Publication

Abstract : The Smart Grid's security and privacy issues create significant challenges, especially in the context of blackouts and cascade failure. Though the recent development of an Intrusion Detection System based on the Federated Learning approach provides scalability, security, and privacy, it often lacks robustness and accuracy in multiclass attack detection. To address these challenges, this study proposes a Location-FedAvg, federated learning framework that integrates three key components of location-aware model aggregation to capture geographically specific attack patterns, Neighborhood Component Analysis (NCA) based feature extraction technique to enhance discriminative capability, and localized hyperparameter tuning at Phasor Measurement Units (PMUs) to improve adaptability across distributed sources. The proposed Location-FedAvg framework incorporates site-specific optimization, resulting in enhanced generalization while maintaining privacy, in contrast to the conventional FedAvg approach, which aggregates and updates the raw weights globally. This framework is validated on a three-bus/two-line transmission system using the publicly available Industrial Control Systems Cyber Attack Power System dataset developed by Mississippi State University and Oak Ridge National Laboratory, evaluated across 15 multiclass datasets from four geographic locations. The experimental findings reveal that Location-FedAvg achieves an average accuracy of 92.40 %, significantly outperforming the FedAvg-RF baseline model, which attained merely 35.74 % accuracy. Further, the model's robustness is validated by statistical analysis, with Cohen's Kappa values achieving greater than 85 % accuracy, which indicates consistent performance across locations. Overall, this study's results demonstrate that Location-FedAvg is a privacy-preserving, resilient framework for multiclass intrusion detection, providing scalable protection for smart grid infrastructures against evolving cyberattacks.