Cloud2Grid frozen models v1: E5c and E5d baselines

Cloud2Grid frozen models v1: E5c and E5d baselines (cloud2grid-models-v1)

What's in it: the frozen E5c and E5d GENCO checkpoints from Cloud2Grid, a held-out synthetic verification set, and a one-command check that re-runs both models.

• No new model or result: the checkpoints are the frozen, registered ones, byte-identical.
• The submission is unchanged: tag cloud2grid-submission-v1 is not modified.

Assets

Asset	Contents	Size	SHA256
cloud2grid-e5c-ft-n10000-s2.tar.gz	checkpoint (original state_dict) + its own normaliser, GraphKit config, model card, I/O schema, manifest, LICENSE (Apache-2.0), NOTICE	74.3 MB	195815e2758f8448e045f2cec23e74d448bf7a32c450e854710f43eac6ba5f6b
cloud2grid-e5d-l1-s0.tar.gz	checkpoint (original state_dict) + its own normaliser, GraphKit config, model card, I/O schema, manifest, LICENSE (Apache-2.0), NOTICE	74.3 MB	8ba6c712c21d4730a78bfd4e056ea3f9281dd424636baca9c2a96733de0845b7
cloud2grid-verification-synthetic-v1.tar.gz	E5d test pairs (3,000 states), E5c test pool (2,804 states), frozen predictions, datasheet, manifest, LICENSE (CC BY 4.0), NOTICE	47.5 MB	13b3b4f7b89cf87025f622a224728d592166d875d9c22c585f63ae2dc72ffec0
SHA256SUMS	SHA256 of the three assets

• Checkpoint SHA256s, identical to the frozen registries in the repository:
  • E5c 31b824241880ffc6807f2dad478c34c5c211ee31c7f403092e586c6373ab4d8a
  • E5d 45ab82525da00844a82a9526bc814de9290efc776159f0d8818703d5e9647435
• Where the asset hashes live: the same hashes are committed in reproduce/assets.json.

Verify

git clone https://github.com/hinox13/grid2cloud && cd grid2cloud
uv sync --extra genco                      # Linux x86-64
uv run python reproduce/verify_models.py   # CPU by default, ~2–3 min after download; --device gpu optional

The script downloads the assets, checks every hash, runs both models with their own normalisers, recomputes the metrics, cross-checks them against the frozen predictions and the published tables, and reports which claims are independently reproduced.

Reproducible from this release (synthetic, held-out data only):

• C1: E5c is blind to PV (import slope ratio 0.014).
• C2: E5d recovers the PV import response (≈ 1.0).
• C3: E5d reproduces the battery and load responses.
• C4: E5d reproduces the distributed branch-flow and voltage responses.
• C5: both models keep an absolute import error of about 14 MW.

Not reproducible from this release:

• The real 2024 results: PV → import slope AC −1.02 / E5c −0.04 / E5d −1.07 MW/MW, E5d's +13 MW import offset, and the E5D-D partial-transfer verdict. They need real operating data that are not published here.
• Training.

Numerical determinism

• GPU inference is not bit-for-bit deterministic. In our CUDA 12.6 setup, two identical GPU runs differed by up to ~0.2 MW in decoded import and branch flow; Vm and Va agreed to ~1e-5. The magnitude may differ on other GPUs.
• CPU inference is deterministic.
• The published values come from one GPU run, so the verifier compares with tolerances set from this measured noise (reproduce/README.md).
• Claim thresholds are the pre-registered E5d gates.

Limitations

See the model cards.

• E5c: blind to PV.
• Generator encoding: a known decoder inconsistency with the generator-aware encoding affects both models.
• E5d: partial transfer to real states (verdict E5D-D: +13 MW import offset; voltage and battery criteria fail), and not seed-robust.
• Both: synthetic training on one network (Sundon 33 kV), flagged selections.
• Research artefacts, not validated for operational use.

Licences

Copyright 2026 Oriol Hinojo.

• Model weights, configs and code: Apache-2.0.
• Synthetic verification dataset: CC BY 4.0.
• Third-party material keeps its own terms:
  • base model gridfm/genco-pf-transfer (Apache-2.0);
  • gridfm-graphkit and gridfm-datakit (Apache-2.0);
  • UK Power Networks open data (CC BY 4.0, attribution in each asset's NOTICE.md).