A game-level difficulty-aware curriculum learning framework for progressive graph neural network training in distribution network fault diagnosis

Abstract

Abstract Graph neural networks (GNNs) have shown considerable potential for distribution network fault diagnosis; however, conventional empirical risk minimization with randomly ordered samples may lead to unstable convergence and substantial performance degradation when complex network topologies and severely corrupted measurements are introduced during the early stages of training. To address this issue, this paper proposes a game-level difficulty-aware curriculum learning framework for progressive GNN training. The framework first performs offline sample difficulty evaluation by combining degree-distribution entropy, which characterizes topological complexity, with inverse signal-to-noise ratio (SNR), which represents measurement corruption. The resulting coupled difficulty score is then used by an online pacing function to progressively introduce training samples from easy to hard. Dynamic loss reweighting is further incorporated to balance previously learned easy samples and newly introduced difficult samples throughout training. The framework is evaluated on seven-category fault diagnosis datasets generated using the IEEE 33-bus and IEEE 123-bus distribution systems under different measurement noise levels, transition resistances, load variations, and missing-data conditions. On the IEEE 33-bus system at 20 dB SNR, the proposed method achieves an accuracy of 91.55% and a Macro-F1 score of 91.20%, outperforming the standard GCN by 11.10 and 11.32% points, respectively. It also enters a relatively stable loss region approximately 25 epochs earlier than standard GCN training and maintains improved diagnostic performance when up to 30% of the nodal measurements are missing. These results demonstrate that the proposed framework improves the accuracy, convergence stability, and robustness of GNN-based distribution network fault diagnosis. The present SNR-based difficulty measure uses clean reference signals available in controlled simulations; practical black-box deployment therefore requires estimated noise indicators, such as filtering residuals, blind SNR estimation, or model-based predictive uncertainty.