To adequately utilize flexible resources on demand side, Virtual Power Plant (VPP) is an effective solution through the aggregation and application of distributed energy resources (DER). While centralized control approaches are easy to achieve global optimum for the scheduling of every DER, they have limitations when dealing with massive number of complex and heterogeneous DERs with time-varying states. Existing decentralized control approaches are mainly based on the assumption that all DERs are completely rational, which is quite far from the reality. In this paper, using a bottom-up approach, we propose a stimulus–response control strategy to realize exploitation of flexibility by VPP. In such a strategy, DERs are dynamically aggregated through autonomous decentralized system, and interact with each other via subscription and publication of topics, regardless of the source and recipient of the messages, thus removing the direct coupling relationship between VPP Operator and DERs. Furthermore, each DER makes an independent decision through edge computing at an agent that has a general End-to-End structure and is driven by the stimulus message received from VPP Operator. We develop a simple yet efficient double deep q-network (DDQN) algorithm to optimize the state sequence of DER agents. A simulation study is conducted with over 1000 DERs including photovoltaics, electric vehicles and air conditioners. Results indicate that the proposed approach can dynamically aggregate DERs and exploit their flexibility with each DER agent dynamically adapting to the change of stimulus signals, thus achieving dynamic, automatic and adaptive exploitation of flexibility by VPP.

为了充分利用需求侧的灵活资源，虚拟电厂（VPP）是通过聚合和应用分布式能源（DER）的有效解决方案。尽管集中控制方法很容易为每个DER的调度实现全局最优，但是在处理大量具有时变状态的复杂和异构DER时，它们具有局限性。现有的分散控制方法主要是基于所有DER都是完全合理的假设，这与实际情况相去甚远。在本文中，我们采用一种自下而上的方法，提出了一种刺激-响应控制策略，以实现VPP对灵活性的利用。在这种策略中，DER是通过自治的分散系统动态聚合的，并通过订阅和发布主题进行交互，而不考虑消息的源和接收者，从而消除了VPP运营商与DER之间的直接耦合关系。此外，每个DER在具有一般端到端结构并由从VPP运营商接收到的激励消息驱动的代理处，通过边缘计算来做出独立决策。我们开发了一种简单而有效的双深度q网络（DDQN）算法，以优化DER代理的状态序列。用超过1000个DER进行了仿真研究，包括光伏，电动汽车和空调。结果表明，所提出的方法可以动态聚合DER，并利用每个DER代理动态适应刺激信号变化的灵活性，从而实现动态，