基于深度强化学习的多能虚拟电厂优化调度

Deep Reinforcement Learning-Based Optimal Scheduling for Multi-Energy Virtual Power Plant

多能虚拟电厂(Multi-Energy VPP,MEVPP)能够聚合电能、热能等多种形式的分布式能源及需求侧灵活性资源。为实现MEVPP的优化调度,文中建立了包含发电单元、制热单元、储能装置以及空调负荷集群、需求响应负荷的MEVPP模型,并面向该模型提出一种基于深度强化学习(Deep Reinforcement Learning, DRL)的优化调度方法,并设计了相应的状态、动作空间与奖励函数。该方法以近端策略优化(Proximal Policy Optimization, PPO)算法为基础,能够根据预测负荷、风/光出力、室外气温等环境信息,对分布式能源和需求侧灵活性资源进行调节,并以最小化运行成本为目标得到MEVPP优化调度策略集。算例结果证明了DRL在MEVPP优化调度中的可行性与策略集的可拓展性。

Multi-energy virtual power plant(MEVPP)aggregates various forms of distributed energy and demand-side flexibility resources including electric and thermal energy.In order to realize the optimal scheduling of MEVPP,in this paper,a MEVPP model including power generation units,heating units,energy storage unit,air conditioning loads cluster,and demand response loads cluster is established,a deep reinforcement learning(DRL)-based optimal scheduling method is proposed for this model,corresponding reward function,state and action spaces are designed.The method is based on the proximal policy optimization(PPO)algorithm,which can regulate distributed energy and demand-side flexibility resources based on environmental information such as forecasted load,wind/light output,outdoor temperature,etc.,and obtain the strategy set of MEVPP optimal scheduling with the objective of minimizing the operating cost.The case study result proves the feasibility of DRL in MEVPP optimal scheduling and the extensibility of the strategy set.