基于CNN与Bi-LSTM的异常用电行为甄别算法研究  

Research on discrimination algorithm of abnormal power consumption behaviorbased on CNN and Bi⁃LSTM

作　　者：余向前 张磊 胡晓祥 YU Xiangqian;ZHANG Lei;HU Xiaoxiang(Zhangye Power Supply Company of State Grid Gansu Electric Power Company,Zhangye 734000,China)

机构地区：[1]国网甘肃省电力公司张掖供电公司,甘肃张掖734000

出　　处：《电子设计工程》2024年第9期96-100,共5页Electronic Design Engineering

基　　金：国网甘肃电力公司科技项目(522707220004)。

摘　　要：为降低非技术损耗给电网运行带来的损失,文中对用户用电行为的甄别技术进行了研究,并提出了一种基于卷积神经网络(CNN)和双向长短期记忆(Bi-LSTM)网络的模型。该模型一方面使用CNN中的卷积、池化运算提升对用电数据中隐性特征的挖掘效率,另一方面利用Bi-LSTM处理长时间序列的优势,弥补了CNN在时序分析上的不足。同时还引入了Dropout机制与Adam优化方法,提升了网络的训练速度,避免了CNN和Bi-LSTM结合后因网络结构复杂而造成的过拟合现象。在自建数据集上进行的仿真结果表明,所提算法的Precision、Recall及F1值均取得了显著提升,且相较于单一的CNN和Bi-LSTM网络,F1值分别提升了10.04%和8.32%。此外,该算法训练与测试的F1值仅相差0.03%,说明算法的鲁棒性较强,未出现过拟合的现象。In order to reduce the losses caused by non-technical losses to power grid operation,the paper studies the discrimination technology of user’s power consumption behavior,and proposes a network model based on Convolutional Neural Network(CNN)and Bi-Long Short-Term Memory(Bi-LSTM)network.On the one hand,the model uses the convolution and pooling operations in CNN to improve the mining efficiency of hidden features in power consumption data,and on the other hand,it uses the processing advantages of Bi-LSTM in long-time series,It makes up for the shortage of CNN in time series analysis.The Dropout mechanism and Adam optimization method are also introduced into the network to improve the training speed of the network and avoid the over fitting phenomenon caused by the complex network structure after the combination of CNN and Bi-LSTM.The simulation results on the self built data set show that the Precision,Recall and F1 values of the algorithm in this paper have been significantly improved.Compared with the single CNN and single Bi-LSTM networks,the F1 values have increased by 10.04%and 8.32%respectively.In addition,the difference between the training and test F1 values of the algorithm is only 0.03%,which indicates that the algorithm is robust and there is no fitting phenomenon.

关 键 词：反窃电 CNN LSTM 时间序列处理 ADAM DROPOUT 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]