Aiming atthe variety of electric larceny means,the efficiency of electric larceny detection remains improvement.Firstly, the fuzzy C mean clustering algorithm is used to construct different load characteristic curves of the user, and the suspiciouselectric larceny user is preliminarily determined by comparing the curves to be detected with the corresponding characteristic curve.Secondly,the particle swarm optimization support vector machine regression model is adopted to detect the behavior of suspected power stealing users.The experiments show that this method can reduce the range of electricity larcenydetection and overcome the influence of less electricity larcenysamples, improve the efficiency ofelectricity larcenydetection, and increasethe mean square error and average absolute error by 00051 and 0.034 respectively.