A massive literature review argues graph-based machine learning is a natural fit for modern power grids, and calls out the field for still lacking the shared benchmarks to prove it out.
The survey, posted to arXiv, examined nearly 800 papers combining graph machine learning with power systems research. It covers use cases across forecasting, state estimation, optimization, control, fault diagnosis, and cybersecurity. The authors argue grid topology - the physical wiring of how substations, lines, and generators connect - makes graphs a natural data structure for these problems, unlike the traditional model-based solvers that struggle to keep up with real-time demands. They also built a requirements catalog meant to guide future dataset development for the field.
Power grids are absorbing more renewables, more distributed generation, and more real-time decision-making than the software managing them was built for. Graph-based methods promise faster, topology-aware approximations that could complement classical solvers without replacing the physics-based rigor grid operators depend on. But the paper's own numbers cut against the hype: despite nearly 800 papers, real-world deployments remain limited and standardized public benchmarks are scarce, which makes many published results hard to reproduce.
In other words, the research pipeline is full, but the field still owes utilities the boring part - open datasets and reproducible tests - before any of this touches a live grid.