"FlowMVGrid: An ML-Ready Synthetic Medium-Voltage Grid Dataset"

"Machine-learning research in power systems is held back by a lack of publicly available datasets that simultaneously provide realistic grid topology, high-resolution temporal load variability, and a graph-native representation suitable for modern deep-learning architectures. Most existing benchmarks are limited to small IEEE test feeders with static snapshots, or provide time-series data without the underlying network structure. FlowMVGrid addresses this gap by combining real Dutch medium-voltage distribution network topologies with realistic synthetic load profiles from the GLASS dataset at 15-minute resolution, and computing full AC power-flow ground-truth labels using the power-grid-model solver. The resulting dataset covers four MV grids (57 to 878~nodes), up to 141{,}696 per-timestep samples, and provides diversity along six axes: topology variants, customer-assignment seeds, PV penetration levels, load multipliers, temporal breadth, and measurement noise. Each sample is stored as a compressed NumPy archive with node-level loads, per-unit voltages, line flows, and losses, alongside a shared graph-topology file in coordinate-list format directly compatible with PyTorch Geometric. We provide four PyTorch Dataset classes supporting power-flow approximation, state estimation, load forecasting, and connection identification, together with automated validation checks, four recommended splitting strategies, and baseline results from analytical, linear, and neural models to support reproducible benchmarking."