Transient stability assessment—determining whether a power system maintains synchronism after large faults—is fundamental to grid reliability and operator decision-making. While machine-learning models offer speed advantages over traditional simulation, their development has been hampered by the scarcity of high-quality, open datasets that integrate realistic system dynamics with network structure and machine parameters.
A new benchmark dataset addresses this gap. Containing 20,000 three-phase-to-ground fault scenarios on the IEEE 9-bus system, the dataset pairs steady-state AC power-flow conditions with detailed electromagnetic-transient simulations showing post-fault rotor-angle and speed trajectories for all three generators. Each record includes the network as an attributed graph with nine buses and eighteen branches, ten node features and twelve edge features, generator constants, fault location and type, and a binary stability label.
The scenarios span eighteen fault locations with varied load and generation levels, yielding a nearly balanced class distribution (48.96% stable, 51.04% unstable) suitable for supervised learning. Full reproducibility is ensured through deterministic generation with fixed random seeds and publicly available code. The dataset is hosted on IEEE DataPort with a persistent digital object identifier (DOI).
This benchmark supports multiple research directions: transient stability classification, generator trajectory prediction, stability margin estimation, and critical-clearing-time calculations. Critically, it enables rigorous comparison of topology-aware graph neural networks, physics-based surrogate models, and hybrid approaches that combine domain knowledge with data-driven learning.
For power system researchers and machine-learning practitioners, the dataset removes a major barrier to advancing autonomous stability assessment tools. As utilities increasingly rely on faster decision support systems and grid operators face tighter stability margins from high renewable penetration, AI surrogates validated on standardized benchmarks become essential infrastructure. This open-source contribution accelerates that transition while maintaining scientific reproducibility and transparency.
