Kazmi, KashifKazi, Omeir Asim [22CO33]Shaikh, Faizan [22CO46]Shaikh, Qizarhayat [22CO56]Shinde, Abhijeet [22CO58]2026-08-202026-08-202026-06http://aiktc.ndl.gov.in/handle/123456789/4568This project addresses the critical challenge of electricity theft in power distribution systems, which causes significant financial loss and grid instability. Given the scarcity of real-world labeled fraud data, the primary objective was to develop a robust detection system capable of identifying illicit consumption patterns using smart meter data. Methodology The objective was achieved through a multi-staged technical pipeline: Data Processing : Raw smart meter data (collected at 3-minute intervals) was aggregated into daily consumption windows. To overcome the lack of labeled data, we designed and injected realistic theft scenarios— including intermittent reduction, sustained drops, and partial manipulation— into the clean dataset. Feature Engineering: We extracted a comprehensive suite of behavioral features, such as day/night consumption ratios, rolling statistical variances, and ”low-usage streaks,” to capture deviations from normal consumer habits. Hybrid Ensemble Modeling: We implemented a two-layered classification architecture. A Supervised Layer (Random Forest and XGBoost) identifies known theft signatures, while an Unsupervised Anomaly Layer (Isolation Forest and Local Outlier Factor) detects novel or statistical outliers. Risk Categorization: Results are merged into a three-tier risk model: Normal, Suspicious, and High Risk.en-USProject Report