oriolh/cloud2grid-verification-synthetic-v1

Cloud2Grid synthetic verification data v1

Held-out synthetic AC power-flow states of the reconstructed Sundon 33 kV network, used by reproduce/verify_models.py to re-run the E5c and E5d checkpoints with real GraphKit inference: 2,804 held-out states and 1,500 AC-solved response pairs (solar, battery and load). Synthetic only: no measured time series and no satellite data.

Part of Cloud2Grid (website · code · GitHub Release · reproduction), a study of adapting the GENCO grid foundation model (gridfm/genco-pf-transfer; Puech et al., 2026) to a real UK distribution network driven by satellite-derived solar.

Files

File	What it is
cloud2grid-verification-synthetic-v1/	the held-out states and response pairs (Parquet)
MANIFEST.json	file hashes and provenance
DATASHEET.md	the released datasheet (also reproduced below)
release/cloud2grid-verification-synthetic-v1.tar.gz	the GitHub Release archive, byte-identical (sha256 in SHA256SUMS)
SHA256SUMS	checksums of every file in this repository

These are the files of the GitHub Release cloud2grid-models-v1, unchanged; the release archive's own SHA256 is the published one.

Load

from huggingface_hub import snapshot_download

path = snapshot_download("oriolh/cloud2grid-verification-synthetic-v1", repo_type="dataset")

Verify

Re-run the synthetic claims with real inference from these Hugging Face copies (the verifier checks every archive's SHA256 first):

git clone https://github.com/hinox13/grid2cloud && cd grid2cloud
uv sync --extra genco
uv run hf download oriolh/cloud2grid-e5c-ft-n10000-s2 --revision v1 --include "release/*" --local-dir hf_assets
uv run hf download oriolh/cloud2grid-e5d-l1-s0 --revision v1 --include "release/*" --local-dir hf_assets
uv run hf download oriolh/cloud2grid-verification-synthetic-v1 --type dataset --revision v1 --include "release/*" --local-dir hf_assets
uv run python reproduce/verify_models.py --assets-dir hf_assets/release   # CPU, about 2-3 minutes

No login is needed. Expected: All checks pass: True, claims C1-C5 reproduced, 44/44 scored metrics within tolerance (run anonymously from a fresh clone on 23 Oct 2026).

The real-state (2024) results need operating data that are not distributed, so they are documented but not independently reproducible from this repository.

Licence and attribution

This dataset: CC BY 4.0 (LICENSE). Synthetic states generated with gridfm-datakit (Apache-2.0) on a network model derived from UK Power Networks open data (CC BY 4.0). When reusing, credit Cloud2Grid and UK Power Networks as set out in NOTICE.md.

---

The released datasheet (verbatim)

Datasheet: Cloud2Grid synthetic verification data v1

Purpose. To let anyone re-run the frozen E5c and E5d checkpoints on the held-out synthetic data used for their published results, and recompute the response and state metrics.

Contents (all synthetic, AC-solved, Sundon 33 kV, generator-aware GraphKit tables):

Folder	What	Size
e5d_test_pairs/	The E5d test pair set: 1,500 pairs (500 each PV, battery, load), 3,000 states. pairs.parquet maps each pair (probe_id, probe_type, magnitude_class, x_mw) to its states A and B	held out from training, validation and selection
e5c_test/	The E5c held-out test pool: 2,804 states (scenarios 0–2803 of the E5c pool, unchanged ids)	held out from training and selection
frozen_predictions/<run>/<set>/	The frozen predictions of each model on each set, trimmed to the columns the metrics use, plus the weights hash they were made with	secondary cross-check
MANIFEST.json	sha256 of every file

How the states were made.

• States: gridfm-datakit 1.1.0 on the E1 Sundon planning case, with N-1 outages and admittance perturbations for the E5c pool.
• Operating envelope: loads, battery and generation within bounds from UKPN planning data (LTDS, ECR).
• Pairs: for the E5d pairs, each pair differs in exactly one controllable input.
• Solver and filtering: AC solutions by the datakit/pandapower pipeline; states failing the pre-registered AC sanity gates were dropped.

Not included:

• No real measurements and no satellite data.
• Nothing used to train or select either model.

Licence. CC BY 4.0 (LICENSE), copyright 2026 Oriol Hinojo.

• Derived from UK Power Networks open data (CC BY 4.0) using Apache-2.0 tools.
• Credit when reusing: credit Cloud2Grid and UK Power Networks as set out in NOTICE.md, link the licence, and indicate changes.
• The frozen predictions are outputs of the Apache-2.0 Cloud2Grid models and are included under the same CC BY 4.0 terms.