Alirezashooshtari/genco-texas-clustered

GENCO fine-tuned for clustered contingencies, with an unresolved-case warning

The released GENCO checkpoint (gridfm/genco-pf-contingency, model/training_bs_16, revision 5c625b0) fine-tuned with its own training objective (gridfm-graphkit MixedLoss) on 6,737 contingencies of the Texas 2,000-bus grid labelled by a verified AC solver: random, nearest-line and corridor outages from N-5 to N-20, at the actual load and with loads scaled by 0.85 to 1.35. One outage construction was held out for testing. These are the weights of training seed 0. A second seed was trained on the same cases (its weights are in seed1/); an earlier run, made before a fix to gradient accumulation, is kept for the record.

Code, notebooks and results: github.com/Alirezashooshtari/genco-trust-but-verify · Demo: genco-contingency-verifier

Alireza Shooshtari, Institut de Recerca en Energia de Catalunya (IREC)

Files

File	Content
model.safetensors	fine-tuned weights for gridfm-graphkit's GNS_heterogeneous (caseTexas configuration)
normalizer_stats.pt	the normalizer of the base model, unchanged
config.json	base model, code revision, training settings and file hashes
warning/warning_released.joblib, warning/warning_finetuned.joblib	the warning head (scikit-learn StandardScaler + LogisticRegression) for each GENCO model
warning/warning_released.json, warning/warning_finetuned.json	the same heads as plain numbers: scaler, coefficients, intercept, threshold
warning/features.json	the 777 input features in order (9 GENCO outputs, 384 + 384 embedding mean and max)
warning/scores.csv	the score and label of every evaluated case
data/features_released.npz, data/features_finetuned.npz	features and labels of every training, validation and test case
training/finetune.ipynb	the notebook that produced the weights (TRAIN_SEED = 0)
loadgrowth/model.safetensors	the same fine-tuning plus a load-growth response term (see below)
training/finetune_loadgrowth.ipynb	the notebook that produced those weights (TRAIN_SEED = 0)
seed1/model.safetensors, loadgrowth/seed1/model.safetensors	the second training seed (TRAIN_SEED = 1) of each, with the results reported below

The .joblib heads were saved with scikit-learn 1.6.1 (the Colab version at the time); with other versions use the JSON files, which hold the same parameters as plain numbers.

model.safetensors SHA-256: a20b05e4dd4f1739fd8170596fc71b89eed305bc4ca9383f59dba5e31da75352. It is converted from the notebook's genco_finetuned_state_dict.pt (SHA-256 cf705d23b87cc0fd1b6ede6a9028925584bd9b00b936a76244a90a75fcb353a2).

Loading follows the base model: build the gridfm-graphkit task from the caseTexas configuration and its normalizer, then load these weights in place of the base checkpoint. In outline:

from safetensors.torch import load_file
# args, normalizer: the caseTexas configuration and normalizer of the base model (gridfm-graphkit, commit 95b90ee)
task = get_task(args, [normalizer])
task.model.load_state_dict(load_file("model.safetensors"))

notebooks/evaluate_finetuned.ipynb in the code repository does this end to end, from installation to predictions.

Results

Paired against the base model on cases neither model saw, with the protocol fixed before the run:

	Base model	Fine-tuned (seed 0)	Seed 1 / first run, before the fix
Decisions missed per case, clustered outages (401 cases)	6.13	4.79 (95 % CI of the difference −1.72 to −0.97)	4.88 / 4.70
Decisions missed per case, held-out construction (153 cases)	2.01	1.56 (−0.66 to −0.24)	1.59 / 1.58
Decisions missed per case, random outages (347 cases)	0.14	0.14 (−0.03 to +0.04)	0.15 / 0.18
Median worst-bus error, released N-2 scenarios (p.u.)	0.0013	0.0036	0.0043 / 0.0035

The fine-tuned model misses fewer decisions on clustered and held-out outages and has a lower worst-bus voltage error on clustered outages (median 0.017 to 0.012 p.u.), but it is less accurate on ordinary scenarios: a specialist, not a replacement for the base checkpoint. A rule that needs only the outage itself picks the right model: use this one when at least 5 lines are out and the buses they touch are at most 11 hops apart, the base model otherwise. On the frozen test it sends 400 of the 401 clustered outages here and none of the 347 random ones, and on random N-2, N-10 and N-20 outages it picked this model in none of 2,000 draws each.

The warning predicts whether the classical Newton-Raphson policy (pre-contingency, DC and flat starts; tolerance 1e-8 p.u.; 30 iterations) will leave a case unresolved. AUROC 0.96 with the base model and 0.98 with the fine-tuned one on the 870-case frozen test; at its threshold it catches 95 to 98 % of unresolved cases with 10 to 16 % false alarms. It predicts solver non-convergence, not the absence of an AC solution, and it is close to chance within 5 % of the fold of the tracked AC solution.

With a load-growth response term (loadgrowth/)

The base model's voltage drops under system-wide load growth are only 0.60 of the AC ones, which is why it glides past the fold of the AC solution. loadgrowth/model.safetensors is fine-tuned on the same cases with the same objective, plus one term: every training case is paired with the same case at 2 % more load (loads only), and the difference between GENCO's and the AC voltage change at the PQ buses is penalised. The decision rule was fixed before the run; both seeds meet it.

	Base model	Fine-tuned (model.safetensors)	Load-growth, seed 0 (loadgrowth/)	Load-growth, seed 1
Load-growth response on 129 held-out scenarios (AC = 1)	0.60	0.78	0.99	1.01
Decisions missed per case, clustered outages	6.13	4.79	4.45 (95 % CI of the difference −2.23 to −1.23)	4.22
Decisions missed per case, held-out construction	2.01	1.56	1.35	1.24
Decisions missed per case, random outages	0.14	0.14	0.21	0.14
Median worst-bus error, released N-2 scenarios (p.u.)	0.0013	0.0036	0.0052	0.0056

It is the stronger specialist and responds to load growth like the AC solution, but it is further from the base model on ordinary scenarios, so the same routing rule applies: with it, missed decisions per clustered case fall to 4.45 and those on random outages stay at 0.14. Near the fold it still follows only part of the AC collapse (14 % of the AC voltage drop over the last 5 % of load, against 5 % for the base model), and Newton never converges beyond a fold from its state.

loadgrowth/model.safetensors SHA-256: 584b15d658fa7dd97b606c5d78ade0f40ae6bbbf6d8f5fe631fbe072c40763eb, converted from the notebook's genco_finetuned_state_dict.pt (SHA-256 9d3f671bf4c171b9327b52e9f100b01c140f76f096fdc35e1318aa3b5401c8b6). It loads in the same way as model.safetensors.

The second seeds load in the same way: seed1/model.safetensors SHA-256 20143b739aef9dca29733a7f92ec6362dd048450adfe53d0d4793463d33552c3 (from 544185154b000f39bbfa53cf7ed1ac6e92fdebf47092f4b864973e418c38d14b), loadgrowth/seed1/model.safetensors SHA-256 cc855f15be3f90257d9850f621ca8b81867a7ed5cc960e414de9d92505b81a09 (from 7bf1ad7b445f7431d7d69e9c7e573cd256e3fad7bd3ce8ebb823761d6130e37c).

Intended use

Research. Every security decision must come from an AC solver: the model proposes a state and the warning orders the work.