Fault Location in Power Distribution Networks With Massive Missing Data: A Graph-Based Imputation and Contrastive Learning Approach

Accurate fault location is essential for reliable operation of distribution networks, yet it remains challenging due to incomplete data caused by insufficient sensor deployment, sensor malfunctions, and communication failures during fault events. Existing Graph Neural Network (GNN)-based methods often assume complete data availability or resort to zero-fill imputation. This work proposes a two-stage fault location pipeline for data deficiency environments. A training-free graph-based imputation algorithm (GDIA) is proposed, capable of restoring missing values by directly minimizing graph Dirichlet energy without relying on prior data distribution. Additionally, a theoretical bound is established, linking changes in Dirichlet energy to imputation error. Subsequently, GNN models are employed to perform fault location using the imputed data. To address class imbalance arising from data deficiencies, a graph contrastive learning technique is integrated, further enhancing fault location accuracy. Experimental results demonstrate the proposed method’s robustness and effectiveness, highlighting its potential to enhance fault location accuracy in practical distribution network applications. Note to Practitioners—This paper addresses the critical challenge of fault location in distribution networks under massive missing data. Traditional methods relying on Graph Neural Networks (GNNs) often assume complete data availability, making them impractical for real-world scenarios. The paper suggests a novel way to address the issue by proposing a training-free graph-based imputation algorithm (GDIA) for data imputation, following the graph constrastive learning framework (GCL) for fault location considering class imbalance. GDIA efficiently reconstructs missing data by minimizing the Dirichlet energy on the network graph, ensuring structural consistency while avoiding reliance on prior data distributions. Theoretical analysis shows that higher Dirichlet energy correlates with reduced imputation accuracy, guiding practical implementation. For data imputation, since the low deployment of sensors and traditional deep learning based imputation methods cannot perform without the prior distribution, this paper proposed GDIA, which minimize dirichlet energy on graph to impute missing data and theoretically prove its viability that when graph Dirichlet energy becomes larger, the imputation accuracy would correspondly becomes lower. For fault location, GCL tackles class imbalance presented in fault datasets by training the GNN model to recognize the nodes that exhibit similar patterns to their neighbors, while faults occurence disrupt these patterns. By amplifying discrepancies between faulty and normal nodes, GCL can address the issue and enhance the fault location accuracy. Validated on the IEEE 33-bus system under massive data missing ( $\geq 60$ %), our GDIA-GCL framework achieves significantly higher fault location accuracy compared to existing methods. This approach enables utilities to enhance grid reliability without costly sensor upgrades, as it operates effectively on existing infrastructure. While the current method assumes homogeneous network, ongoing work focuses on adapting to heterogeneous grid reconfigurations and scaling to larger networks.