Abnormal detection of electricity theft using a deep auto-encoder Gaussian mixture model
DOI:DOI: 10.19783/j.cnki.pspc.211659
Key Words:stealing electricity  unsupervised learning  deep auto-encoder Gaussian mixture model  augmented Dickey Fuller test  decoupling
Author NameAffiliation
LIU Zhaorui 1. College of Electrical and Information Engineering, Hunan University, Changsha 410082, China
2. Hangzhou Haixing Electrical Co., Ltd, Hangzhou 310011, China) 
GAO Yunpeng 1. College of Electrical and Information Engineering, Hunan University, Changsha 410082, China
2. Hangzhou Haixing Electrical Co., Ltd, Hangzhou 310011, China) 
GUO Jianbo 1. College of Electrical and Information Engineering, Hunan University, Changsha 410082, China
2. Hangzhou Haixing Electrical Co., Ltd, Hangzhou 310011, China) 
LI Yunfeng 1. College of Electrical and Information Engineering, Hunan University, Changsha 410082, China
2. Hangzhou Haixing Electrical Co., Ltd, Hangzhou 310011, China) 
GU Dexi 1. College of Electrical and Information Engineering, Hunan University, Changsha 410082, China
2. Hangzhou Haixing Electrical Co., Ltd, Hangzhou 310011, China) 
WEN Yizhang 1. College of Electrical and Information Engineering, Hunan University, Changsha 410082, China
2. Hangzhou Haixing Electrical Co., Ltd, Hangzhou 310011, China) 
Hits: 661
Download times: 101
Abstract:Considering the applicability of unsupervised methods for user-side electricity theft detection, this paper studies how to solve the decoupling problem between feature extraction and anomaly detection. It proposes a user-side electricity theft detection method based on the deep auto encoder Gaussian mixture model (DAGMM). First, the electricity consumption data dimension with stationarity is obtained according to the augmented Dickey Fuller test. Then, potential characteristics of data are extracted by compressing the network. An estimation network and Gaussian mixture model are used to obtain sample energy. This reflects the degree of anomaly. Finally, network parameters are optimized jointly based on end-to-end learning to avoid model decoupling, and identify users whose sample energy exceeds the abnormal threshold as electricity thief. In this way theft of electricity can be detected. The experimental results show that the detection method based on DAGMM is less affected by the sample of electricity theft, and the extracted features can effectively reflect the user's electricity consumption law with higher detection accuracy. Compared with the existing methods, the detection rate, false detection rate, F1 measurement and AUC of the proposed method are significantly improved. This work is supported by the National Natural Science Foundation of China (No. 51777061).
View Full Text  View/Add Comment  Download reader