Advanced deep learning-based electricity theft detection in smart grids using multi-dimensional analysis with Convolutional Autoencoder and Transformer

Electricity theft poses a significant threat to power grid stability, leading to substantial economic losses and safety hazards for utility providers. Therefore, effective electricity theft detection (ETD) is critical for ensuring the cost-efficiency and security of smart grids. While machine learning (ML) techniques utilizing electricity consumption (EC) data from smart meters have shown potential, traditional methods often rely on single-dimensional analysis, lacking the granularity to capture complex, periodic EC patterns, which results in suboptimal detection performance. To address these shortcomings, we propose a novel ETD system that utilizes deep learning (DL) through a combination of Convolutional Autoencoder (CAE) and Transformer models for advanced multi-dimensional analysis of EC data. We further tackle EC data imbalance with the K-means-based Synthetic Minority Oversampling Technique (K-means SMOTE). Our approach preprocesses EC data, employing the CAE to extract detailed features from 28-day consumption intervals, followed by the Transformer to analyze sequences and capture comprehensive patterns. The proposed ETD enables precise differentiation between anomalous, non-periodic theft signatures and consistent, periodic, legitimate usage. Evaluated on the State Grid Corporation of China dataset, our proposed model achieves an impressive F1 score of 0.9918, surpassing existing benchmarks. Notably, its lightweight design ensures scalability and efficiency, broadening its applicability to other smart grid anomaly detection tasks, such as identifying equipment failures or cyber threats, thus reinforcing modern power grid security.