Visualization and Root Cause Analysis of Power System Faults Based on Explainable Machine Learning

Addressing the challenges of unidentifiable coupling mechanisms and the "black-box" nature of models in power system fault analysis, this study develops a causal exploration framework based on Random Forest surrogate models and SHAP visualization technology. The framework transforms the fault prediction task into a deconstruction of the internal interaction structures of the data, utilizing interpretable machine learning to assist in mechanism analysis. Through empirical research on multi-source heterogeneous data—including oil temperature, ambient temperature, electrical load, and power quality—this paper employs global contribution measurement and nonlinear interaction dependency visualization to intuitively reveal a critical disaster mechanism: high ambient temperatures significantly reduce the load tolerance threshold of electrical equipment. Furthermore, by applying a time-series micro-attribution method, a potential evidence chain for specific oil temperature trip faults is successfully constructed, tracing the progression from environmental background to specific trigger factors. The results demonstrate that the proposed scheme effectively uncovers complex physical laws hidden within massive monitoring data, achieving a transition from associative pattern recognition to mechanistic causal inquiry . This provides a novel visual perspective for fault tracing and proactive prevention in power systems.