TRANS-DEEP: an optimized deep learning framework for electricity theft detection in smart grids

Abstract

Electricity theft significantly impacts power utilities, causing financial losses and compromising system reliability. Traditional detection methods, such as manual inspections and basic comparisons, lack scalability and accuracy when applied to large datasets. This article introduces TRANS-DEEP, an advanced methodology that integrates sophisticated techniques to effectively detect electricity theft and overcome these limitations. The proposed approach includes data collection and preprocessing, followed by Variational Autoencoder (VAE)-based anomaly detection to identify irregularities in consumption data. The Adaptive Synthetic Sampling (ADASYN) method is employed for data balancing to address class imbalance, complemented by Generative Adversarial Networks (GANs) for dataset enrichment. A Bidirectional Long Short-Term Memory Network (BiLSTM) is used for feature extraction, capturing sequential and temporal dependencies. Finally, a Transformer-based model leverages self-attention mechanisms for effective theft detection by focusing on the critical features identified by BiLSTM. TRANS-DEEP was trained and tested to evaluate its ability to enhance detection accuracy and robustness, outperforming existing computer-based methods. The experimental results demonstrate the superiority of the proposed methodology, achieving high accuracy rates of 94.66% for Smart Meter Datasets (SMD)-I, 96.34% for SMD-II, and 95.12% for SMD-III in detecting electricity theft.