Hybrid Edge–Cloud CNN Framework for Real-Time Fault Detection and Localization in Distribution Networks

Abstract

Timely and accurate fault detection and localization are essential for reliable operation of distribution networks. This paper presents a hybrid edge–cloud framework that integrates convolutional neural networks (CNNs) with edge computing to achieve real-time performance. The proposed method distributes computational tasks such that edge devices handle data acquisition, preprocessing, and CNN-based inference, while cloud servers manage model retraining and historical data storage. The CNN architecture comprises three convolutional layers with ReLU activation, max-pooling, and two fully connected layers optimized for lightweight inference. A 33 kV distribution network model was used to generate fault scenarios, including single line-to-ground, double line-to-ground, line-to-line, three-phase, and three-phase-to-ground faults under varying resistances and loads. Experimental results show that the proposed framework achieves 100% fault-type classification accuracy, an average fault localization error of 0.18 km (vs. 1.25 km for impedance-based methods), and a 50% latency reduction compared to cloud-only implementations. These results confirm that the framework enhances both responsiveness and resilience, offering a scalable solution for modern distribution network fault management.