cardef/genco-small-texas

GENCO small, fine-tuned for Texas2k heat waves and hurricanes

GENCO small (a grid foundation model, arXiv 2608.09921) fine-tuned from its official GOC 2000 OPF checkpoint to predict the AC optimal power flow of TAMU's Texas2k (a synthetic 2,000-bus grid, not ERCOT's network) under heat domes and hurricane damage. It is the model of the grid lab Space, where it runs in the browser.

For each what-if it predicts every bus voltage (magnitude and angle) and every generator's active output. A full AC power flow then makes the point physical and a linearised check certifies it (see the Space); the model's prediction alone is not a certified dispatch.

Training

• python -m gridfm_graphkit finetune --config config_small.yaml --data_path <train>, gridfm-graphkit commit 95b90ee, from GENCO's GOC 2000 small OPF checkpoint: 3.5 hours on one GPU.
• Data: train/heat_domes and train/hurricanes of cardef/grid-lab-whatifs, 23,297 what-ifs with an exact AC-OPF optimum (random heat domes, load, wind, sun, prices and outages; damage of 12 historical hurricanes), split by load scenario as in scenario_splits.json.

Results

With the Space's fast correction, on test what-ifs never seen in training:

• 1,484 heat-dome what-ifs (test/heat_domes, mostly at 0.6-0.75 of peak load): 97.0% certified AC-feasible, 0.27% above the optimal cost (median);
• 1,149 Harvey (2017) and Beryl (2024) damage draws (test/harvey, test/beryl): 92.1%, 0.73% above.

This model and the correction settings were picked after seeing these test sets; our pre-registered hurricane test did not pass. test_metrics.csv and test_RMSE.csv are the fine-tuning run's own in-distribution test (the model alone).

Files

file	what
best_model_state_dict.pt	the weights (PyTorch state dict, gridfm-graphkit GNS_heterogeneous, hidden size 24, 12 layers)
config_small.yaml	the fine-tuning config, model included
normalizer_stats.pt, normalization_stats.txt	the input normaliser's statistics, needed for inference
scenario_splits.json	the train, validation and test split of the 23,297 what-ifs
test_metrics.csv, test_RMSE.csv	the fine-tuning run's test metrics
onnx/genco_small_texas.onnx, onnx/model.json	the export the Space runs in onnxruntime-web (ONNX opset 18, float32)

Licence: Apache-2.0, as GENCO.