An AI explained systematic modular approach for enhanced Electricity Theft Detection for urbanized Smart Grid environment

Electricity theft is a major issue that disrupts utilities and compromises the stability of urbanized Smart Grids (SG), leading to significant financial losses, operational disruptions, and compromised system stability. Traditional detection systems often face challenges like severe data discrepancies, high class imbalance, and volatile consumption patterns, which hinder their ability to detect abnormal movements and limit system dependability. To tackle these challenges, we propose a systematic modular approach that consists of three modules, Module-1: Data Pre-processing, Module-2: Data Balancing, and Module-3: Blending Classification. Initially, the data pre-processing module imputes missing values with interpolation, handles outliers with the three-sigma rule, and normalizes data with min–max scaling methods to enhance data quality for robust model training. Subsequently, the data balancing module leverages the localized randomized affine shadow sampling to maintain a balanced data distribution by addressing the class imbalance problem. Finally, the proposed SATBlend in the classification module utilizes AlexNet for feature extraction, ShuffleNet for efficient computation, and a temporal convolutional network for temporal correlation detection to enhance the reliability of advanced ETD systems. In addition to its application on the state grid corporation of China dataset, we evaluate SATBlend on the Pakistan residential electricity consumption dataset, demonstrating its robustness and scalability across varied real-world datasets. The experimental results reveal that SATBlend outperforms state-of-the-art machine and deep learning techniques with improved accuracy of 4% and 2%, recall of 2% and 1%, F1-score of 2% and 1%, and receiver operating characteristic area under the curve of 3% and 2% for both datasets, respectively.Further, the 10-fold cross-validation method validates the main findings to prove the generalizability of the SATBlend model. Finally, to improve both local and global interpretability and explainability of SATBlend’s predictions, explainable artificial intelligence techniques are employed, including local interpretable model-agnostic explanations and Shapley additive explanations. The results show that the well-constructed SATBlend model is effective in improving the detection of electricity theft while strengthening SG’s operational and infrastructure capabilities.