Research On Abnormal Power Consumption Detection Method Based On Deep Learning

With the rapid growth of the economy and the rapid development of scientific and technical,residents have grown to electricity,and the power grids capacity is expanding.The development speed of power grids is also increasing.However,in recent years,the user’s abnormal electrical behavior is increasingly frequent,and it is timely and effectively detect an abnormal electrical behavior in the power system to discover malicious abnormal electrical users,reducing economic losses,ensuring that the power system is safe and reliable,maintaining the steady development of national economies is critical.With the deployment of senior metering infrastructure,massive large-scale high-dimensional complex power consumption data is accumulated,but there are problems of missing value and low abnormal electrical detection accuracy.How to effectively impute the missing value of electricity consumption data to improve data quality,and effectively detect the exceptional electricity user,has been widely concerned in recent years.The purpose of the paper is to improve the accuracy of imputing the power consumption data missing value and the accuracy of detection.Considering the correlation between high-dimensional power energy data proximity features and potential spatial characteristics,this paper analyzes the electricity consumption data in combination with deep learning,improves abnormal electricity detection accuracy,the main work content is as follows:(1)Focusing on the problem that the existing missing data imputation model lacks considering neighboring data characteristic correlation,and the missing data imputation value does not meet the real data distribution,the multi-scale sliding window is introduced on the basis of the missing value imputation algorithm.The multi-scale sliding window aims to adaptively extract the characteristics of timing data,capture potential correlation in the neighboring data.We propose a generative adversarial network model with a multi-scale sliding window(MSW-GAN)for missing value imputation.We introduce a prompt matrix in the MSW-GAN generator input to speed up the model training,suppress the discriminator capability.The experiment is carried out on a plurality of data containing timing characteristics,verifying that the proposed model is more accurate to impute the missing value of electricity consumption data.(2)For the existing electricity theft detection model,the timing characteristics and spatial characteristics of the electricity consumption data set are not taken into account,a mixed deep learning model(CAEs-LSTM)combined of convolutional autoencoder and long short-term memory network is proposed to detect users with abnormal electrical behaviors.The model learns the timing characteristics of electricity consumption data through long short-term memory networks,spatial characteristics through convolutional autoencoder,and perform Gaussian white noise processing to improve model anti-noise capacity.The experiment is conducted by the real data set published by the Power Grids,and the results show that the mixed model effectively improves the detection accuracy.