Researchers have trained one AI model to predict power flow across three separate electrical grids, using about 100 times fewer training updates than the reference method it improves on.
The system builds on GENCO, an existing power-flow correction network, by swapping two of its local correction steps for a module that shares information through two reduced graphs. The team tested three ways of building those graphs: a Kron-derived transport, a same-anchor Quotient construction, and a flat baseline with no hierarchy. All three were trained under one shared protocol, with fewer than 1,900 scenarios per grid and 200 epochs, then tested on 200 new scenarios per grid the models had not seen during training. The Kron version came out on top, cutting voltage error 51.3% below a simple per-bus reference and beating that reference on 98.5% of 600 fresh test scenarios; the flat baseline, by contrast, never beat the reference on any grid.
That gap matters because power-flow models are usually trained separately for each grid, which is expensive and does not scale as utilities add more grids to monitor. Here, one set of learned parameters handled three different grid topologies at once, and the version that organized information hierarchically, rather than flatly, won consistently across all three seeds tested. It is a small but concrete signal that grid topology, not just grid-specific data, can shape a useful inductive bias for these models.
The catch: none of the models yet beat the reference on two topologies they were not trained on, so this is generalization across scenarios on known grids, not a plug-and-play model for grids nobody has trained it on.