基于决策空间裁剪强化学习的连锁故障调切结合紧急控制  

Emergency Control Combined with Adjustment and Tripping Against Cascading Failures Based on Decision Space Pruning Reinforcement Learning

作　　者：陈戈 张俊勃 彭颖 王明扬 CHEN Ge;ZHANG Junbo;PENG Ying;WANG Mingyang(School of Electric Power Engineering,South China University of Technology,Guangzhou 510641,China)

机构地区：[1]华南理工大学电力学院,广东省广州市510641

出　　处：《电力系统自动化》2025年第6期144-156,共13页Automation of Electric Power Systems

基　　金：国家自然科学基金资助项目(52277101);国家自然科学基金企业创新发展联合基金项目(U22B6007);广州市应用基础研究计划项目(2024A04J9892)。

摘　　要：新能源快速功率控制技术的发展使新能源功率调节具有参与过载主导型连锁故障紧急控制的潜力,而现有基于深度强化学习的连锁故障紧急控制方法未考虑调切结合的控制策略,且存在因决策空间较大而难以收敛的问题。为此,文中提出一种基于决策空间裁剪图深度强化学习的电网连锁故障调切结合紧急控制方法。首先,构建调切结合的映射策略模型,提出连锁故障紧急控制框架;其次,提出基于图卷积深度网络的决策空间裁剪模型及学习方法,根据灵敏度大小保留有控制贡献的控制地点以裁剪决策空间;然后,提出基于图深度强化学习的映射策略模型学习方法,在给定控制地点下实现对控制量的学习;最后,在IEEE 39节点和IEEE 300节点系统中验证所提方法的有效性和泛化性。The development of fast power control technologies for renewable energy makes the power adjustment of renewable energy has the potential to participate in the emergency control against overload-dominated cascading failures.However,the existing deep reinforcement learning based emergency control methods against cascading failures do not consider the control strategies combined adjustment and tripping,and there exists a problem of difficulty in converging with large decision spaces.Therefore,an emergency control method combined with adjustment and tripping against cascading failures in the power grid based on the decision space pruning graph deep reinforcement learning is proposed.First,a mapping strategy model combined adjustment and tripping is established,and an emergency control framework against cascading failures is proposed.Subsequently,a decision space pruning model and its learning method based on the graph convolutional deep network are proposed to prune the decision space by retaining the control locations with control contributions through sensitivity.Then,a learning method for the mapping policy model based on the graph deep reinforcement learning is proposed to learn control quantities under the given control locations.Finally,the effectiveness and generalization of the proposed method are verified in the IEEE 39-bus and IEEE 300-bus systems.

关 键 词：新型电力系统 新能源 连锁故障 图深度强化学习 紧急控制 决策空间裁剪 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]