型,刻画了多逆变器协同调控的配电网无功优化特性;其次,设计了物理信息嵌入的条件扩散概率模型,有效扩充强化学习回放经验池;然后,提出了基于时序差分误差的自适应混合经验回放机制,动态调节增强数据样本比重;最后,通过构建集中式训练-分布式执行框架,离线训练智能体的无功控制策略并将其在线部署以实现分布式自主调控.仿真结果表明,所提方法在训练样本有限的场景下显著提升了强化学习方法的性能,具有优异的实时响应能力和无功优化效果.
Abstract:
Data-driven reactive power optimization methods for distribution networks face the dual challenges of high costs in acquiring training samples and limited adaptability of traditional physics-based models.To address these issues,this paper proposes a reinforcement learning method for reactive power optimization enhanced by conditional diffusion models.First,a Markov game model incorporating measurement errors is formulated to characterize the cooperative regulation of multiple inverters in reactive power optimization.Second,a physics-informed conditional diffusion probabilistic model is designed to effectively augment the experience replay buffer for reinforcement learning.Third,an adaptive hybrid experience replay mechanism based on temporal-difference(TD)error is introduced to dynamically adjust the proportion of augmented data samples.Finally,a centralized training with decentralized execution(CTDE)framework is constructed to train the agents'reactive power control policies offline and deploy them online for distributed autonomous control.Simulation results demonstrate that the proposed method significantly improves the performance of reinforcement learning under scenarios with limited training samples,exhibiting excellent real-time response capabilities and reactive power optimization performance.
刊 名 智慧电力 ISTICEIPKU
年,卷(期): 2026, 54(9)
分类号: TM721
机标分类号: G633.7 TP391 TM721
在线出版日期: 2026年11月9日