Machine Learning for Online Transient Stability Assessment

Machine learning (ML) offers a computationally tractable alternative to time-domain simulation for online transient stability assessment (TSA), yet inconsistent evaluation practice limits reproducibility and cross-study comparison. We benchmark 17 ML classifiers on IEEE 9-bus and New England 39-bus power systems under a unified evaluation protocol, showing that tree-based ensemble models—Extra Trees (99.22%/98.15% accuracy) and LightGBM (99.10%/98.12%)—consistently outperform parametric and non-parametric approaches across accuracy, recall, and area under the receiver operating characteristic curve (AUC) with training times under two seconds. We further quantify the latency-accuracy tradeoff between snapshot and trajectory-based TSA. A five-second post-fault observation window recovers 85-100% accuracy using physics-informed temporal features, while classical ensembles match Long Short-Term Memory (LSTM), Temporal Convolutional Network (TCN), and Transformer architectures within 1.4% at 30- 100× lower training cost. We introduce Tranpy, an open-source Python package with pre-trained models, labeled datasets, and PowerFactory integration for custom dataset generation, addressing critical reproducibility gaps in ML-based TSA research.