An edge-cloud collaborative fault diagnosis method for high-resistivity faults in low-voltage distribution networks considering metering uncertainties

To address the issue of weak characteristics and susceptibility to environmental noise and measurement errors in high-impedance faults (HIFs) in low-voltage distribution networks, this paper proposed an edge-cloud collaborative fault diagnosis method (ECCFD) that considers the metering uncertainty of smart meters. First, to address the common problems in low-cost smart meters such as transformer errors, clock asynchrony, and quantization noise, a statistical modeling method for measurement uncertainty is established. This uncertainty is encoded into multi-dimensional node features through time window partitioning and feature statistics, improving the robustness of feature representation. Second, an edge-cloud collaborative diagnostic architecture for low-voltage distribution networks is constructed. At the edge, high-order harmonic ratios, order components, and statistical features are extracted and normalized to reduce communication load and mitigate the impact of load fluctuations on feature stability. Then, a graph model is constructed in the cloud based on the distribution network topology. A Bayesian graph convolutional neural network (BGCN) is introduced to perform joint inference on node states. Variational inference and Monte Carlo sampling are used to estimate fault probability and prediction uncertainty. Finally, simulation and comparative experiments verified that the constructed method can still maintain high diagnostic accuracy and stability under conditions of high noise and significant measurement uncertainty, providing technical support for robust and reliable diagnosis of high-resistance faults in the actual operation of smart grids.