Case 2

Jointly predict bus active/reactive power and voltage magnitude plus directional branch active/reactive flows on a real provincial grid. Accuracy requires every voltage and branch-error threshold to be met for a sample.

Case II uses 366 days of operational data from an unnamed Chinese provincial grid. The first 300 days generate 500,000 training and 100,000 test samples; the remaining 66 days yield about 150,000 generalization samples with 300–690 buses.

Nodes represent buses with electrical quantities and edges represent transmission elements with network parameters. Inputs, targets, units, slack-bus treatment, and feasibility status follow the benchmark case generator.

The catalog records these protocols or variants: Standard split. Evaluation uses Accuracy (higher is better). Exact masks or folds must come from the cited release.

Only converged, feasible operating points should be compared under the documented policy. Per-unit conventions, topology changes, solver tolerances, and bus-type handling can silently alter labels; these cases do not establish transfer to real grids.