SafePowerGraph: A safety-aware benchmarking framework for scalable graph neural networks in power flow and optimal power flow

Power grid operations require solving power flow and optimal power flow problems efficiently under continuous perturbations. Existing graph neural network benchmarks for power systems overlook safety constraints and realistic failure scenarios, limiting their operational relevance. We present SafePowerGraph, the first simulator-agnostic, safety-oriented framework and benchmark for graph learning in power grid operations. The framework integrates four widely used simulators: PandaPower, MATPOWER, PowerModels.jl, and OpenDSS, and evaluates models under three perturbation scenarios: in-distribution load variations, energy price fluctuations, and N-1 power line outages. SafePowerGraph introduces heterogeneous graph representations with physics-informed loss functions that enforce alternating current power flow equations and operational boundary constraints. Experiments across five benchmark networks ranging from 9 to 13,659 buses demonstrate that Graph Attention Networks consistently achieve the lowest supervised error and constraint violation rates. Physics-informed learning reduces power flow equation residuals by up to one order of magnitude under line outage scenarios. Energy price variations constitute the most challenging perturbation type, causing slack bus reactive power errors to increase by up to five orders of magnitude relative to in-distribution settings. The open-source code, datasets, and pretrained models are publicly available to standardize safety-critical graph neural network research in power systems (https://github.com/LISTEnergyIntelligence/SafePowerGraph_Powergrid).