Research On Identification Technology Of Abnormal Power Consumption Based On Data-driven

As an important support for the high-quality development of the national economy,electric power must be guaranteed during the rapid development of the smart grid.However,with the development of the economy and the increase in power load in recent years,the abnormal behavior of power consumption,mainly based on stealing,has become a factor that cannot be ignored affecting the development of the power system.The proliferation of abnormal power consumption behaviors seriously damages the economic interests of power companies,affects the fair and equitable power supply order,and even threat to the safe operation of the power grid.The popularity of smart meters has accumulated a large amount of power consumption data in the power metering system.Using a data-driven method to efficiently identify abnormal power consumption can improve the efficiency of power system staff in dealing with abnormal behavior.And change the current situation of time-consuming and labor-intensive regular manual inspections,and improve the operating efficiency of the entire power system.This paper starts from the problems in data-driven abnormal identification of power consumption.To solve the two problems of a large number of missing values and the lack of marked abnormal samples in the data set,establishes a supervised machine learning model suitable for power load data to realize the identification of abnormal power consumption.First of all,aiming at how to deal with the missing values in the power data set,proposes a model for filling missing values of electricity load data based on clustering and K-nearest neighbors.The model uses K-means clustering to mine potential patterns of power consumption data.Improve the KNN interpolation process according to the clustering results.The K value is dynamically optimized during the KNN operation,and the K value is self-adapted when the KNN interpolation is performed in an orderly manner.The verification shows that the missing value filling model of power consumption data in this paper performs better than the comparison model on different index,and realizes more reliable estimation of missing values.Next,this paper introduces the Time GAN network to solve the problem of the lack of labeled abnormal samples of power consumption.Combines it with OCSVM,and proposes a data augmentation model of abnormal power load based on Time GAN and OCSVM.Time GAN and OCSVM are trained using real abnormal power consumption samples,and then the sequence generated by Time GAN is input into OCSVM for classification and screening.The selected sequence is then verified by TSNE to expand the number of abnormal power consumption samples.It has been verified that the data generated by the model has certain usability.Finally,establishing a supervised learning power consumption anomaly recognition model based on feature engineering XGBoost.Using Tsfresh and the improved Boruta algorithm,the model extensively extracts and selects the time series features that contribute to the XGBoost classification decision.Using the selected features to train the XGBoost model,the identification of abnormal power consumption is realized based on the time series features of the power load data.The simulation results on the SGCC dataset show that the proposed method achieves an accuracy of0.9329,a specificity of 0.9371,and an AUC of 0.9505,and the Kappa coefficient achieves a score of 0.8722.Compared with the comparison model,it shows better recognition ability.