Unsupervised Anomaly Detection for Energy Theft Identification in Smart Grids using Using Isolation Forest and Autoencoder Models

The paper is concerned with the critical problem of energy theft in smart grids, which is an issue causing a substantial loss of money and implying the inability to ensure a high level of operational efficiency. The large amount of data generated by smart meters has rendered conventional detection techniques ineffective. In this work, we will use two contemporary unsupervised anomaly detection techniques on UCI Smart meter data: Isolation Forest and Autoencoder. These models will be developed, tested, and their performance in detecting fraudulent energy consumption behavior will be compared. Strict preprocessing, feature engineering, and an integrated thresholding method are used in the study to provide robust evaluation metrics including accuracy, precision, and recall. The results also show that the Isolation Forest outperforms the Autoencoder in a number of areas, including accuracy and recall, which can be utilized to create intelligent monitoring systems that improve the dependability of energy network operations. Future improvements, including hybrid and deep learning models, can be based on the research, despite issues, including the use of opensource data and massive computations.