Detection and localization of power theft based on CNN-LSTM algorithm

Abstract

Electricity theft is an extremely prominent problem plaguing the power sector in smart grids. As smart meters become more popular, comparing the electricity consumption data from the customer's smart meter and the main meter can present a possibility for detecting and locating electricity theft. In this work, a deep learning method based on long short-term memory (LSTM) and improved convolutional neural network (CNN) are proposed to detect and locate possible electricity theft behaviors. By extracting feature vectors related to electricity consumption behavior from historical intelligent meter data, a model for determining possible electricity theft behavior is learned and trained, and electricity consumption behavior is predicted from new real-time electricity consumption data to identify smart meters with large deviations in electricity consumption behavior. The CNN-LSTM algorithm is verified to have good prediction ability for electricity consumption behavior and possible electricity theft behavior under normal conditions through comparative analysis of different algorithm simulations.