A new preprint on arXiv benchmarks power-flow solvers across a wide range of network sizes, from 14 to 70,000 buses. The authors compare seven open-source CPU-based solvers as well as GPU and machine-learning approaches.
The central finding is that the common practice of evaluating learned surrogates against a single CPU baseline is fragile. When the same problem set is run across multiple CPU solvers, the apparent speed advantage of ML models changes substantially.
The paper implies that researchers should adopt more comprehensive benchmarking standards before claiming speedups for learned power-flow solvers. Since only the abstract is available in the source, specific numerical results are not yet detailed.