The grid can't be protected by AI it can't train on: a new open dataset of 40,000 faults
Only 1.7 percent of 119 papers reviewed on AI-enabled power system protection actually published the data behind their results. That single number — pulled from a recent survey cited in this paper — explains why machine learning is transforming almost every other industry faster than it transforms the electric grid. Researchers keep building clever neural networks to detect faults before they cascade into blackouts, but they cannot compare their work against anyone else's, because nearly everyone keeps their training data locked away.
Now a team from Friedrich-Alexander-Universität Erlangen-Nürnberg and the Ostbayerische Technische Hochschule Amberg-Weiden has built EvEMTBench, an open, physically grounded simulation dataset of roughly 40,000 grid fault and event recordings, and released it for the whole research community. It is not a small thing. The researchers generated waveform data at 9600 Hz across four literature-standard grid topologies spanning voltage levels from 20 kV to 345 kV, plus 105 randomly generated grids per voltage level, then labeled every recording with what happened and where. It is, by their own assessment, the first public dataset to meet the field's own stated requirements for diverse topologies, raw signal export, and comprehensive documentation (Kordowich et al., 2026).
The problem EvEMTBench addresses is quietly urgent. The global push to decarbonize is rewiring the physics of the power system. Inverter-based resources — solar farms, wind turbines, battery storage — do not behave like the spinning synchronous generators that grids were engineered around for more than a century. Their fault currents are smaller, faster, and harder for traditional protection relays to read. As the paper puts it, the rapid integration of distributed renewable energy resources "increases the operational complexity of the grid and causes significant challenges." Machine learning is the leading candidate to take over the job of watching millions of synchronized measurements and deciding, in milliseconds, what just happened. But machine learning starves without data.
The Science
The conventional protection relays that keep the grid safe today are designed around a physics that is disappearing. A synchronous generator, when a fault hits, delivers a large, predictable surge of current that a distance relay can triangulate into a location. An inverter-based resource behaves differently: it caps its current, shifts its phase, and responds on timescales conventional relays were never built to interpret. This is why researchers increasingly want to hand the problem to machine learning models that can extract features from the raw waveforms themselves — features "inaccessible to conventional methods," the authors note.
The obstacle has never been the algorithms. It has been the absence of a shared training ground. Real-world fault data is nearly impossible to collect at scale: grid faults are rare by design, and Evdakov and colleagues (cited in the paper) point out that manual labeling is expensive and error-prone. So researchers have turned to simulation. But most simulation datasets fall short in some way. Some target steady-state load flow, with minute-level resolution that cannot capture events lasting milliseconds. Others work in the phasor domain, which smooths away exactly the transient detail a protection algorithm needs. Several excellent datasets cover only a single grid topology, so models trained on them cannot be shown to generalize. None, the authors argue after a careful survey, satisfies the full checklist their colleagues laid out: diverse topologies, fault scenarios and operating conditions; raw signal export; and comprehensive labeling and documentation.
EvEMTBench is built to satisfy all three. The framework treats DIgSILENT PowerFactory as its simulation engine, automated through a Python API. The authors ran electromagnetic transient (EMT) simulations — physics-accurate, sub-millisecond models of how voltage and current actually swing during an event — rather than simpler phasor-domain approximations. This matters because phasor models "cannot capture transients from many events relevant for power system protection, and phasor domain data do not allow sub-cycle analysis," limiting how fast a detection algorithm can possibly react. The EMT simulations run at a 10-microsecond step, capturing the frequency range of 0.1 Hz to 3 kHz where the most relevant protection transients live.
The team focused on modeling "Group I" frequencies (0.1 Hz to 3 kHz) because these cover the transient phenomena that matter for extended protection applications, and because accurately modeling higher frequencies would demand a level of detail that is both computationally prohibitive and impossible to parameterize from published literature. Each simulation runs for 1.5 seconds; the first second is discarded to let initialization transients settle, and the final half-second is exported. At each monitored cubicle the dataset records synchronized point-on-wave voltages and currents for all three phases — six channels per measurement point, sampled at 9600 Hz, yielding 4801 samples per channel. The entire measurement is captured in a tensor form, with one dimension per cubicle, one per channel, and one for time.
Crucially, the team validated their simulation parameters against established literature rather than inventing them. "The realism of synthetic datasets critically depends on appropriate model selection and parameterization," they write, so they conducted an extensive literature review to fix realistic ranges for every component — transformers, synchronous machines, cables, overhead lines, loads. Where the true distribution of a parameter is unknown, they sample from uniform distributions, which is the established way to maximize diversity.
The dataset comes in three flavors, each serving a different machine-learning purpose. The "multigrid" dataset contains 10,000 simulated events across 105 randomly generated topologies per voltage level, with every class of parameter — event, operating state, and grid structure — varied to train models that can generalize to grids they have never seen. The "benchmark" dataset runs events in four grids drawn directly from the literature, with event parameters varied systematically at discrete values (fault location at 1%, 20%, 50%, 80%, and 99%), so that any two models can be compared fairly. And the "adaptgrid" dataset, generated for each literature grid, offers a middle path for fine-tuning: fixed topology, randomly varied events and operating states.
The four literature grids were chosen deliberately. The CIGRE Benchmark MV Grid operates at 20 kV. The Protection Reference Grid runs at 110 kV subtransmission. The IEEE 39-bus New England system — the paper's only 60 Hz grid — operates at 345 kV transmission. And a simple double-line grid at 110 kV provides a stripped-down starting point for proof-of-concept work. Some famous test systems were excluded simply because their modeling depth is too shallow for EMT simulation; the IEEE 9-bus system is out for exactly that reason.
What They Found
The dataset simulates a sweeping catalog of events — 22 distinct types, all listed in the paper's event table. Operating events mirror the mundane and consequential moments of everyday grid life: load switches opening and closing, capacitors connecting, overhead lines and cables energizing, transformers inrushing, induction motors starting, and the loss of generation from both conventional synchronous machines and inverter-based resources. Fault events cover the full spectrum of short circuits — single line to ground, line to line, double line to ground, and three-phase — with impedances up to 50 ohms.
But two event types deserve special attention because they target the specific weaknesses of modern protection. High-impedance faults (HIFs) occur when a conductor contacts something that resists current flow — a tree branch, dry soil, pavement. The resulting fault current can be so small that conventional relays simply never notice, which is how energized lines end up lying on the ground undetected. EvEMTBench simulates these faults with resistances up to 150 kilo-ohms, sampled in four distinct impedance bands from 50 ohms to 150,000 ohms, with equal probability across bands. The authors treat these as primarily a medium-voltage challenge, so they are only simulated at MV level.
Even more unusual are "incipient faults," a relatively new and thorny problem in the field. These are self-extinguishing, high-impedance faults that flash and vanish within four cycles — a staple precursor to permanent failures, but one that leaves little trace. Detecting them early is like hearing a small crack before the dam gives way. EvEMTBench simulates them with and without arc modeling.
For the faults, the physics is rendered with unusual care. Single-line-to-ground faults often produce arcs, and the team models those arcs using the Kizilcay arc model. The arc voltage and resistance scale with arc length, following equations the paper reports: arc voltage and arc resistance .
shows the overhead line models used in the simulations alongside photographs of their real-world equivalents, grounding the abstract models in physical reality.
The operating constraints are realistic too. The team verifies load flow convergence, caps transformer, line, and machine loading at 110%, and holds voltage deviations within conventional bounds: 0.95 to 1.05 per unit for medium-voltage grids, 0.90 to 1.10 per unit for high- and extra-high-voltage grids. Simulations that fail to converge or go numerically unstable are discarded, along with their parameter sets, so every recording in the published dataset is a physically valid one.
Why This Changes Things
The chart below tells the urgent story: for decade-defining tasks like fault detection, the review literature published usable data in fewer than 2 in 100 papers (1.7%). EvEMTBench is designed to break that deadlock by giving the whole field a common, open test track.
The deeper significance is about generalization — the thing that has quietly stymied real-world machine learning in power systems. A model trained on a single simulated grid tends to fail on any grid that differs even slightly from the one it learned. Real grids have "parameter imperfections, and models of connected elements such as consumers or IBR controllers are typically unknown," the authors note. A model that only ever saw idealized simulations treats the real world as out-of-distribution noise and stumbles.
EvEMTBench's multigrid dataset attacks this head-on using domain randomization — deliberately varying topology, operating state, and event parameters so widely that real-world data lands somewhere inside the distribution the model has seen, rather than outside it. The three dataset types together support the full machine-learning lifecycle: pretraining for cross-grid generalization, benchmarking for fair comparison, and fine-tuning for deployment on a specific utility's own grid. This mirrors how foundation models work in language and vision, and it could carry the same transformative logic into protection engineering.
There is also a practical-scale achievement here. The authors build their synthetic transmission and distribution grids not as full-scale networks but as small sections, on the sensible reasoning that "transients from electrically distant events generally do not influence protection systems, which are therefore relatively local." This is what lets them monitor every bus without drowning in data — a modeling philosophy that keeps the dataset tractable while preserving physical fidelity where it matters. The team also publishes each grid's topology as a machine-readable NetworkX graph and as a PowerFactory file, so researchers can inspect the exact structure they are training on, not just the waveforms.
The dataset's timing is fortunate. Virtualized protection platforms and the IEC 61850-9-2 Sampled Values protocol are already streaming synchronized point-on-wave data from digital substations in the real world. EvEMTBench speaks the same native language — synchronized per-phase voltage and current at high sample rates — so models trained on it can in principle transfer to the data modern substations actually produce. The simulation-to-real gap is real, and the authors are honest about it: performance "may deteriorate in real-world applications" and often requires domain adaptation or fine-tuning with grid-specific data. But a physically grounded, openly documented training set is precisely the foundation a field needs to start closing that gap.
What's Next
EvEMTBench is an infrastructure contribution, and like all good infrastructure, it opens more doors than it closes. The immediate opportunity is for the machine-learning community to use the benchmark dataset to start settling long-debated questions: which architectures generalize across topologies, which tasks are genuinely solvable from waveform data alone, and where the limits of simulation-based training really lie.
Several open questions remain. The dataset emphasizes European grid configurations, with the IEEE 39-bus system as its lone 60 Hz representative — a reasonable start, but a reminder that the world's grids are not one grid. The authors deliberately exclude cables from their high- and extra-high-voltage grids, noting most transmission circuits are overhead, which is true in general but not everywhere. The simulation-to-real gap, while honestly acknowledged, is measured nowhere yet; the dataset's ultimate validation will come when models trained on it are tested against real oscillograms like those in the existing Grid Event Signature Library or RTE's waveforms.
Most of all, EvEMTBench recasts what is possible when a field stops rationing its data. The 1.7% statistic was not a technical failure — it was a cultural one. Every researcher who builds on EvEMTBench, extends it, or simply benchmarks against it is helping normalize the idea that in power system protection, as in every other field, progress should be reproducible, comparable, and open. The grid of the future will be watched by algorithms. This dataset is an early, careful attempt to make sure those algorithms are trained honestly.