Research On Abnormal Power Users Detecting Methods

According to studies,operating losses caused by non-technical problems in the power system are up to 10 billion dollars every year.Non-technical losses refer to operational losses caused by a series of false electricity customers' behaviours,such as electricity stealing and traud by power customers on the distribution network side.With the continuous advancement of smart grids and the rapid development of sensor acquisition technology,power companies increase the amount of electricity load data,this will lead the detection of abnormal customers more and more difficult.In recent years,some intelligent detection algorithms have been proposed to overcome the disadvantages of high blindness and low precision of original manual detection.These algorithms improve the hit rate of field detection and reduce the operation cost.At present,most of the intelligent detection algorithms are based on supervised learning,and a large number of tagged training sets are required.However,in reality,the initial stage of data analysis and detection does not have a large number ot training sets for model training,so this thesis proposes an abnormal electricity behavior detection model based on unsupervised and semi-supervised which aims to form a list of customers' suspicious degree sorting,providing the inspection focus for the field detection and improving the accuracy of field detection.The work of this thesis mainly includes:(1)In order to solve the problem of lack of customers' electricity types(normal and abnormal)in the early stage of detection.An outlier based anomaly detection model framework is proposed by using the unsupervised learning method.The framework includes the first level grey list generation algorithm based on clustering analysis and the two level grey list generation algorithm based on outlier calculation.Using this model can generate a list of customers with suspicious rank.Experiments using real power data show that the proposed model generates a list of suspicious degrees without training set,and the first 30%customers of the list can get a high recall rate of exception.That is,the higher the ranking of customers in the list,the more suspicious.(2)In order to detect crimes committed by customers groups,this thesis uses Anomaly Detection Model Based outlier,and combines a semi supervised learning.It proposes an abnormal customers' detection model based on semi-supervised learning.In this model,the three level grey list generation algorithm based on behavior similarity calculation is proposed.This algorithm is used to detect the suspicious customers with similar behavior of the abnormal behavior in the blacklist database.The experimental results show that the detection efficiency of the anomaly detection model using semi-supervised learning is significantly higher than that with unsupervised learning.