A stacking ensemble with Pareto optimization for scalable electricity theft detection via hybrid data repair and lightweight deployment.

Electricity theft poses a significant challenge to grid reliability and utility revenues, while its detection using smart meter data is constrained by data quality issues, severe class imbalance, and practical deployment limitations. This study presents a stacking ensemble framework, termed Scalable Trustworthy Lightweight Network (STL-Net), for electricity theft detection (ETD) using smart meter data. The proposed framework integrates hybrid data repair, class imbalance handling, and temporal dimensionality reduction with a heterogeneous stacking ensemble composed of NGBoost, CatBoost, LightGBM, and XGBoost. Hyperparameters of the base learners are optimized using the Non-dominated Sorting Genetic Algorithm II (NSGA-II) to identify Pareto-optimal configurations that jointly consider predictive performance and model complexity before stacking. Model interpretability is supported through SHapley Additive exPlanations (SHAP), which provide transparent analysis of detection outcomes. Experiments conducted on real-world smart meter data demonstrate that STL-Net achieves an ROC-AUC of 0.9869 and an F1-score of 94.47%, outperforming a wide range of machine learning, ensemble, and deep learning baselines across multiple evaluation metrics. A lightweight variant, STL-Lite, preserves comparable detection performance (ROC-AUC: 0.9858) while reducing inference latency by approximately 40%, making it suitable for resource-constrained environments. These results indicate that the proposed framework effectively integrates accuracy, computational efficiency, and interpretability for ETD in smart grid applications.