Collaborative Machine Learning Model Improves Grounding Fault Localization in Distribution Networks

Engineers have developed a collaborative fault localization framework combining an improved Wasserstein Generative Adversarial Network with Gradient Penalty and edge federated learning to pinpoint single-phase grounding faults in medium- and low-voltage distribution networks. The research, conducted by Chun Ma, Tian Zhang, Wei Kang, Yu Guo, Yongzhan Ma, and colleagues, targets the growing operational complexities in modern power systems where traditional fault localization techniques exhibit low accuracy and poor generalization. In intelligent distribution networks requiring rapid self-healing and continuous operational safety, conventional approaches struggle against limited fault records, multi-source data heterogeneity, inadequate extraction of temporal features, and constraints on computational resources at edge terminals.

To overcome these challenges, the investigators designed an architecture that addresses data scarcity and feature extraction simultaneously while preserving network operational privacy. An improved Wasserstein Generative Adversarial Network with Gradient Penalty is integrated into the workflow to generate synthetic fault samples and extract temporal characteristics from electrical disturbances.

Alongside the generative model, the architecture integrates edge federated learning to enable distributed model training across edge terminals located throughout the power distribution grid. Rather than requiring participating terminals to transmit raw electrical measurements across the network to a central server, the system coordinates collaborative updates locally. The researchers introduced adaptive weighting and a dedicated region-weighted module directly into the Federated Averaging aggregation algorithm, enhancing the global model aggregation accuracy while effectively handling heterogeneous data originating from diverse network sectors.

The evaluation framework utilized two distinct datasets to benchmark model behavior: the standard IEEE 123-node simulation test feeder and a dataset of field-recorded events. The simulation verification dataset contained more than 10,000 electrical samples, providing a broad range of simulated operational conditions. To assess performance under practical operating environments, the authors incorporated an on-site measured dataset comprising 326 real single-phase grounding fault events collected directly from physical distribution systems.

Analysis of the generative network demonstrated reliable sample synthesis, achieving a Fréchet Inception Distance of 12.36 and a Peak Signal-to-Noise Ratio of 27.94 dB during fault sample generation. Across the combined verification data, the overall collaborative localization framework achieved an average fault localization accuracy of 94.85%.

The researchers also tested the system under challenging physical conditions to assess operational robustness against typical field disturbances. Under scenarios involving high grounding resistance, the fault localization accuracy reached 82.70%, while performance during arc fault events achieved 84.06%. When subjected to sudden load variations across the network, the localization accuracy remained at 89.37%, and under high-noise operating conditions, the system maintained an accuracy of 82.25%, demonstrating measurable anti-interference capabilities across diverse operating states.

In comparative analyses against existing localization techniques, the proposed model exhibited advantages in total parameter size, inference runtime, and communication bandwidth consumption. Because raw electrical measurements remain at local edge devices, the federated approach eliminates the requirement to upload sensitive grid telemetry, balancing data privacy constraints with the resource limits of field-deployed edge computing terminals. The authors report that the lightweight design and anti-interference profile make the framework suitable for complex on-site operating environments, providing practical technical support for rapid fault identification, on-site troubleshooting, and autonomous self-healing in medium- and low-voltage distribution networks.

Subject of Research: Electrical engineering and machine learning for power distribution network fault localization

Article Title: Collaborative localization of WGAN-GP and edge federated learning for single-phase grounding faults in distribution networks

Article References: Ma, C., Zhang, T., Kang, W., Guo, Y., & Ma, Y. (2026). Collaborative localization of WGAN-GP and edge federated learning for single-phase grounding faults in distribution networks. Discover Applied Sciences. https://doi.org/10.1007/s42452-026-09782-9

Image Credits: AI Generated

DOI: 10.1007/s42452-026-09782-9

Keywords: distribution networks, fault localization, single-phase grounding faults, Wasserstein Generative Adversarial Network, WGAN-GP, edge federated learning, Federated Averaging, adaptive weighting, region-weighted module, IEEE 123 simulation dataset, Fréchet Inception Distance, Peak Signal-to-Noise Ratio, smart grids, edge terminals, power system reliability, data privacy