A Heterogeneous Graph-Based Multi-Task Learning for Fault Event Diagnosis in Smart Grid

TL;DR

The paper tackles the challenge of timely and accurate fault diagnosis in distribution grids by introducing a heterogeneous multi-task learning graph neural network (MTL-GNN) that jointly performs fault detection, localization, type classification, resistance estimation, and current estimation. It leverages a common GCN-based backbone to learn topological and feature representations from voltage phasors on the IEEE-123 feeder and employs a GNNExplainable method to identify a sparse, informative subset of measurement nodes. The approach demonstrates robustness to measurement noise, variable fault resistance, topology changes, and limited observability, while achieving strong performance across all tasks and enabling practical deployment with reduced sensing requirements. This work advances real-time grid fault management by unifying multiple diagnostics in a single model and providing actionable explainability for sparse sensing and operator decision-making.

Abstract

Precise and timely fault diagnosis is a prerequisite for a distribution system to ensure minimum downtime and maintain reliable operation. This necessitates access to a comprehensive procedure that can provide the grid operators with insightful information in the case of a fault event. In this paper, we propose a heterogeneous multi-task learning graph neural network (MTL-GNN) capable of detecting, locating and classifying faults in addition to providing an estimate of the fault resistance and current. Using a graph neural network (GNN) allows for learning the topological representation of the distribution system as well as feature learning through a message-passing scheme. We investigate the robustness of our proposed model using the IEEE-123 test feeder system. This work also proposes a novel GNN-based explainability method to identify key nodes in the distribution system which then facilitates informed sparse measurements. Numerical tests validate the performance of the model across all tasks.

Figures (10)

• Figure 1: Diagram of IEEE-123 node feeder system. The highlighted blue blocks represent the node pairs that are considered connected. The number of voltage regulators, transformers, and switches is mentioned in ($\cdot$) and the number of buses, their phases and load connectivity are mentioned in the table. The active substation (source bus) is connected to the node $150r$.
• Figure 2: t-Distributed Stochastic Neighbor Embedding (t-SNE) visualization of all the data points. (Left) shows the dataset generated with a variable range of fault resistance sampled from a uniform distribution $\mathcal{U}(0.05 \Omega, 20 \Omega$). (Right) shows the dataset generated with a constant fault resistance $20 \Omega$.
• Figure 3: Visualization of the sequence of procedures for a single data point and label generation. The double-circled digits represent the loop iteration points in the algorithm. By performing all the iterations in a hierarchical manner (the digits specify the order of the hierarchy) the entire dataset is generated.
• Figure 4: Architecture of the proposed heterogenous MTL-GNN. The input features go through the common backbone GNN to generate graph embeddings. For visual clarity message passing across the layers for a couple of nodes (highlighted in green) is shown, where the blue highlighted sections signify nodes included in the message passing process (Left). The embeddings generated by $128$ nodes are flattened and concatenated together to convert to a one-dimensional vector (Middle). The concatenated feature vector is passed to $5$ heads, three classification heads and two regression heads. The corresponding loss is computed based on the predicted output ($\hat{y}_{t}^{k}$) and ground truth label ($y_{t}^{k}$). The computed loss for each task is weighted and summed together (Right).
• Figure 5: Confusion matrix for fault type classification task. The first three classes are asymmetric faults, the next two classes are symmetric faults and the final class corresponds to non-fault events.