Abstract：User power theft is the main cause of abnormally high losses on voltage lines in distribution networks. With the improvement of power grid inspection, the effect of power theft control on high-loss lines is remarkable. However, non-high-loss lines still exist in electricity theft by users, the existence of such users to steal electricity is relatively small, easy to be affected by the normal fluctuations in electricity consumption and external noise, caused by the anomalous fluctuations of the daily average line loss in the blurring of the characteristics of the line loss is relatively hidden and difficult to find. Aiming at the problems of hidden behavior of power theft in non-high-loss scenarios of medium-voltage (MV) lines, easy to confuse the differential features caused by power theft, and limited recognition accuracy of traditional detection techniques, a method for detecting power theft in non-high-loss lines is proposed by integrating the causal inference of line loss and the cyclic load features. The core of the approach involves constructing a hybrid graph-based model that synergistically combines a temporal feature extraction network with a causality-driven graph attention mechanism. First, by constructing WaveNet, load sequences with different cycle lengths are divided to capture the load features of different cycles and fused into the node feature matrix of the model. Second, the graph attention neural network based on causal inference is established to determine the directional connection relationship between each node and line loss, and the causal correlation between user behavior and line loss is mined by using causal sampling, which effectively improves the anti-interference ability of the model, and further calculates multi-attention coefficients to obtain the domain information in the causal paths, and constructs the adjacency matrix. Finally, the graph network structure is used to undertake two kinds of electricity behavior characteristics and refine the contribution of each user to the abnormal fluctuation of line loss. By simulating different types of power theft behavior with examples and analysis, the experimental results show that the model proposed has good detection accuracy and robustness. Simulation experiments on line non-high-loss power theft scenarios were conducted using actual collected user data. The experimental results show that the proposed model analyzes the non-high-loss scenarios of single-user power theft, and its detection indexes are all above 95%, while the recognition accuracy Acc for average power theft, peak-shaving power theft, and identification of zero power theft is close to 99%. In various non-high-loss scenarios of mixed multi-user power theft, the detection indexes of the proposed model are all over 94%, of which the recognition accuracy Acc is close to 96%, which proves the effectiveness of the proposed model. Through the perturbation experiment analysis, it can be seen that under 20% perturbation, combining the C-GAT model to analyze the non-high-loss scenarios, the evaluation indexes are still higher than 90%, which has strong anti-interference ability. In the ablation experiment, comparing the detection performance of multiple models verifies that the proposed dual-module cooperative power theft detection model has good recognition effect in non-high-loss scenarios. The following conclusions can be obtained from the analysis: (1) The analysis of the user's small amount of electricity theft causes the 15 min scale line loss to show a local high loss state, but the phenomenon of the daily scale full-volume line loss blurring masks the characteristics of the theft of electricity. (2) The model proposed can effectively distill the causal connection between the user's longitudinal cycle law and the abnormal line loss caused by electricity theft, and combine the two characteristics to evaluate the extent of the user's behavioral abnormalities. (3) The model proposed has a strong robustness and anti-jamming capability in different electricity theft scenarios, and it is more resistant to interference.
[1] 王佳仁, 谢宁, 王承民, 等. 提升现代配电系统效率和效益的研究与实践[J]. 电网技术, 2018, 42(11): 3543-3549. Wang Jiaren, Xie Ning, Wang Chengmin, et al.Research and practice of enhancing efficiency and benefit of modern distribution system[J]. Power System Technology, 2018, 42(11): 3543-3549. [2] 王守相, 张丙杰, 赵倩宇, 等. 基于集成树和MoE的馈线统计线损率双层估计模型[J]. 电工技术学报, 2024, 39(3): 774-784. Wang Shouxiang, Zhang Bingjie, Zhao Qianyu, et al.Double-layers stacking estimation model for feeder statistical line loss rate based on tree-based ensemble learning and MoE[J]. Transactions of China Electro-technical Society, 2024, 39(3): 774-784. [3] 黄南天, 程铎, 蔡国伟. 基于改进时空图神经网络的高渗透率有源配电网故障定位[J]. 电力系统自动化, 2025, 49(10): 112-122. Huang Nantian, Cheng Duo, Cai Guowei.Fault location for active distribution network with high penetration rate based on improved spatio-temporal graph neural network[J]. Automation of Electric Power Systems, 2025, 49(10): 112-122. [4] 吴涵, 袁越, 侯语涵, 等. 配电网理论线损概率分布函数的计算与分析[J]. 中国电机工程学报, 2024, 44(16): 6444-6455. Wu Han, Yuan Yue, Hou Yuhan, et al.Computation and analysis of theoretic line loss probability distribution function of distribution network[J]. Proceedings of the CSEE, 2024, 44(16): 6444-6455. [5] 陈启鑫, 郑可迪, 康重庆, 等. 异常用电的检测方法: 评述与展望[J]. 电力系统自动化, 2018, 42(17): 189-199. Chen Qixin, Zheng Kedi, Kang Chongqing, et al.Detection methods of abnormal electricity consumption behaviors: review and prospect[J]. Automation of Electric Power Systems, 2018, 42(17): 189-199. [6] 唐冬来, 刘友波, 熊智临, 等. 基于时空关联矩阵的配电台区反窃电预警方法[J]. 电力系统自动化, 2020, 44(19): 168-176. Tang Donglai, Liu Youbo, Xiong Zhilin, et al.Early warning method of electricity anti-theft in distribution station area based on spatiotemporal correlation matrix[J]. Automation of Electric Power Systems, 2020, 44(19): 168-176. [7] 林振智, 崔雪原, 金伟超, 等. 用户侧窃电检测关键技术[J]. 电力系统自动化, 2022, 46(5): 188-199. Lin Zhenzhi, Cui Xueyuan, Jin Weichao, et al.Key technologies of electricity theft detection at consumer side[J]. Automation of Electric Power Systems, 2022, 46(5): 188-199. [8] 金晟, 苏盛, 薛阳, 等. 数据驱动窃电检测方法综述与低误报率研究展望[J]. 电力系统自动化, 2022, 46(1): 3-14. Jin Sheng, Su Sheng, Xue Yang, et al.Review on data-driven based electricity theft detection method and research prospect for low false positive rate[J]. Automation of Electric Power Systems, 2022, 46(1): 3-14. [9] Glauner P, Meira J A, Valtchev P, et al.The challenge of non-technical loss detection using artificial intelligence: a survey[J]. International Journal of Computational Intelligence Systems, 2017, 10(1): 760-775. [10] Amin S, Schwartz G A, Cardenas A A, et al.Game-theoretic models of electricity theft detection in smart utility networks: providing new capabilities with advanced metering infrastructure[J]. IEEE Control Systems Magazine, 2015, 35(1): 66-81. [11] Raggi L M R, Trindade F C L, Cunha V C, et al. Non-technical loss identification by using data analytics and customer smart meters[J]. IEEE Transactions on Power Delivery, 2020, 35(6): 2700-2710. [12] 王昕, 田猛, 赵艳峰, 等. 一种基于状态估计的新型窃电方法及对策研究[J]. 电力系统保护与控制, 2016, 44(23): 141-146. Wang Xin, Tian Meng, Zhao Yanfeng, et al.A new kind of electricity theft based on state estimation and countermeasure[J]. Power System Protection and Control, 2016, 44(23): 141-146. [13] Zheng Zibin, Yang Yatao, Niu Xiangdong, et al.Wide and deep convolutional neural networks for electricity-theft detection to secure smart grids[J]. IEEE Transactions on Industrial Informatics, 2018, 14(4): 1606-1615. [14] Biswas P P, Cai Hongyun, Zhou Bin, et al.Electricity theft pinpointing through correlation analysis of master and individual meter readings[J]. IEEE Transactions on Smart Grid, 2020, 11(4): 3031-3042. [15] 李秋红. 基于改进粒子群算法的前馈神经网络识别用户窃电行为[J]. 电气技术, 2022, 23(11): 44-48. Li Qiuhong.Feedforward neural network based on improved particle swarm optimization algorithm for identification of user electricity stealing behavior[J]. Electrical Engineering, 2022, 23(11): 44-48. [16] 高昂, 郑建勇, 梅飞, 等. 基于三元组孪生网络的窃电检测算法[J]. 中国电机工程学报, 2022, 42(11): 3975-3986. Gao Ang, Zheng Jianyong, Mei Fei, et al.Electricity theft detection algorithm based on triplet network[J]. Proceedings of the CSEE, 2022, 42(11): 3975-3986. [17] 王炜韬, 赵健, 王小宇. 基于对冲对抗机制与图注意力网络的异常用电检测[J]. 电力系统自动化, 2022, 46(22): 120-128. Wang Weitao, Zhao Jian, Wang Xiaoyu.Abnormal electricity consumption detection based on hedge-antagonism mechanism and graph attention network[J]. Automation of Electric Power Systems, 2022, 46(22): 120-128. [18] 金涛, 王万豪, 黄钦瑜, 等. 一种利用轻量级多通道注意力融合网络的窃电检测边缘计算方法[J]. 中国电机工程学报, 2026, 46(6): 2420-2433. Jin Tao, Wang Wanhao, Huang Qinyu, et al.An edge calculation method of electricity stealing detection using lightweight multi-channel attention fusion network[J]. Proceedings of the CSEE, 2026, 46(6): 2420-2433. [19] 陈敏, 张逸, 邹阳, 等. 基于稳健回归和卷积神经网络的中压窃电类型检测方法[J]. 电网技术, 2024, 48(11): 4729-4738. Chen Min, Zhang Yi, Zou Yang, et al.A medium-voltage stealing type detection method based on robust regression and convolutional neural network[J]. Power System Technology, 2024, 48(11): 4729-4738. [20] 刘康, 李彬, 薛阳, 等. 基于传递熵密度聚类的用户窃电识别方法[J]. 中国电机工程学报, 2022, 42(20): 7535-7546. Liu Kang, Li Bin, Xue Yang, et al.User electric theft detection method based on transfer entropy density clustering[J]. Proceedings of the CSEE, 2022, 42(20): 7535-7546. [21] 邹念, 张颖, 苏盛, 等. 基于小时尺度周期特征自编码器的用户窃电识别方法[J]. 电网技术, 2023, 47(6): 2558-2567. Zou Nian, Zhang Ying, Su Sheng, et al.User electric theft detection method based on hour scale periodic feature LSTM-transformer[J]. Power System Technology, 2023, 47(6): 2558-2567. [22] Liao Wenlong, Zhu Ruijin, Yang Zhe, et al.Electricity theft detection using dynamic graph construction and graph attention network[J]. IEEE Transactions on Industrial Informatics, 2024, 20(4): 5074-5086. [23] 陈静, 郑垂锭, 李桂敏, 等. 考虑行业关联度的工业用户用电异常识别研究[J]. 仪器仪表学报, 2023, 44(4): 72-81. Chen Jing, Zheng Chuiding, Li Guimin, et al.Research on power consumption anomaly identification of industrial users considering industry relevance[J]. Chinese Journal of Scientific Instrument, 2023, 44(4): 72-81. [24] 李青, 张新燕, 摆志俊, 等. 基于MQ-WaveNet的风电集群发电功率多步概率预测[J]. 电力系统自动化, 2023, 47(8): 156-168. Li Qing, Zhang Xinyan, Bai Zhijun, et al.Multi-step probability prediction of power generation for wind power clusters based on multi-horizon quantile-WaveNet[J]. Automation of Electric Power Systems, 2023, 47(8): 156-168. [25] 胡博, 张鹏飞, 黄恩泽, 等. 基于图WaveNet的电动汽车充电负荷预测[J]. 电力系统自动化, 2022, 46(16): 207-213. Hu Bo, Zhang Pengfei, Huang Enze, et al.Graph WaveNet based charging load forecasting of electric vehicle[J]. Automation of Electric Power Systems, 2022, 46(16): 207-213. [26] 涂潮, 刘万军, 赵琳琳, 等. 有限训练样本下的多尺度空洞密集网络高光谱影像分类[J]. 仪器仪表学报, 2024, 45(4): 206-216. Tu Chao, Liu Wanjun, Zhao Linlin, et al.Multiscale dilated dense network for hyperspectral image classification with limited training samples[J]. Chinese Journal of Scientific Instrument, 2024, 45(4): 206-216. [27] 陈晓龙, 孙丽蓉, 李永丽, 等. 基于图注意力网络和一致性风险控制的配电网故障区段定位方法[J]. 电网技术, 2023, 47(12): 4866-4877. Chen Xiaolong, Sun Lirong, Li Yongli, et al.A fault section location method based on graph attention network and conformal risk control in distribution network[J]. Power System Technology, 2023, 47(12): 4866-4877. [28] 郭美仑, 寇鹏, 田润泽, 等. 基于贝叶斯图卷积神经网络的风电场内多风机风速概率预测[J]. 电工技术学报, 2025, 40(17): 5539-5552. Guo Meilun, Kou Peng, Tian Runze, et al.Speed probabilistic forecasting of multiple wind turbines in wind farm based on Bayesian graph convolutional neural network[J]. Transactions of China Electro-technical Society, 2025, 40(17): 5539-5552. [29] 李云松, 张智晟. 考虑综合需求响应的Transformer-图神经网络综合能源系统多元负荷短期预测[J]. 电工技术学报, 2024, 39(19): 6119-6128. Li Yunsong, Zhang Zhisheng.Transformer based multi load short-term forecasting of integrated energy system considering integrated demand response[J]. Transactions of China Electrotechnical Society, 2024, 39(19): 6119-6128. [30] 张翼, 刘富州, 朱永利, 等. 广域行波信息与图注意力网络相结合的输电网故障定位[J]. 仪器仪表学报, 2022, 43(6): 140-150. Zhang Yi, Liu Fuzhou, Zhu Yongli, et al.Fault location of transmission network combining Wide-area traveling wave information and graph attention network[J]. Chinese Journal of Scientific Instrument, 2022, 43(6): 140-150. [31] 郑毅, 王承民, 刘保良, 等. 基于多层级时空图神经网络的风电机组在线异常检测[J]. 电力系统自动化, 2024, 48(5): 107-119. Zheng Yi, Wang Chengmin, Liu Baoliang, et al.Online anomaly detection of wind turbines based on hierarchical spatio-temporal graph neural network[J]. Automation of Electric Power Systems, 2024, 48(5): 107-119. [32] 李佳玮, 王小君, 和敬涵, 等. 基于图注意力网络的配电网故障定位方法[J]. 电网技术, 2021, 45(6): 2113-2121. Li Jiawei, Wang Xiaojun, He Jinghan, et al.Distribution network fault location based on graph attention network[J]. Power System Technology, 2021, 45(6): 2113-2121. [33] 张涛, 程毅飞, 孙欣煦. 基于因果推断的图注意力网络[J]. 计算机科学, 2023, 50(增刊1): 157-165. Zhang Tao, Cheng Yifei, Sun Xinxu.Graph attention networks based on causal inference[J]. Computer Science, 2023, 50(S1): 157-165. [34] 彭昱, 符琛, 郭昕, 等. 基于FCM-BOA-TCN-GRU的分布式光伏出力异常检测方法[J]. 电工技术学报, 2025, 40(17): 5389-5401. Peng Yu, Fu Chen, Guo Xin, et al.Distributed photovoltaic power output anomaly detection method based on FCM-BOA-TCN-GRU[J]. Transactions of China Electrotechnical Society, 2025, 40(17): 5389-5401. [35] 葛磊蛟, 李京京. 面向先进计量基础设施的人工智能驱动窃电检测方法综述[J]. 中国电机工程学报, 2026, 46(10): 4086-4107. Ge Leijiao, Li Jingjing.A review of artificial intelligence-driven electricity theft detection methods under advanced metering infrastructure[J]. Proceedings of the CSEE, 2026, 46(10): 4086-4107. [36] 杨天鹏, 龙葵, 李秋惠, 等. 基于时域信号超分辨率的串联电弧检测方法[J]. 电工技术学报, 2026, 41(13): 4554-4568. Yang Tianpeng, Long Kui, Li Qiuhui, et al.Series arc fault detection method based on time-domain signal super-resolution[J]. Transactions of China Electro-technical Society, 2026, 41(13): 4554-4568.
2026, Vol. 41 
