oriolh/cloud2grid-e5d-l1-s0

Cloud2Grid E5d: physics-response-aware adaptation of GENCO (e5d_l1_s0)

GENCO (gridfm/genco-pf-transfer) fine-tuned like E5c, plus a paired-response loss: pairs of AC-solved synthetic states that differ in one input (solar, battery or one load) teach the model the AC change in grid import, branch flows and voltages, normalised per MW of perturbation. Total loss L_E5c + λ·κ·L_response, λ = 1, seed 0. On held-out synthetic pairs it recovers the AC solar response (import slope ratio 0.997, against 0.014 for E5c and 0.001 for the λ = 0 control); on 2,829 reconstructed 2024 states its fitted solar slope is −1.07 MW per MW against −1.02 for the AC reference, but its import carries a near-constant +13.1 MW offset, and only this seed of three learned the response this well. A research model, not an operational one.

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
best_model_state_dict.pt	the frozen checkpoint (original torch.save state_dict)
normalizer_stats.pt	the exact normaliser fitted with it; use only with this checkpoint
graphkit_predict_config.yaml	GraphKit configuration used for prediction
IO_SCHEMA.md	model architecture, inputs and outputs
checkpoint_manifest.json	file hashes, frozen registry entry, base-model and dependency provenance
response_sigma.json	training-time response-loss scale (transparency only; not needed for inference)
MODEL_CARD.md	the released model card (also reproduced below)
release/cloud2grid-e5d-l1-s0.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

import torch
from huggingface_hub import hf_hub_download

repo = "oriolh/cloud2grid-e5d-l1-s0"
state = torch.load(hf_hub_download(repo, "best_model_state_dict.pt"), map_location="cpu", weights_only=True)
norm = torch.load(hf_hub_download(repo, "normalizer_stats.pt"), map_location="cpu", weights_only=True)
# build GraphKit's GNS_heterogeneous model with graphkit_predict_config.yaml and load `state` (see IO_SCHEMA.md)

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

Cloud2Grid's fine-tuned weights, configuration and cards: Apache-2.0 (LICENSE). Fine-tuned from gridfm/genco-pf-transfer (Apache-2.0); all its weights were modified, the architecture is unchanged. Training data: synthetic states from gridfm-datakit (Apache-2.0) on a network derived from UK Power Networks open data (CC BY 4.0); attribution in NOTICE.md.

---

The released model card (verbatim)

Model card: Cloud2Grid E5d (e5d_l1_s0)

What it is. GENCO (GraphKit GNS_heterogeneous) fine-tuned like E5c, plus a paired-response loss. Pairs of AC-solved states that differ in one input (PV, battery or load) teach the model the AC change in import, branch flows and voltages. λ = 1, seed 0. It is the frozen, selected E5d checkpoint: the final model in Cloud2Grid v1.1, shown alongside zero-shot, E5, E5b and E5c in its Explorer.

Checkpoint sha256	45ab82525da00844a82a9526bc814de9290efc776159f0d8818703d5e9647435
Normaliser sha256	9ae23c469edd4fe5cf894e2e02264d801013015fe5dcd170c7d35d18f69f5cec. Use only with this checkpoint (baseMVA 36.741)
Response σ	response_sigma.json (sha256 6c87e81b…59fd): the training-time response-loss scale, provided for transparency and not needed for inference
Selection	Pre-registered rule, synthetic validation only. FLAGGED: no run passed all 23 criteria, so the fallback rule chose it. It passes all 15 response and 3 retention criteria
Training	200 epochs, best epoch 198. Pretrained init from gridfm/genco-pf-transfer (Apache-2.0). Same data, splits, normaliser rule and recipe as E5c; loss = L_E5c + λ·κ·L_response
Frozen record	metadata/e5d_checkpoint_registry.yaml in hinox13/grid2cloud, tag cloud2grid-submission-v1
Formats	Original torch.save state_dict, unchanged (loads with torch.load(weights_only=True))
Schema, licences	IO_SCHEMA.md, LICENSE (Apache-2.0), NOTICE.md

Verified results (synthetic; reproducible with reproduce/verify_models.py)

E5c held-out test (2,804 states):

Metric	Value
PBE	0.795 MVA
Active residual	0.672 MW
Vm error	0.0125 pu
Import MAE	14.2 MW
Import sign accuracy (|AC| ≥ 5 MW)	95.0 %

Held-out E5d test pairs (500 per type), GENCO response vs AC:

	Import slope ratio	Import rel. error	Branch P	Vm	Va
PV	0.997	0.026	0.057	0.049	0.026
Battery	0.998	0.010	0.18	0.26	0.28
Load	0.962	0.049	0.23	0.31	0.27

E5d recovers the PV response. The import ratio goes from 0.014 for E5c to 0.997, including small steps of 0.05–0.375 MW. The λ = 0 control trained with the same code stays blind (0.001), so the gain comes from the response loss.

Real-state results (not reproducible from this package)

On 2,829 real Sundon states from 2024:

• PV response: import responds to satellite-derived PV with slope −1.07 MW/MW, against −1.02 in AC (E5c −0.04).
• Load response: slope ratio 0.99.
• Import: r 0.994.
• Verdict E5D-D (partial / other):
  • Import offset: a near-constant +13.1 MW offset (MAE 13.2 MW), R3 fails.
  • Voltage: Vm MAE 0.0109 pu against a 0.010 limit, R4 fails.
  • Battery: slope overshoots (ratio 1.17), R6 fails narrowly.

These numbers depend on source data not included here (see the release notes). The 2024 population was examined during development, so it is not an independent confirmation.

Licence and provenance

• This package (Cloud2Grid's fine-tuned weights, configuration, model card and schema): Apache License 2.0, see LICENSE. Copyright Oriol Hinojo.
• Base model: fine-tuned from gridfm/genco-pf-transfer pretraining/best_model_state_dict.pt (sha256 b2c9cf9b…5d41, Apache-2.0; Puech et al. 2026, arXiv:2608.09921). All of its weights were modified by fine-tuning; the architecture is unchanged.
• Training data: synthetic states generated with gridfm-datakit (Apache-2.0) on a Sundon network model derived from UK Power Networks open data (CC BY 4.0). The attribution is in NOTICE.md.
• Exact provenance: checkpoint_manifest.json holds the file hashes, the frozen registry entry, the public source tag and the dependency versions.

Reproducing the numbers

• What can be reproduced: reproduce/verify_models.py in hinox13/grid2cloud re-runs this checkpoint on the held-out synthetic verification set (release asset cloud2grid-verification-synthetic-v1.tar.gz). It recomputes every synthetic number in the "Verified results" section. Independent numerical reproduction is synthetic-only: the real-state numbers need data not in this release.
• Numerical determinism: GraphKit/PyG inference on GPU is not bit-for-bit deterministic (most likely the varying order of GPU scatter-add reductions; not isolated, and measured on one CUDA 12.6 GPU setup).
  • How large the noise is: two identical GPU runs differ by up to ~0.2 MW in decoded import and branch flows (mean ~0.03 MW); Vm and Va agree to ~1e-5. CPU inference is deterministic.
  • Why: the near-ideal branches of the network amplify float32 angle rounding into flows.
  • Effect on the published numbers: they come from one GPU run, so small response errors carry this noise.
  • Tolerances: the verifier's tolerances are set from it (reproduce/README.md). Claim thresholds are the pre-registered gates.

Limitations

• Partial transfer (E5D-D): correct sensitivities, but the absolute import level carries a ~13 MW offset on real states. Synthetic import MAE is 14.2 MW; the ST3 / ST4 import gates fail.
• Not seed-robust: of three λ = 1 seeds, only this one learned the distributed (branch, voltage) response and the load response to gate level. The other seeds are much worse.
• Generator-encoding issue (inherited from E5c): GraphKit's internal physics decoder sets Pg = 0 at PQ buses, so the known LBUZ generation (buses 8–10) is missing from the model's internal residual feedback. It was found after freezing. Evaluation uses the correct generation, so the reported metrics are valid for the model as implemented; the effect of correcting it is not evaluated in this release.
• One network, synthetic training only; no real data was used for training or selection.
• Research artefact, not validated for operational use.