Federated Learning-Based Anomaly Detection Framework for Enhancing Security in Smart Grid Environments

Abstract

The growing integration of smart grids into modern energy infrastructures presents both unprecedented opportunities for intelligent power management and serious concerns regarding data privacy and cyber-security. Traditional anomaly detection methods, although effective, often rely on centralized data collection, thereby increasing the risk of exposing sensitive user information. To overcome these limitations, this thesis introduces FED-XID, a novel Federated Learning-based framework for privacy-preserving anomaly detection in smart grid systems. The framework enables decentralized model training using the XGBoost algorithm and incorporates embedded Intrusion Detection Systems (IDS) at the edge level, ensuring localized monitoring while safeguarding user data confidentiality. In addition, advanced deep learning techniques based on Temporal Convolutional Networks (TCN) and Long Short-Term Memory (LSTM) are applied to handle missing or incomplete smart meter data. The proposed model is both robust and efficient, leveraging a hybrid edge–cloud architecture and a lightweight classification core to ensure high performance, low latency, and scalable deployment in real-world smart grid environments. FED-XID achieved an AUC of 93.60 and a training time of only 25.97 seconds, demonstrating strong detection capability and computational efficiency.