Abstract
<jats:p>Power grids are increasingly exposed to extreme events, creating challenges for fault diagnosis and resilient operation. This paper proposes an interpretable adaptive feature-enhanced random forest (AFE-RF) framework for fault classification, localization, and severity assessment. Three-phase voltage magnitudes and phase angles are used as inputs, while an accuracy-guided feature-enhancement strategy improves difficult fault categories and locations. A severity ranking for resilience-oriented assessment is further introduced to rank critical fault–bus pairs. Validation on the IEEE 9-bus and IEEE 39-bus systems shows classification accuracies of 99.89% and 99.74% and localization accuracies of 99.14% and 98.86%, respectively. Feature-importance and decision-tree analyses provide transparent and physically meaningful interpretations of the model decisions. The proposed framework provides a lightweight and interpretable tool for resilience-oriented fault assessment under extreme-event conditions.</jats:p>