Data Driven Electricity Theft Detection Based on Federated Learning

Electricity theft is the main cause of non-technical loss of the power network, and it is a long-term problem faced by distribution network operators. The electricity consumption information acquisition system collects a large amount of data from electricity users, which makes it possible to carry out data driven detection of abnormal electricity consumption and identify users with electricity theft behaviors accurately. Thus, this paper presents a data-driven method to perform electricity theft detection, which combines deep learning with federated learning effectively. It is capable to not only detect electricity theft precisely, but also achieve secure data sharing among distribution network operators. The proposed method is verified on an actual Irish open-access dataset, and the results of the case studies show that it can obtain a comparatively higher detection accuracy, meanwhile ensuring data privacy of different distribution network operators.