Resource-Efficient Federated Deep Learning Framework for Real-Time Electrical Fault Diagnosis and Localization

The increasing complexity of modern power systems and the growing penetration of distributed energy resources have intensified the need for accurate and real-time electrical fault diagnosis. However, traditional centralized deep learning approaches require large volumes of data to be aggregated at a single server, leading to high communication costs, privacy risks, and significant computational burden. This paper presents a resource-efficient federated deep learning framework designed to perform real-time electrical fault classification and fault location prediction without requiring raw data transfer. The proposed architecture integrates lightweight model partitioning, selective layer freezing, and gradient quantization techniques to minimize resource consumption while maintaining high diagnostic accuracy. Experiments conducted on simulated IEEE bus systems and real-world grid fault datasets demonstrate that the proposed framework achieves 97.3% classification accuracy and reduces localization error to 1.18 km, outperforming standard federated learning (FL) and centralized methods. Furthermore, communication overhead is reduced by 68%, and edge-level energy consumption decreases by 39%, enabling deployment on low-power substation and field devices. These results confirm that the optimized federated learning framework offers a highly scalable, privacy-preserving, and computationally efficient solution for next-generation intelligent grid fault diagnosis and location prediction.