A novel combined deep learning-based technique for electricity theft detection in a smart grid

Abstract

Energy theft represents a major challenge for electricity distribution companies worldwide and constitutes a key component of non-technical losses (NTLs), alongside faulty meters and billing inaccuracies. Despite the rise of artificial intelligence techniques for electricity theft detection, existing approaches exhibit limitations in classification and data analysis when applied to large-scale datasets. These shortcomings can reduce the effectiveness and precision of the theft detection systems. Therefore, this article proposes a bidirectional long short-term memory (Bi-LSTM) with the lookahead adaptive moment estimation (LADAM) optimizer and a support vector machine (SVM) with a bidirectional gated recurrent unit (Bi-GRU) for the detection of energy theft in a smart electricity grid to assess energy supplier companies. In this work, the proposed model is implemented on the State Grid Corporation of China (SGCC) dataset. In the proposed framework, the Bi-LSTM is used for feature extraction, while the LADAM optimizer is employed for feature selection and normalization. The refined feature set is subsequently fed into the SVM classifier to distinguish between theft and non-theft instances. The obtained findings highlight the superior efficiency of the proposed strategy compared with those in the literature, highlighting its potential for practical deployment in real-world smart grid systems.