ABDHAM/p013-grid-contingency-screening-data

P-013 grid contingency screening data (IEEE 57-bus)

This is the generated data for the P-013 project: a graph neural network that ranks every single-branch outage on the IEEE 57-bus grid by how dangerous it looks, while the exact AC power flow solver checks the cases that need checking.

Code: GitHub link to be added.

What the data is

All samples are synthetic. Every one is an AC power flow solved with pandapower (Newton-Raphson) on the PGLib-OPF version of the IEEE 57-bus case, pglib_opf_case57_ieee.m (IEEE PES Power Grid Library, v23.07). The grid has 57 buses and 80 branches (lines and transformers). An operating point is a random load level and pattern with generation scaled to match. No real operating data is included.

• data/train.npz: 20,000 training samples. One operating point and one random branch outage per sample.
• data/test_sweep.npz: 250 days (operating points), each with the intact grid plus all 80 single-branch outages solved, so 250 x 81 = 20,250 rows.
• data/practice/test_sweep.npz: 500 more days in the same format (40,500 rows). No day repeats one in the two files above.
• eval/cache/*.npz: the model's predictions next to the solver's answers for every day and every outage, so the evaluation can be rerun without running the model again. Each .npz has a .json beside it saying which sweep file and which weights built it (with SHA-256 of both).
  • predictions.npz and practice_predictions.npz: built with the main model (the weights in the code repo's model/ folder).
  • score_model.npz and practice_score_model.npz: built with a second, retrained model variant from a side experiment. Its weights are not in the code repo; rebuilding these two files needs that model retrained first (see below).

Arrays in each sweep or training file: vm_pu, va_degree (bus voltage magnitude and angle), loading_percent (branch loading), in_service (which branches are in), load_p_mw, load_q_mvar, gen_p_mw, status (solver outcome, for example islanded or no solution), operating_point_id, operating_point_hash, gen_bus, seconds. Load all files with numpy.load(path); no pickles are needed.

Download

hf download ABDHAM/p013-grid-contingency-screening-data --repo-type dataset --include "data/*" --include "eval/*" --local-dir .

Run it from the root of the code repo. The files land at the same paths the code expects (data/... and eval/cache/...). The two --include flags keep this card from being written over the code repo's own README.

Regenerate it from code instead

Fixed seeds, so the data comes out the same. Each step is slow on a laptop (on a dual-core laptop, as recorded in each file's seconds array: training set 24.7 min, 250-day sweep 18.7 min, 500-day sweep 56.1 min); run them one at a time, from the code repo root, after pip install -e .:

python -m p013.generate
python -m p013.generate --only sweep --test-points 500 --seed 20260916 --out data/practice
python experiments/build_cache.py
python experiments/build_cache.py --sweep data/practice/test_sweep.npz --split practice --out eval/cache/practice_predictions.npz

The first command writes data/train.npz and data/test_sweep.npz, the second the 500 practice days, the last two the main model's caches.

For the two score_model caches, first retrain the variant model (30 epochs, slow on a CPU), export it, then build its caches:

python experiments/train_score.py --data data/train.npz --epochs 30 --seed 20260911 --out model_score
python -m p013.export_weights --weights model_score/weights.pt --out model_score/weights.npz --hidden 64 --rounds 6
python experiments/build_cache.py --weights model_score/weights.npz --out eval/cache/score_model.npz
python experiments/build_cache.py --weights model_score/weights.npz --sweep data/practice/test_sweep.npz --split practice --out eval/cache/practice_score_model.npz

Files

File	Size (bytes)	SHA-256
data/train.npz	39,591,787	26ddde343af0f71606a74d7c1e11cc64bf75383a71a59c4f34619e8c01f2c00d
data/test_sweep.npz	27,176,458	91b0ceb5b3bf7cbeb5bf806fa5314626c12bf0df2977ad1446e5fffe4f85935a
data/practice/test_sweep.npz	54,338,776	290ad0f2454194988ec82652898b4525bc831f477092ecfcca76fd6d7b70839f
eval/cache/predictions.npz	24,319,494	7e4bfd41445676ef42f8de755fcc68eadfe8c65dd10c05ad4ac4b3d0b8098ada
eval/cache/predictions.json	530	250d096fddbe8b8d9937e54f5009d2d94ddd8346a46515abaef175df51c2c567
eval/cache/practice_predictions.npz	48,630,338	0023a30297be55305417bf1bb7ee28d0d0a82081d668013e8751c8216468ab03
eval/cache/practice_predictions.json	661	50b6c51a2b2c831ba41e24920ec6fe8fe0ecc5541ac732c7c09a7c1cbbf10910
eval/cache/score_model.npz	24,270,466	215fb3a4a2092be9b3dc8f0eb3c798434f6bcfd0dfb5937b7366275674a755b9
eval/cache/score_model.json	536	04cbec2cc8f74ba961c04058116282ccd48072b342c55d6b305a22de544b1c02
eval/cache/practice_score_model.npz	48,523,232	523faa451665a021de4884ab90852a8fa8c89d04e32574ac7d2a4c0cbd8d2b13
eval/cache/practice_score_model.json	666	f82972b6fab0cf34da63f857a8536b519b52f1d476d475ac2d9067195c123d08

To check a download: sha256sum (Linux/macOS) or certutil -hashfile <file> SHA256 (Windows).

Licence

CC BY 4.0. The data is derived from pglib_opf_case57_ieee.m from the IEEE PES Power Grid Library (PGLib-OPF), which is itself released under CC BY 4.0. Please credit PGLib-OPF as the source of the grid case, and pandapower (L. Thurner et al., IEEE Transactions on Power Systems, doi:10.1109/TPWRS.2018.2829021) as the solver.