Voltage stability in modern smart grids faces increasing challenges due to the widespread use of renewable energy and diminished reactive-power margins. While power flow analysis remains the most precise method, it is often too slow and resource-intensive for exploring extensive operating spaces. This paper introduces a physics-based machine learning approach that combines MATPOWER simulations with an ensemble classifier to efficiently generate clear and interpretable instability risk maps for the IEEE-14 system. By varying load levels, renewable penetration (represented as negative PQ-bus injections), and specific network stress factors, operating scenarios are created; a scenario is deemed unstable if power flow fails to converge or if the lowest bus voltage falls below 0.94 p.u. Trained on a balanced dataset with approximately 40% unstable cases, the model achieved ROC-AUC = 0.973 and PR-AUC = 0.715 through five-fold cross-validation, with well- calibrated probabilities. Feature analysis identified load level and renewable penetration as primary causes of instability. The model delivers results thousands of times faster than traditional methods while maintaining high accuracy, enabling practical screening, enhanced risk understanding, and targeted use of CPF for final margin assessment.
Wattegama et al. (2026) studied this question.