Abstract:

Electricity theft is a critical issue causing significant revenue losses for utility companies and disrupting energy distribution. Traditional detection methods often rely on manual inspections or simplistic analytics, which are inefficient and prone to errors. This project presents a multi-faceted approach for electricity theft detection, combining anomaly detection, hybrid machine learning algorithms, and audit analytics. A synthetic dataset representing customer consumption, meter tampering, and voltage fluctuations is utilized to simulate real-world scenarios. The project proposes three hybrid models: VoltGuard Ensemble, AmpereShield Detector, and GridSentinel Hybrid, which integrate Isolation Forest, One-Class SVM, KNN, LightGBM, Random Forest, and XGBoost techniques. These models are designed to capture subtle anomalies, local neighborhood patterns, and nonlinear feature interactions. Preprocessing steps include feature scaling, label encoding, and engineered anomaly scores. The system is implemented with a Python Flask backend, HTML/CSS/JS frontend, and MySQL database for storage. Model evaluation uses accuracy, precision, recall, F1-score, and confusion matrices. This hybrid approach improves detection performance compared to existing machine learning methods. The project emphasizes explainability, allowing insights into feature contributions for theft predictions. It provides a scalable, efficient, and automated framework to detect electricity theft in diverse consumption environments. The proposed methodology demonstrates high accuracy and robustness, minimizing false positives and enabling timely intervention. This system can assist utility providers in revenue protection, smart grid monitoring, and customer behaviour analysis. Future work can extend to real-time monitoring, integration with IoT smart meters, and adaptive learning with evolving consumption patterns.

Keywords

Electricity theft, anomaly detection, hybrid machine learning, Isolation Forest, One-Class SVM, KNN, LightGBM, Random Forest, XGBoost, consumption analysis