Deep Attention-based Neural Network for Electricity Theft Detection

Abstract

Electricity theft causes significant harm to social and economic development. In recent years, as a powerful technique in data mining, deep learning has attached much attention and become popular in electricity consumption sequence analysis. Nevertheless, existing methods mainly focus on short-term numerical data modeling, while the records in real-world scenarios (1) usually consist of multiple temporal features and (2) are often of large scale. In this paper, to overcome the two fundamental challenges, we propose a novel method called Deep Attention-based Neural Network for Electricity Theft Detection (DANN-ETD). Specifically, we first respectively decompose the electricity sequences into the trend, seasonal and residual views to fully exploit the temporal features. To effectively and efficiently model the large-scale time series, we then split the series into several snapshots and further design the deep attention-based recurrent neural networks which can detect the fine-grained evolution of electricity consumption. Experimental results on realworld datasets demonstrate that our method outperforms the state of the arts.