Research On Data-driven Abnormal Electricity Consumption Detection Method

The increasing loss of electrical energy is one of the significant challenges faced by global power distribution facilities.Reducing energy waste and economic losses caused by abnormal electricity consumption behaviors is of great practical significance.The growing installation of smart meters has enabled power companies to collect massive amounts of electricity usage data,which includes user consumption patterns and behaviors,and provides the foundation for datadriven research on abnormal electricity consumption detection,In this study,focusing on abnormal electricity consumption detection,we propose corresponding solutions to issues such as feature extraction and detection methods,imbalanced data,and reuse of multiple user models.The main contributions and innovations of this study are as follows:(1)To discover deep and multi-scale information in electricity consumption data,an empirical mode decomposition and entropy features based feature extraction method is proposed.This method decomposes the original electricity consumption data using empirical mode decomposition,extracts entropy features from the sub-wavelet and original data at different scales to construct feature vectors.To improve the utilization of data,a multi-scale convolution module based feature fusion method is proposed,which fuses feature vectors with the original data through multiple different convolution layers.In addition,we use graph convolutional networks to explore the correlation information between electricity consumption data and establish the mapping relationship between electricity consumption data and consumption behaviors.Experimental results on an electricity consumption dataset demonstrate that the proposed method has high accuracy and robustness.(2)In practical applications,the amount of abnormal electricity consumption data are far less than that of normal data,which will lead to the model tends to learn the majority class data and ignore the minority class data during training.However,the insufficient learning ability of the model for minority class data will seriously affect the effect of abnormal electricity consumption detection.To solve the problem of imbalanced data,this paper proposes a synthetic minority oversampling based abnormal electricity consumption detection method.This method employs the synthetic minority oversampling technique to achieve the expansion of minority class data to obtain the balanced dataset,and then uses the balanced dataset to train the abnormal electricity consumption detection model.In the experiments,the abnormal electricity consumption detection results under different imbalanced ratio are compared,and this method is compared with other data generation methods.Experimental results show that this method is superior to other methods in the accuracy of anomaly electricity consumption detection.(3)When detecting abnormal electricity consumption of multiple users,different users have different electricity consumption rules and behaviors,so it is necessary to establish different detection models for different users.This process undoubtedly increases the burden of model training,prolongs the training time,and raises the requirement of hardware equipment.To tackle this problem,this paper proposes a transfer learning and self-attention mechanism based abnormal electricity consumption detection method.The self-attention layer in this method can pay different attention to the information of different importance degrees in the input data,so that the model can focus on the more important information and improve the ability of the model to process data.With the help of the "pretraining-finetuning" method in transfer learning,the model is trained with the source domain data to get the pretrained model.Then,a small amount of electricity consumption data of multiple users are used as the target domain data to finetune the model and to get the detection model of each user.This process does not need to retrain the detection models of different users,but only needs to use a small amount of data to finetune the pretrained model to get the final model,which greatly reduces the training burden of the model and shortens the training time.To verify the performance of the proposed method,the electricity consumption data of several users are used as the target domain dataset to carry out abnormal electricity consumption detection experiments,and experimental results prove that the proposed method has a high detection accuracy.