ENGAGE Dataset: Evaluating Network Generalization for AC Grid Estimation

ENAGAGE is a framework for evaluating the generalization potential of GNN-based power flow solvers, as introduced in the original paper and the accompanying git repository. Overview This dataset is the accompanying dataset for model evaluation. It was generated using SimBench and Powerdata-gen. For each grid of the 10 base distribution grids from the Simbench scenario 1 (future grid with normal increase of distributed energy resources), we generate 300 new synthetic network test cases with varying loads, generations, and topologies. Therefore, the dataset contains 3000 distribution grids for model training and evaluation. The base grids are as follows: LV-rural1 LV-rural2 LV-rural3 LV-semiurb4 LV-semiurb5 LV-urban6 MV-comm MV-rural MV-semiurb MV-urban The "complete" dataset contains a subdirectory for each base grid. In each subdirectory, there is a directory 'train/' containing all the data for the grids, including the PyTorch Geometric input graphs (dataset.pt), the pandapower network of each synthetic grid (*.json), and a csv file connecting the graph index in the dataset file with the name of the pandapower network file it came from (dataset_src.csv). The "graphs-only" dataset is the same as the complete dataset, but excluding the pandapower source networks and the src csv file, as they generally have a lot of extra information and take up a lot more storage. As most users will not need to access or modify the original grids, we recommend to download the graphs-only version, as it is much smaller. Data Attributes Each graph in the input dataset is represented as a PyTorch Geometric Data object. Each object has the following attributes and dimensions: x: Dimension: (N, 7) N is the number of nodes/buses in the network There are 7 node features: [Slack?, PV?, PQ?, p_mw, q_mvar, vm_pu, va_degree] edge_index: Dimension: (2, 2E) E is the number of lines in the network 2 because each the edge_index list assumes directed edges edge_attr: Dimension: (2E, 4) E is the number of lines in the network There are 4 edge features: [trafo?, r_pu, x_pu, sc_voltage] y: Dimension: (N, 4) N is the number of nodes/buses in the network There are 4 true node attributes: [p_mw, q_mvar, vm_pu, va_degree] dc_pf: Dimension: (N, 4) N is the number of nodes/buses in the network There are 4 DC-PF approximated node attributes: [p_mw, q_mvar, vm_pu, va_degree] Note: Some input node features of x have "nan" values, representing the unkown node features for the task. The correct values are then included in the output labels of y, with the corresponding names.