A Benchmark Graph Dataset for Transient Stability Assessment of the IEEE 9-Bus System: 20,000 Scenarios with Full Generator Trajectories

Power grids are the invisible circulatory system of modern life, a vast network of wires and machines that must move electricity from where it is made to where it is needed without ever stumbling. At the heart of this system are massive spinning generators, heavy steel rotors that turn in perfect unison, like a team of rowers keeping time on a single stroke. When a sudden shock hits the grid—a lightning strike, a broken wire, or a short circuit—these machines are jolted. The critical question for engineers is whether they can absorb that shock, stay in step with one another, and keep the lights on, or if they will fall out of sync and cause a cascading failure that plunges cities into darkness. This moment of truth, known as transient stability, happens in a fraction of a second, and predicting it has traditionally required running slow, complex computer simulations for every possible disaster scenario.

For years, researchers have hoped that artificial intelligence could learn to spot these dangerous moments instantly, acting as a fast substitute for the slow simulations. But progress has been stalled by a lack of good data. Most available datasets were like snapshots: they showed the state of the grid before a problem and a simple "pass or fail" result after, but they left out the vital details of how the machines actually moved during the crisis. Without seeing the full motion of the rotors and the specific physical properties of the machines, it is difficult to teach a computer the deep physics of why a grid holds together or falls apart. To solve this, a team of researchers at the Institut de recherche d'Hydro-Québec has released a massive, open collection of twenty thousand simulated disaster scenarios, designed specifically to bridge the gap between raw data and the laws of physics.

The researchers built their testbed on a standard, small-scale model of a power grid known as the IEEE 9-bus system, which contains nine connection points and three generators. Instead of just looking at the grid in a calm state, they subjected this model to a rigorous series of stress tests. They created thousands of different starting conditions, varying how much electricity was being used and how much was being generated, ensuring the grid was tested under both light and heavy loads. Then, for each unique starting point, they introduced a sudden, three-phase short circuit at one of fifteen different locations on the network. These faults were allowed to last for different amounts of time, ranging from a fraction of a second to nearly half a second, before being cleared. The team then ran detailed, high-speed computer simulations to watch exactly how the three generators reacted, recording their speed and position every millisecond for three full seconds after the fault.

The result is a dataset that captures the complete story of each event. Unlike previous resources that might only say "stable" or "unstable," this collection includes the full trajectory of every generator, the exact electrical conditions before the fault, and the specific physical constants of the machines, such as their weight and resistance. The team found that the outcomes were not random; they followed a clear physical logic. When the fault lasted too long or hit a critical spot, the generators would accelerate away from each other, losing their synchronization. The researchers defined a clear boundary for this failure: if the angle between any two generators swung apart by more than one hundred and eighty degrees, the system was declared unstable. In their simulations, the stable scenarios stayed well within this limit, while the unstable ones spun wildly out of control, sometimes rotating more than three full turns relative to each other.

What makes this work particularly valuable is the balance and richness of the data. The team generated a near-perfect split between stable and unstable outcomes, with roughly forty-nine percent of the scenarios ending in a stable grid and fifty-one percent resulting in a loss of synchronism. This balance prevents artificial intelligence from simply guessing the answer based on a majority vote. Furthermore, the data reveals that the location of the fault matters immensely. Some spots on the grid are far more dangerous than others; for instance, a fault on one specific line caused instability in nearly seventy-three percent of the cases, while a fault on another line caused failure in only twenty-seven percent. This proves that the grid's structure itself carries a hidden signal about its vulnerability, a nuance that machine learning models can now learn to recognize because the data includes the network's shape alongside the machine movements.

The researchers also took great care to ensure that the data was trustworthy and reproducible. Every single scenario was generated using a fixed, deterministic process, meaning that if another scientist runs the same code, they will get the exact same results. The dataset is organized as a graph, a way of representing the network where the connection points are nodes and the wires are edges, allowing modern computer programs to study the grid's topology directly. By providing the full history of the machines' movements alongside the network map and the physical constants, the team has created a unique resource that allows scientists to test different types of learning methods. They can now compare approaches that rely purely on data patterns against those that try to embed the laws of physics into the learning process, or hybrid methods that combine both.

This release does not claim to solve the problem of grid stability for the entire world, as the test was performed on a small, specific model. However, it provides a rigorous, shared benchmark that was previously missing. It allows researchers to stop arguing over whether their methods work and start comparing them on a level playing field. The dataset is now available to the public, offering a clear, detailed view of twenty thousand moments where the grid teetered on the edge of collapse. By making the full motion of the machines and the structure of the network available in one place, the researchers have given the scientific community the tools needed to build faster, smarter, and more reliable ways to keep the power flowing when the unexpected happens.