军事知识服务平台
考虑多主体交互数据增强的充电站智能定价策略
Multi-Agent Interaction-Based Data Augmentation for Intelligent Pricing Strategies at Charging Stations
摘要 [目的]为解决大规模电动汽车(electric vehicle,EV)随机接入下多充电站协同定价面临的交互影响复杂、训练样本匮乏等问题,提出一种基于Transformer去噪扩散概率模型的数据增强与多智能体深度强化学习的协同定价方法,以扩充复杂工况下的定价学习样本空间,优化多充电站智能体协同定价策略.[方法]首先,设计基于Transformer架构的去噪扩散概率模型,引入稀疏注意力机制捕捉充电站时空分布特征,并结合Top-k价值筛选与Softmax概率采样机制,在历史电价、充电需求和时间特征等多维条件引导下,重构并扩充高质量、物理一致的交互样本空间;然后,构建计及 EV 时空分布与配电网运行约束的交互式环境,利用多智能体深度确定性策略梯度算法(multi-agent deep deterministic policy gradient,MADDPG)在增强样本集上进行协同定价策略训练.[结果]仿真结果表明,所提方法在保证配电网安全运行的前提下,能够有效引导充电负荷合理时空分布,提升充电站运营收益与用户服务水平.[结论]所提基于Transformer的样本增强方法能够有效扩充定价学习的样本空间,增强策略对复杂交互的适应性;实现了配电网安全、充电站效益及用户体验的协调优化,为电动汽车充电市场多主体智能协同定价提供了新的技术途径.
Abstract:
[Objective]To address complex interactions and scarce training samples in collaborative pricing among multiple charging stations under large-scale random EV access,this paper proposes a method integrating Transformer-based DDPM data augmentation with multi-agent deep reinforcement learning to expand the sample space for pricing policy learning under complex conditions and optimize multi-agent collaborative pricing strategies.[Methods]A Transformer-based denoising diffusion probabilistic model(DDPM)is first designed.By employing a sparse attention mechanism to capture spatiotemporal characteristics of charging stations and combining Top-k selection with Softmax sampling,the model reconstructs and augments a physically consistent interaction sample space conditioned on historical prices,demand,and temporal features.Then,an interactive environment considering EV spatiotemporal distribution and network constraints is constructed,and a multi-agent deep deterministic policy gradient(MADDPG)algorithm trains collaborative pricing strategies on the augmented set.[Results]Simulation results show the proposed method effectively guides the rational spatiotemporal distribution of charging loads,enhances station revenue and service quality while ensuring distribution network safety.The Transformer-based sample augmentation method significantly expands the sample space for pricing learning and strengthens the strategy's adaptability to complex interactions.[Conclusions]The proposed approach achieves coordinated optimization among distribution network security,station profitability,and user experience,offering a novel pathway for intelligent multi-agent collaborative pricing in the EV charging market.
刊 名 电力建设 ISTICPKU
年,卷(期): 2026, 47(10)
分类号: TM72
机标分类号: TP391 TP18 F426.61
在线出版日期: 2026年10月19日