Objective The increasing structural complexity and uncertain operating conditions of modern power grids have made the identification of critical nodes essential for preventing cascading failures and enhancing system resilience. Traditional topology-based and model-driven approaches rely heavily on static assumptions, accurate physical models, and extensive prior knowledge, limiting their applicability in dynamic and nonlinear power-flow environments. Existing reinforcement-learning-based identification methods also exhibit several limitations, including inadequate representation of disturbance propagation, Q-value overestimation, and limited capability to capture nonlinear interactions among nodes. To address these challenges, this study aims to i) enhance the state representation of power-grid nodes by incorporating disturbance diffusion characteristics, ii) improve the stability and accuracy of reinforcement-learning-based node evaluation, and iii) provide a robust framework for identifying nodes whose perturbation causes significant structural and operational degradation of the power grid. Methods An improved diffusion double-DQN-based reinforcement-learning framework is developed. The critical-node identification problem is formulated as a Markov decision process, in which each state represents a multi-attribute description of a node, including electrical, topological, and operational features. Actions correspond to selecting candidate nodes for evaluation, while the reward function integrates indicators of voltage deviation, load sensitivity, power-flow redistribution, and disturbance propagation. A diffusion mechanism is introduced to model the propagation of perturbations across the network topology. Starting from a given node, disturbance information is iteratively diffused through the adjacency structure to generate a multi-hop neighborhood representation. The resulting diffusion state captures both direct and indirect interactions, enhancing the agent's ability to perceive local vulnerability, load-redistribution potential, and cascading-failure risk. A multi-head attention (MHA) module is incorporated into the feature-extraction stage to capture nonlinear dependencies among node features. This module enables the model to assign adaptive importance to different node attributes and neighboring interactions, thereby improving the representation of structurally subtle yet operationally influential nodes. To improve training stability, an advantage-constrained update strategy is integrated into the double-DQN framework. This strategy reduces Q-value estimation bias, mitigates instability during policy updates, and improves learning efficiency in high-dimensional state spaces. The proposed framework integrates diffusion-enhanced state representation, attention-based feature extraction, and stabilized Q-learning into a unified model for evaluating and ranking node importance. Results and Discussions Simulation studies were performed on the IEEE 30-bus and IEEE 118-bus systems using network efficiency, the size of the largest connected component, the voltage stability index (VSI), and the load-shedding amount (LSA) as evaluation metrics. In the IEEE 30-bus system, removing the top 10 nodes ranked by each method resulted in a pronounced decline in both network efficiency and largest connected component. The nodes selected by the proposed Diffusion-DDQN consistently produced greater reductions, particularly during the initial node removals, indicating that the method effectively identifies nodes whose perturbation has a greater structural and operational impact. Because the IEEE 30-bus system is relatively small and the top-10 node sets identified by different methods partially overlap, the performance curves converge after all 10 nodes are removed. For the IEEE 118-bus system, nodes were removed according to percentage thresholds because of the larger network size. When the top 10% of nodes identified by each method were removed, the proposed Diffusion-DDQN caused the most significant degradation, reducing network efficiency to 32.8% and the largest connected component to 0.44 of their original values. Compared with the conventional DDQN, the reduction in network efficiency was 3.5% greater, while the decrease in the largest connected component exceeded that of the closeness-based method by 0.06. When the top 20% of nodes were removed, network efficiency and the largest connected component further decreased to 14.6% and 0.15, respectively. These reductions exceeded those achieved by the improved PageRank baseline by 4.7% and 0.24, respectively, demonstrating stronger agreement with disturbance-impact simulations in larger systems. Further evaluations were performed by applying controlled perturbations (5% voltage disturbance and 10% load disturbance) to the top-ranked nodes identified by each method. For the VSI, the nodes selected by the Diffusion-DDQN produced the largest voltage-stability deviations, with changes approximately 10% greater than those produced by the conventional DDQN. In terms of LSA, perturbing the top 5% of nodes identified by the Diffusion-DDQN and DDQN resulted in load shedding approximately 60% greater than that of the improved PageRank method. When the perturbation was extended to the top 20% of nodes, the Diffusion-DDQN yielded the highest LSA among all evaluated methods. Across all evaluation metrics, the diffusion-enhanced state representation enabled the agent to identify nodes that induce greater cascading stress, producing rankings that are more consistent with the outcomes of physical disturbance simulations. Conclusions This paper presents an improved Diffusion-DDQN framework for critical node identification in complex power grids. By incorporating a diffusion mechanism into the state representation, the proposed framework captures spatial disturbance-propagation characteristics that are overlooked by conventional reinforcement-learning and graph-theoretic approaches. The integration of a multi-head attention module improves feature-interaction modeling, while the advantage-constrained update strategy stabilizes Q-value estimation and enhances learning efficiency. Simulation results on the IEEE 30-bus and IEEE 118-bus systems demonstrate that the proposed framework provides more accurate, robust, and interpretable assessments of node importance. The proposed framework offers a practical and extensible tool for vulnerability analysis, protection planning, and resilience assessment in modern power systems.
MorenoJ, ShawD.Community resilience to power outages after disaster:A case study of the 2010 Chile earthquake and tsunami[J].International Journal of Disaster Risk Reduction,2019,34:448-458. doi:10.1016/j.ijdrr.2018.12.016
[2]
YanDaobo, WenJinyu, DuZhi,et al.Analysis of Texas blackout in 2021 and its enlightenment to power system planning management[J].Power System Protection and Control,2021,49(9):121-128.
ZhangPengfei, MaChanghui, LiWei,et al.Analysis of two splitting accidents of European power grid in 2021 and consideration on power grid security in China[J].Automation of Electric Power Systems,2021,45(24):22-29.
LiTao, RongLili, YanKesheng.Vulnerability analysis and critical area identification of public transport system:A case of high-speed rail and air transport coupling system in China[J].Transportation Research Part A:Policy and Practice,2019,127:55-70. doi:10.1016/j.tra.2019.07.008
[7]
Arias VelásquezR M, Mejía LaraJ V.Root cause analysis for shunt reactor failure in 500 kV power system[J].Engineering Failure Analysis,2019,104:1157-1173. doi:10.1016/j.engfailanal.2019.06.076
[8]
CaiZexiang, WangXinghua, RenXiaona.A review of complex network theory and its application in power systems[J].Power System Technology,2012,36(11):114-121.
ChenShiming, WangCong.Crucial nodes identification of cyber physical power system considering power flow characteristics[J].Computer Applications and Software,2021,38(12):50-55.
LiChangchao, KangZhongjian, YuHongguo,et al.Identification of key nodes in power communication network considering the importance of power businesses[J].Transactions of China Electrotechnical Society,2019,34(11):2384-2394.
ZouYanli, TanShuyi, LiuXinyan,et al.Power system critical node identification based on subnetwork partition[J].Chinese Journal of Computational Physics,2022,39(3):361-370.
WuHao, ZhuZiwei.Key lines identification in power grid based on comprehensive index calculated by the entropy weight-analytical hierarchy process[J].Electric Power,2020,53(5):39-47.
AdebayoI G, SunYanxia.A novel approach for the identification of influential node in an electric power grid[C]//Proceedings of the 2019 14th IEEE Conference on Industrial Electronics and Applications(ICIEA).Xi'an:IEEE,2019:718-723. doi:10.1109/ICIEA.2019.8834042
[19]
WeiGang.Evaluation method of key nodes in power grid based on medium weight entropy degree[J].Application Research of Computers,2020,37(Supp1):27-30.
[20]
魏刚.基于介权熵度的电网关键节点评估方法[J].计算机应用研究,2020,37():27-30.
[21]
WangHaoxiang, ChenJunxi, WeiZhenlin,et al.A novel algorithm for identifying key nodes in complex networks based on relative entropy and neighbor influence clustering[J].Journal of Beijing Jiaotong University,2024,48(2):154-164.
AhmadW, WangBang, ChenSi.Learning to rank influential nodes in complex networks via convolutional neural networks[J].Applied Intelligence,2024,54(4):3260-3278. doi:10.1007/s10489-024-05336-x
[24]
ZhangMin, LiuZesan, FuChenghua,et al.Identifying critical nodes in the power grid via deep learning methods[C]//Proceedings of the Eighth International Conference on Energy System,Electricity,and Power(ESEP 2023).Wuhan:SPIE,2024:15. doi:10.1117/12.3024227
[25]
LiXujie, JiPu, SunYing,et al.A node influence ranking algorithm in complex networks based on deep reinforcement learning[J].Telecommunication Engineering,2024,64(10):1644-1652.
WangLiming.Identification of vital nodes in complex networks based on deep reinforcement learning[D].Bengbu:Anhui University of Finance & Economics,2020.
[28]
汪黎明.基于深度强化学习的复杂网络关键节点识别[D].蚌埠:安徽财经大学,2020.
[29]
FanChangjun, ZengLi, SunYizhou,et al.Finding key players in complex networks through deep reinforcement learning[J].Nature Machine Intelligence,2020,2(6):317-324. doi:10.1038/s42256-020-0177-2
[30]
ChenNa.Algorithm for detection of crucial nodes in complex networks based on graph neural network[J].Radio Engineering,2023,53(1):64-72.
SunZejun, ChangXinfeng, WangFeifei,et al.Critical nodes identification for power grid based on deep reinforcement learning[J].Electronic Design Engineering,2024,32(15):166-171.
WangWangbing, WangXianpei, YouZezhang,et al.Research on key node identification method in electric power communication network[J].Power System Protection and Control,2018,46(1):44-49.
LiYunkun, FeipengLü, JiangKe,et al.Method for power system operation mode combination based on shortest electrical distance[J].Power System Protection and Control,2010,38(15):24-27.
YangXiongping, DuanXianzhong, ShiDongyuan.Method for operation mode combination based on electric coupling of mesh network[J].Automation of Electric Power Systems,2005,29(23):64-68.
XuWeiting, LiuJunyong, YaoLiangzhong,et al.Clustering system islanding analysis considering voltage controllability[J].Automation of Electric Power Systems,2013,37(21):72-79.
BialekJ.Tracing the flow of electricity[J].IEE Proceedings—Generation,Transmission and Distribution,1996,143(4):313. doi:10.1049/ip-gtd:19960461
[43]
XuJingyou, ChenChong, LuoChunjian,et al.Identification of power grid key parts based on improved complex network model[J].Automation of Electric Power Systems,2016,40(10):53-61. doi:10.7500/AEPS20150824006
LiuWenying, DanYangqing, ZhuYanwei,et al.Research on physical indicators to identify power system self-organ-ized critical state[J].Transactions of China Electrotechnical Society,2014,29(8):274-280. doi:10.3969/j.issn.1000-6753.2014.08.035
SunQiming, ShiLibao, NiYixin,et al.An enhanced cascading failure model integrating data mining technique[J].Protection and Control of Modern Power Systems,2017,2(1):5. doi:10.1186/s41601-017-0035-3
[48]
MaLijia, ShaoZengyang, LiXiaocong,et al.Influence maximization in complex networks by using evolutionary deep reinforcement learning[J].IEEE Transactions on Emerging Topics in Computational Intelligence,2023,7(4):995-1009. doi:10.1109/tetci.2021.3136643
[49]
FengBin, HuYijie, HuangGang,et al.Review on optimization methods for new power system dispatch based on deep reinforcement learning[J].Automation of Electric Power Systems,2023,47(17):187-199.
YangYanhong, MaTengfei, LiHaitao,et al.Federated double DQN based multi-energy microgrid energy management strategy considering carbon emissions[J].Global Energy Interconnection,2023,6(6):689-699. doi:10.1016/j.gloei.2023.11.003
[52]
LiuJianwei, GaoFeng, LuoXionglin.Survey of deep reinforcement learning based on value function and policy gradient[J].Chinese Journal of Computers,2019,42(6):1406-1438. doi:10.11897/SP.J.1016.2019.01406
ZhuDarui, WangRui, ChengWenji,et al.Critical transmission node identification method based on improved PageRank algorithm[J].Power System Protection and Control,2022,50(5):86-93.
RuanYirun.Research on node importance ranking and influence maximization in complex networks[D].Changsha:National University of Defense Technology,2018.
[60]
阮逸润.复杂网络节点重要性排序和影响最大化问题研究[D].长沙:国防科技大学,2018.
[61]
LeiCheng, LiuJunyong, WeiZhenbo,et al.Integrative evaluation model of node vulnerability considering network transmission ability and anti-interference ability[J].Electric Power Automation Equipment,2014,34(7):144-149. doi:10.3969/j.issn.1006-6047.2014.07.025