Large-scale renewable energy integration and climate-induced extreme weather events increase uncertainty in power system operations, calling for tools that can accurately and efficiently simulate a wide range of scenarios and operating conditions. Machine learning approaches offer a potential speedup over traditional solvers, but their performance has not been systematically assessed on benchmarks that capture real-world variability. This paper introduces PF∆, a benchmark dataset designed to evaluate power flow methods under variations in load, generation, and topology. We evaluate traditional and graph neural network-based approaches, and demonstrate key areas for improvement in existing methods.