| 引用本文: | 温琳,宋懿洋,洪启腾,林俐,王剑晓.基于机理驱动强化学习的虚拟电厂应急防灾调度方法[J].电力自动化设备,2025,45(6):200-207. |
| WEN Lin,SONG Yiyang,HONG Qiteng,LIN Li,WANG Jianxiao.Emergency disaster response dispatching method for virtual power plants based on mechanism-driven reinforcement learning[J].Electric Power Automation Equipment,2025,45(6):200-207. |
|
| 摘要: |
| 应急调度场景下需要应对强不确定性、快速制定灾后应急运行方案，但传统的强化学习求解算法在物理约束的安全保障方面存在不足。考虑台风灾害影响机理，提出基于机理驱动的强化学习算法引导虚拟电厂应急调度策略高效生成。在以虚拟电厂作为智能体、灾后主网系统作为交互环境的基础下，在多智能体深度确定性策略梯度算法中把线路潮流和日负荷最低需求量的安全约束条件以损失函数的形式内嵌到动作网络和评价网络的学习过程中，利用损失函数值的大小引导神经网络中参数的有效更新方向，若动作越限，则利用相应惩罚令其加速回到安全域内，使其生成的策略可以更好地满足系统物理约束及应急调度的时效性需求。基于改进IEEE 5节点系统和IEEE 30节点系统验证所提算法在加速收敛方面的效果与虚拟电厂的加入对电力系统韧性提升的影响。 |
| 关键词: 电力系统韧性 多主体强化学习 物理信息引导 虚拟电厂 应急调度 |
| DOI：10.16081/j.epae.202412002 |
| 分类号:TM73 |
| 基金项目:国家自然科学基金面上项目(52277092)；北京大学-思克莱德大学联合科研基金资助项目；中国科协青年人才托举工程项目(YESS20210227) |
|
| Emergency disaster response dispatching method for virtual power plants based on mechanism-driven reinforcement learning |
|
WEN Lin1, SONG Yiyang2, HONG Qiteng3, LIN Li1, WANG Jianxiao4
|
|
1.State Key Laboratory of Alternate Electrical Power System with Renewable Energy Sources, North China Electric Power University, Beijing 102206, China;2.Department of Electrical and Electronic Engineering, University of Hong Kong, Hong Kong 999077, China;3.College of Electronic and Electrical Engineering, University of Strathclyde, Glasgow G11RD, UK;4.National Engineering Laboratory for Big Data Analysis and Applications, Peking University, Beijing 100871, China
|
| Abstract: |
| In emergency dispatching scenarios, it is necessary to address high uncertainty and quickly formulate post-disaster emergency operation schemes. However, traditional reinforcement learning algorithms have shortcomings in the security guarantee of physical constraints. Considering the impact mechanism of typhoon disasters, a mechanism-driven reinforcement learning algorithm is proposed to guide the efficient generation of emergency dispatching strategy of virtual power plant. Based on the interaction between the virtual power plant as the agent and the post-disaster main grid system as the environment, within the multi-agent deep deterministic policy gradient algorithm, safety constraints such as line power flow and minimum daily load demand are embedded into the learning process of action network and evaluation network in the form of loss functions. The magnitude of these loss functions guides the effective updating of neural network parameters, and if an action exceeds the specified limits, corresponding penalties are applied to accelerate its return to the safe operating domain. This integration ensures that the generated strategies better satisfy physical constraints and the timeliness requirements of emergency dispatching. The effectiveness of the proposed algorithm in accelerating convergence and the impact of the inclusion of virtual power plants on the resilience of power system are validated using the improved IEEE 5-bus and IEEE 30-bus systems. |
| Key words: power system resilience multi-agent reinforcement learning physics-informed guidance virtual power plants emergency dispatching |
