Detection Of Electricity Theft In Smart Grid Based On K-SMOTE And Improved Random Forest

With the rapid development of smart grid and the continuous advancement of electric energy measurement technology,the data on the user side of the grid presents a situation of high complexity and high redundancy,which also provides an opportunity for some illegal users of the smart grid to steal electricity.The losses caused by power theft on the user side of the power grid cannot harm the economic benefits of the power grid companies or even the steady development of the national economy.The traditional methods of power theft detection cannot meet the analysis and processing requirements of user power characteristics under diverse data.In order to assist power grid companies to improve the efficiency of power consumption inspection and manage users’ standardized power consumption,this paper proposes a user-side power theft detection method based on improved SMOTE oversampling technology and improved random forest method based on the analysis of the characteristics of power theft.First,based on the analysis of the electricity stealing behavior of a user based on a certain distribution network area,the user’s electricity stealing characteristics are extracted,and related electrical parameter characteristics such as voltage,current,and power factor are proposed.Secondly,considering the impact of power grid imbalance data on the detection accuracy of machine learning methods,a K-SMOTE oversampling method is proposed,that is,the K-means method is used to cluster user data sets to find the cluster centroids of the imbalanced data.Then,based on the combination of the centroid and the SMOTE oversampling method,clustering is used as the area for interpolation,and a balanced user-side power consumption data sample is constructed,which provides a data basis for the later detection of power theft to achieve the desired accuracy.Finally,a balanced forest dataset is used to train a random forest classifier.At the same time,in view of the limitations of the existing algorithms on the number of decision trees in a random forest,a particle swarm improved random forest method(P-RF)is proposed.The swarm optimization algorithm solves the tree of the optimal decision tree in a random forest.Through simulation and comparison experiments,and comparison with multiple detection models,it is verified from different perspectives that the power user stealing behaviordetection model proposed in this paper has good accuracy and effectiveness,so as to provide power companies with power stealing detection for users.More information reduces the cost of grid risk.