This paper proposes a novel hybrid ensemble machine learning framework integrated with Explainable AI (XAI) to overcome the fundamental trade-off between non-detection zones (NDZ) and false positives in conventional islanding detection methods. The approach introduces a cascaded dual-stage architecture: first, an Isolation Forest algorithm serves as a high-sensitivity anomaly trigger to minimize the NDZ by detecting nearly all operational deviations; second, an optimized multi-class XGBoost classifier, activated only upon an anomaly, precisely discriminates islanding events from other disturbances like faults and switching operations. Validated on a comprehensive public dataset, the framework demonstrated a high islanding detection rate of 96.0% while reducing false positives to a near-zero rate of 2.7% through adaptive confidence thresholding. Integrated SHAP-based interpretation provides operational transparency by quantifying feature contributions to each decision. This work introduces a novel cascaded framework that uniquely integrates an unsupervised anomaly detector (Isolation Forest) with a multi-class XGBoost classifier and SHAP-based explainability for islanding protection. Unlike prior hybrid methods, the proposed architecture explicitly decouples sensitivity near-zero NDZ and specificity (near-zero false positives) while providing full decision transparency, offering a more reliable and interpretable solution compared to conventional single-model approaches.
Shahade et al. (Fri,) studied this question.