Research On The Identification Method Of Electricity Theft Users Based On Machine Learning

The application of smart grids and Advanced Metering Infrastructure(AMI)enhances real-time communication between different components of the grid,but smart grids are vulnerable to electricity theft attacks,including physical methods such as tampering with meters or change wiring,as well as novel methods such as fake data injection,the diversity of electricity theft methods makes it significantly more difficult to detect electricity theft.A large amount of electricity consumption data provided by AMI system and smart meter make machine learning technology popular in the field of electricity theft detection.The user’s electricity consumption characteristics are mined from a large number of users’ electricity consumption data,and the users who are suspected of electricity theft are screened out by the classifier.However,the existing methods have problems such as incomplete extraction of electricity consumption characteristics,and low accuracy of electricity theft detection due to unbalanced electricity theft data.Therefore,an appropriate algorithm is designed to improve the identification accuracy of electricity theft user.Firstly,the research background and significance of this paper are described,and the research status of identification methods for electric theft users at home and abroad is summarized,and the main work is divided into the following four categories:method based on game theory,method based on state estimation,method based on hardware equipment and data-oriented method.After analyzing the shortcomings of existing methods,the main ideas of this paper are put forward.Two data sets commonly used in the field of electricity theft are analyzed: ISET(Irish Smart Energy Trail)and SGCC(State Grid Company of China),the data processing methods are introduced and the two data sets are preprocessed.Finally,the evaluation indexes used in the discrimination and detection of electricity theft are given.In view of the problem that the existing discrimination methods for electricity theft are not comprehensive in extracting users’ power consumption characteristics and cannot detect suspected users of electricity theft as soon as possible,the principles of Sideout Fusion Convolutional Neural Network(SFCNN)and Gated Recurrent Unit(GRU)are analyzed,SFCNN is used to extract local features and detail features of electric quantity signals,and GRU is used to extract dynamic features at the time level.Finally,the features of the two are fused to construct an electricity theft discrimination method based on the SFCNN-GRU hybrid model.The short-term electricity consumption data of the ISET data set is used to identify suspected users,and the model parameters are determined through comparative experiments.Simulation experiments are carried out on ISET data sets,and different models,different attack types of electricity theft and the model performance under unbalanced conditions are compared.The simulation results show that this method has a higher detection rate for identifying electricity theft users,and its performance is higher than other methods in the case of unbalanced conditions,and it can screen out the suspected electricity theft users with a small amount of historical data.Considering that single classifier such as support vector Machine(SVM)and decision tree(DT)have limited ability to process high-dimensional data and slow training speed,the integrated learning Light GBM(Light Gradient Boosting Machine)algorithm is proposed to detect electricity theft with high-dimensional power consumption data.The low detection accuracy is due to the imbalance between the number of normal users and electricity theft users.Focal Loss is used to improve the Light GBM Loss function to reduce the impact of imbalance,an electricity theft detection algorithm based on FLLight GBM(Focal Loss-Light GBM)is constructed.FLLight GBM algorithm is used to perform secondary discrimination on suspected users screened out by SFCNN-GRU hybrid model to achieve higher accuracy of electricity theft detection.The long-term electricity consumption data of ISET and SGCC data sets are simulated.The simulation experiments are carried out on long-term electricity consumption data of ISET and SGCC data sets.First,the experimental results of Focal Loss under different parameters are compared to select appropriate parameters.Then,the performance of commonly used classifiers such as Logistic Regression(LR)and Multilayer Perceptron(MLP)models under different training sets is compared,and the running time of each model is compared and the performance of each model under different proportion of ISET power consumption data is compared.It is verified that the FLLight GBM algorithm proposed in this paper improves the accuracy of the detection of electricity theft users and its excellent performance in processing high-dimensional data,which provides effective ideas for the accurate detection of electricity theft users.Finally,the developed multi-category laboratory reproduction test platform for electricity theft behavior is used to collect the data of electricity theft,and the electricity theft user identification software based on virtual instrument is constructed.Experiments are conducted on the measured data based on SFCNN-GRU electricity theft discrimination model and Focalloss-Lightg BM electricity theft detection model respectively.Experimental results of different training set ratios and common models are compared to prove the effectiveness and feasibility of the proposed method.