From black box to decision support: An interpretable clustering-based ensemble for dynamic security assessment in modern power systems

Real-time Dynamic Security Assessment (DSA) is a critical step toward achieving resilient operation in modern power systems. Although machine learning (ML) techniques are regarded as promising tools for this task, conventional ML approaches are often limited by their “black-box” nature, failing to provide the interpretability required for high-stakes operational decisions. This paper proposes a novel DSA framework that integrates contingency clustering, a stacking ensemble model, and Explainable AI (XAI) to enable reliable, real time dynamic security assessment of power systems. Unlike standard ensembles, our approach employs “Dynamic Signature Decomposition,” utilizing Critical Clearing Time (CCT) profiles to partition contingencies into dynamically similar groups. Specialized base learners trained for each cluster are then intelligently aggregated by a meta-learner. We leverage a multi-level application of SHapley Additive exPlanations (SHAP) to unlock the model's reasoning, transforming it into a transparent decision-support tool that provides actionable explanations identifying the key physical parameters driving instability. Validated on the IEEE 68-bus test system and scaled to the IEEE 127-bus system, the proposed framework achieves superior F1-Scores (0.9500 and 0.9380, respectively), outperforming strong monolithic baselines. By uniting high predictive accuracy with physics-aware feature attribution, this work provides operators with evidence-based intelligence for proactive grid control, proving that interpretability and performance can be mutually reinforcing.