FedTheftDetect: Optimizing Anomaly Detection in Smart Grid Metering Systems Using Federated Learning

Abstract

The detection of anomaly energy consumption patterns in smart grid metering systems remains a critical issue. This is due to data imbalance, privacy constraints, and the dynamic nature of consumption patterns. To address these concerns, we present a privacy-preserving and scalable anomaly detection framework named as FedTheftDetect framework. The proposed framework integrates deep learning algorithms into a federated learning (FL) architecture through the incorporation of advanced ensemble classifiers to detect behavioral anomalies in daily consumption patterns. A real-world smart meter dataset with significant class imbalance is used to assess the suggested framework. The dataset had significant preprocessing to identify consumption-related anomalies in behavior. Experimental results demonstrate that the suggested framework outperforms the competitive centralized and distributed models. It achieves significant improvements in Accuracy, Precision, Recall, and F1-score, all of which are close to 0.95, which indicates a great predictive capability and reliability.