GANs generate power grid layouts from geographic data

A new paper from Juan Manuel Garcia-Perez and Carlos Mateo proposes using Generative Adversarial Networks (GANs) to automatically create power distribution network layouts from image-based geographic data. The framework is trained on rasterized views of existing distribution systems extracted from GIS sources. It supports two modes: an unconditional mode that learns general topology patterns from the training dataset, and a conditional mode that incorporates street maps and the spatial distribution of consumers. The model was tested on low voltage (LV), medium voltage (MV), and high voltage (HV) feeder designs, successfully reproducing realistic topologies that align with the underlying geography. The approach offers a data-driven complement to traditional heuristic-based synthetic network generators, which often fail to reflect real-world patterns.

The study also identifies key limitations: training stability remains a challenge, resolution-dependent artifacts appear in generated layouts, and the framework currently lacks explicit electrical constraints such as power flow validation. The authors note that future work must integrate electrical modeling to ensure generated networks are physically operable. Despite these hurdles, the method shows promise for automating distribution network design in new electrification projects—enabling utilities to quickly propose layouts based on existing infrastructure patterns and geographic context.