Abnormal Electricity Consumption Detection Based On Deep Semi-supervised Learning

Non-technical losses caused by abnormal electricity consumption have always affected the normal operation of the power grid in terms of economic and safety.Abnormal consumption detection,as the core issue of reducing the harm of nontechnical losses,has received continuous attention.With the comprehensive construction of China’s smart grid,the grid has accumulated a large amount of historical electrical data,laying a foundation for data-driven abnormal power detection methods.Although in recent years,big data analysis and machine learning technology have made great progress,data-driven abnormal consumption detection still has the following problems to be solved: 1)how to reduce the impact of poor data quality;2)how to integrate detecting algorithms with electrical knowledge and theory;3)How to enable the detecting algorithm to accurately and effectively recall abnormal samples with the insufficiency of cases;4)How to ensure that the model effectively learns the characteristics of abnormal consumption behavior with the presence of label noise.This article first converts electrical data which includes multiple data types.We innovatively use the probability distribution of electrical indexes to characterize the user’s consumption behavior,and visualize the distribution.This not only solves the common problem of missing data,but also converts the traditional time series anomaly detection problem into image detection and classification problems.Secondly,this paper designs a neural network structure that combines multi-channel features based on object detection,which can dig and understand the electricity consumption behavior in depth.In order to solve the problem of limited sample labels,this paper trains a deep semi-supervised learning model that is suitable for abnormal electricity detection based on consistency loss.With the aid of unlabeled data,the model’s generalization ability is improved.Finally,by modeling the probability distribution of gradients of neural network during training,this paper designs an algorithm that can effectively separate label noise and improve the signal-to-noise ratio of gradients,which effectively reduces the effect of label noise on the model generalization ability.The main contribution of this article is that it provides a new way of applying deep learning algorithms to a wider range of scenarios.And to solve the problem of label insufficiency and label noise,an effective method is proposed,which has certain reference significance.Through a series of experiments,this paper validates the effectiveness of each part of the detection method.The final test results show that the proposed method exceeds the state-of-the-art in the abnormal detection problem.The average AUC score can reach 0.94.