In this paper, we present a study on multi-label fault classification in smart grid power systems using a LightGBM-based approach for fault identification. The proposed method is capable of detecting simultaneous faults involving ground (G) and phase conductors (A, B, and C) using three-phase current and voltage measurements obtained from existing datasets, as well as, derived datasets generated from real-world-inspired, large-scale synthetic data. Additionally, we conduct an ablation study in which twenty-five features are ranked according to their contribution to classification accuracy. The study reveals that symmetric component features, including zero-sequence and negative-sequence components, along with vector norms, are among the most discriminative feature groups for smart grid fault identification. The experimental results demonstrate that models trained on the 4.2-million-row synthetic bulk dataset achieve a subset accuracy of 97.89%–98.55% and a Hamming accuracy exceeding 99.4%, substantially outperforming models trained solely on existing datasets. The benchmark dataset used in this study was obtained from Kaggle, where cross-dataset evaluation achieved a subset accuracy of approximately 85.76%. These findings establish best-practice guidelines for scalable, accurate, and overfitting-resistant smart grid fault diagnosis systems.