Abstract: Accurate identification method of abnormal electricity users can provide reference for power supply

enterprises to lock in electricity theft or other violations of power users. Most abnormal user identification models

based on machine learning adopt unsupervised algorithms，but the accuracy of the models is low. To solve the

above problems，a two-stage abnormal power user identification method combining unsupervised local outlier

factor（LOF）algorithm and supervised support vector machine（SVM）algorithm was proposed. Based on the

analysis of the current and voltage performance of the abnormal energy meter different from the normal energy

meter，the input characteristics of the abnormal identification model were constructed. The LOF algorithm was used

to sample，and the suspicious samples were selected and handed over to manual labeling. Then the supervised SVM model was trained by the labeled samples. In the subsequent detection work，the suspicious samples screened by LOF algorithm were directly sent to the SVM model for identification. The example results show that this method

has high identification accuracy for power abnormal users，and has guiding significance for the power stealing

inspection of power supply enterprises.

      Key words: abnormal electricity user identification；machine learning；local outlier factor（LOF）；support

vector machine（SVM）