Research On Electricity Theft Behavior Detection In Neighborhood Area Network Of Smart Grid

A large number of intelligent hardware devices,such as smart meters,are deployed in modern power systems.While smart devices provide more advanced technologies,they also bring a lot of loopholes to the power system,so that malicious users can launch various cyber/physical attacks anytime and anywhere to steal electricity.Electricity theft not only brings huge economic losses to utility companies,but also may cause fires and other accidents,endangering public safety.Therefore,the research purpose of electricity theft behavior detection in smart grid neighborhood area network is to detect the electricity theft behavior in the neighborhood area network in time,and identify all malicious users.The main difficulties and challenges in the identification of electricity theft users are as follows: how to design a data organization method to process the one-dimensional historical electricity consumption sequences,so that the data can contain more regular information;how to more efficiently and comprehensively extract the users’ normal electricity consumption pattern characteristics from the input data;for the new sample data,how to design a detection algorithm to distinguish users’ normal/abnormal behaviors with the lowest possible cost,the highest possible accuracy,the lowest false positive rate and the lowest false negative rate.There are still some shortcomings in the proposed methods of electricity theft detection.The main contributions of this thesis are summarized as follows:(1)In view of the problems that the existing machine learning-based electricity theft detection methods do not fully learn the periodic characteristics of the users’ electricity consumption pattern,the detection accuracy is relatively low,and the false alarm rate is relatively high,Electricity Theft Detector based upon Convolutional Long Short-Term Memory(ETD-Conv LSTM)neural networks with high accuracy is proposed.For suspicious users in the community,first of all,it is necessary to reconstruct the one-dimensional time sequences of electricity consumption data with temporal correlations into a two-dimensional matrix with spatio-temporal correlations.This matrix is then divided into a series of submatrix sequences and fed into the ETD-Conv LSTM network.The network model combines the advantages of long short-term memory neural network and convolution mechanism,which can not only capture the time-dependent features(i.e.,global features)contained in the last row vector of matrices,but also can learn user’s spatial features(i.e.,local features)of differences in electricity consumption between adjacent dates and time periods.The experimental results show that the proposed ETD-Conv LSTM can more effectively capture the periodic characteristics of users’ electricity consumption,and the detection accuracy is as high as 93%.Compared with the existing algorithms,the detection accuracy is higher,while the false negative rate and false positive rate are relatively low.(2)In view of the limitation that the existing detection methods generally suffer from an implicit assumption that the malicious users always steal a large amount of electricity,and the detection accuracy is greatly reduced or even fails in the case of Small-amount Electricity Theft(SET),a Shewhart-CUSUM joint control chart based detector for SET electricity theft detection is proposed.It can effectively detect small-amount electricity theft attacks.The proposed detector consists of an electricity theft detection phase and a malicious user identification phase.During the electricity theft detection phase,the Shewhart and CUSUM control charts are applied to analyze the difference change between the electricity produced measurement of the central observer meter and the summation of all users’ reported electricity consumptions.If the Shewhart control chart detects data anomalies,it indicates that there is at least one malicious user in the community has launched Large-amount Electricity Theft(LET)attacks and/or there are several malicious users have launched SET attacks.If the CUSUM control chart detects data anomalies,it indicates the existence of at least one malicious user launching SET attacks.In the malicious user identification phase,the two control charts are combined to analyze the fluctuation of each user’s daily electricity consumption,aiming to exactly locate all malicious users of electricity theft in the community.On the whole,if the Shewhart/CUSUM control chart detects data anomalies,the corresponding user is a malicious user who launched LET/SET attacks.Extensive experiments are conducted to evaluate this method.Results show that it can detect SET attacks quickly and efficiently,and has lower false positive rate and false negative rate than existing algorithms.