华东师范大学学报（自然科学版） ›› 2023, Vol. 2023 ›› Issue (5): 193-204.doi: 10.3969/j.issn.1000-5641.2023.05.016

• 数据分析 • 上一篇    

基于残差网络和深度可分离卷积增强自注意力机制的窃电识别

段志尚1(), 冉懿1, 吕笃良1, 祁杰2, 钟佳晨2, 袁培森2,*()   

  1. 1. 国网新疆电力有限公司 营销服务中心(资金集约中心、计量中心), 乌鲁木齐, 830000
    2. 南京农业大学 人工智能学院, 南京　210031
  • 收稿日期:2023-07-05 出版日期:2023-09-25 发布日期:2023-09-20
  • 通讯作者: 袁培森 E-mail:15739576170@163.com;peiseny@njau.edu.cn
  • 作者简介:段志尚, 男, 硕士, 工程师, 主要研究方向为电能计量、数据分析. E-mail: 15739576170@163.com
  • 基金资助:
    国家自然科学基金(61877018); 上海市大数据管理系统工程研究中心开放基金(HYSY21022)

Identifying electricity theft based on residual network and depthwise separable convolution enhanced self attention

Zhishang DUAN1(), Yi RAN1, Duliang LYU1, Jie QI2, Jiachen ZHONG2, Peisen YUAN2,*()   

  1. 1. The Marketing Service Center (Capital Pooling Centers, Metering Centers), State Grid Xinjiang Electric Power Co. Ltd., Urumqi,　830000, China
    2. College of Artificial Intelligence, Nanjing Agricultural University, Nanjing　210031, China
  • Received:2023-07-05 Online:2023-09-25 Published:2023-09-20
  • Contact: Peisen YUAN E-mail:15739576170@163.com;peiseny@njau.edu.cn

RichHTML

7

PDF

131

摘要：

窃电行为严重危害着电力设备和人身安全, 并造成重大经济损失. 对窃电行为实现准确识别是供电企业降损增效的一项重要工作. 在残差网络 (residual network, ResNet) 结构的基础上, 将二维卷积神经网络与深度可分离卷积增强的自注意力 (depthwise separable convolution enhanced self attention, DSCAttention) 机制相结合并构建模型, 用于提升窃电用户的正确分类. 此外, 由于窃电数据常存在缺失值、异常值和正负样本不平衡的问题, 故采用补零法、分位数变换和分层拆分法对以上问题分别处理. 在真实窃电数据集上进行了大量实验, 实验结果表明, 所提模型的AUC指标达到了91.92%, MAP@100指标达到了98.58%, MAP@200指标达到了96.77%. 与其他窃电分类模型相比, 所提模型在窃电分类任务上亦有很好的表现, 可以在窃电智能化识别中推广使用.

关键词: 残差网络, 卷积增强, 自注意力机制, 深度可分离卷积, 窃电识别

Abstract:

Power theft seriously endangers power equipment and personal safety, and causes significant economic losses for energy suppliers. Hence, it is important for these suppliers to accurately identify instances of power theft to reduce losses and increase efficiency. In this paper, based on the residual network (ResNet) structure, a 2D convolutional neural network is combined with a depthwise separable convolution enhanced self-attentive (DSCAttention) mechanism to improve the number of correctly-classified electricity theft users. In addition, electricity theft data often contains missing values, outliers, and positive and negative sample imbalance. Each of the above problems are treated separately using the zero-completion method, quantile transformation, and hierarchical splitting method, respectively. The proposed model has been extensively tested using real power theft data sets. The results show that the area under curve (AUC) index of the proposed model reaches a value of 91.92%, while mean average precision values MAP@100 and MAP@200 are measured reaching 98.58% and 96.77%, respectively. Compared with other electricity theft classification models, the proposed model performs the electricity theft classification task better. The method in this paper can be extended to electricity theft intelligent identification.

Key words: residual network, convolution enhancement, self-attention mechanism, depthwise separable convolution, identifying electricity theft

中图分类号: