GENCO Foresight: Package power-fm-foresight

• MODEL: Extends pre-trained GENCO from AC-OPF solver to a probabilistic forecaster

• DATA: Temporal dimension from EIA ERCOT 2025, operating-stress calibration for voltage and line events

• TRAINING: Synthetic temporal dataset: 10,080 Ipopt-labeled hours driven by EIA ERCOT 2025 demand

• OPTIMIZER Anticipatory solver warm starts evaluated against cold starts (no solver benefit observed at 2.93× cost)

• SYSTEM Reproducible on Aurora (XPU) HPC and NVIDIA (CUDA) systems, with CPU fallback

The v0.2.0 tag collects all weekend experiments' code and the run index. The final consolidation gate passed the full test suite and the existing case30 regression. Failed and budget-capped experiments retain their recorded outcomes; this tag does not assert that every training gate passed.

Motivation

Method

Empirical Highlights

Conclusion

References

Below, the key references.

GridFM graphkit - A library to train, finetune and interact with a foundation model for the electric power grid. (Puech, et al., 2025).

@article{puech2025gridfmdatakitv1pythonlibraryscalable,
  title={gridfm-datakit-v1: A Python Library for Scalable and Realistic Power Flow and Optimal Power Flow Data Generation},
  author={Alban Puech and Matteo Mazzonelli and Celia Cintas and Tamara R. Govindasamy and Mangaliso Mngomezulu and Jonas Weiss and Matteo Baù and Anna Varbella and François Mirallès and Kibaek Kim and Le Xie and Hendrik F. Hamann and Etienne Vos and Thomas Brunschwiler},
  journal={arXiv preprint arXiv:2512.14658},
  year={2025},
  url={https://arxiv.org/abs/2512.14658}
}

GENCO - A Unified Neural Solver Embedded in a Development Framework for Steady-State Grid Analysis; (GEometric Neural Corrective Optimizer) (Puech et. al, 2026)

@article{puech2026gencounifiedneural,
  title={GENCO - A Unified Neural Solver Embedded in a Development Framework for Steady-State Grid Analysis},
  author={Alban Puech and Matteo Mazzonelli and Tamara R. Govindasamy and Mangaliso Mngomezulu and Héctor Maeso-García and Thomas Tolhurst and Javad Bayazi and Ali Moeini and Naomi Simumba and Celia Cintas and David Nelischer and Romeo Kienzler and Jonas Weiss and Anna Varbella and Florian Dörfler and Gabriela Hug and Martin Mevissen and Juan Bernabé-Moreno and François Mirallès and Hendrik F. Hamann and Etienne Vos and Thomas Brunschwiler},
  journal={arXiv preprint arXiv:2608.09921},
  year={2026},
  url={https://arxiv.org/abs/2608.09921}
}

IPOPT - On the implementation of an interior-point filter line-search algorithm for large-scale nonlinear programming (Wächter, Biegler, 2005)

@article{wachter2006implementation,
  title={On the implementation of an interior-point filter line-search algorithm for large-scale nonlinear programming},
  author={W{\"a}chter, Andreas and Biegler, Lorenz T},
  journal={Mathematical programming},
  volume={106},
  number={1},
  pages={25--57},
  year={2006},
  publisher={Springer}
}

PowerModels.jl - An Open-Source Framework for Exploring Power Flow Formulations (Coffrin, et al., 2017)

@inproceedings{coffrin2018powermodels,
  title={Powermodels. jl: An open-source framework for exploring power flow formulations},
  author={Coffrin, Carleton and Bent, Russell and Sundar, Kaarthik and Ng, Yeesian and Lubin, Miles},
  booktitle={2018 Power Systems Computation Conference (PSCC)},
  pages={1--8},
  year={2018},
  organization={IEEE}
}

ERCO: EIA's Hourly Electric Grid Monitor from the Electric Reliability Council of Texas, Inc. (ERCO)

@misc{eia_erco,
  author = {{U.S. Energy Information Administration}},
  title  = {Hourly Electric Grid Monitor: Electric Reliability Council of Texas ({ERCO})},
  url    = {https://www.eia.gov/electricity/gridmonitor/dashboard/electric_overview/balancing_authority/ERCO},
  note   = {Accessed: 2026-10-24}
}

V2 production policy

Determinism OFF is adopted with disclosure for new V2 production runs; use configs/v2/production.yaml. Historical v3/v4 configurations and manifests retain ON, as do evaluation and resume-mechanism checks. The case30 OFF selection-ES mean is 0.290849888 versus 0.285553075 for historical ON seeds (Δ = +0.005296813). Mean test ES is 0.307292007 versus 0.300798303 (Δ = +0.006493704); the three-seed comparison does not establish statistical equivalence, and some voltage metrics also differ. Two OFF seeds exceed the historical selected-ES range. This operational adoption does not turn the earlier failed gates into passes. See the complete Stage 1 verdict, including seed identities, stopping reasons, all test metrics and the separate confirmation-job exceedance.

Multigrid CPU format support

The V2 A0 and λ ablations use separate run identities and a 3.3-hour worker cap. Apply configs/v2/a0.yaml, lambda01.yaml, or lambda10.yaml as overrides through load_config; they are not standalone complete configurations. A0 freezes latent conditioning and runs GENCO under no_grad, then adds final-state innovations and recomputes differentiable physics. Its demand head still receives direct demand-ES gradients. The conditioned adapter remains the production model. The frozen GENCO path and the default adapter forward were checked before these experiments.

This repository supports case14, case30, case57, case118, case500, and case2000. Select a case with GENCO_CONFIG=configs/multigrid/case118_ieee.yaml; the selector uses the v4 recipe and separate output paths. GridSpec supplies dimensions, reference buses, generator mappings, and offline masks.

Run python -m scripts.multigrid_smoke for CPU loading and formatting checks, then python -m scripts.multigrid_report to refresh the support report. These checks launch no solver, training, or PBS job. Bulk artifacts stay on /lus/flare; the existing Torch/Transformers environment is retained.

Python package

The distribution is power-fm-foresight (development version 0.3.0.dev0); import it as power_fm_foresight. The Python API is also exposed by the five pff verbs below. genco, temporal, grids, data, solver, scoring, and training expose existing implementations. Private submodules retain the original module boundaries. Experiment drivers, PBS scripts, gates and controllers live in experiments/ and are excluded from the wheel. Historical module paths remain deprecated aliases. Install the package before using those paths from a checkout.

Install matrix

Provision the torch stack first. Torch, PyG, scatter and Graphkit are deliberately absent from install_requires; [cuda] and [xpu] are empty selection markers. The commands below are pinned installation recipes. The packaging gate runs on Aurora; CUDA and CPU wheel compatibility is documented, not hardware-validated here. See the PyTorch version matrix and PyG wheel matrix.

Target	Python	Torch stack	Installation policy
CUDA 12.8	3.12	Torch 2.10.0, PyG 2.8.0, scatter 2.1.2	Fresh environment, commands below
CPU	3.12	Torch 2.10.0, PyG 2.8.0, scatter 2.1.2	CPU wheels, commands below
Aurora XPU	vendor 3.12.12	Torch 2.13.0a0+gitcf30153, PyG 2.8.0.post1, scatter 2.1.2	Preserve frameworks/2026.1.0 and base constraints

CUDA, optionally after conda env create -f environment-cuda.yml and conda activate power-fm-foresight-cuda:

python -m pip install torch==2.10.0 torchvision==0.25.0 torchaudio==2.10.0 --index-url https://download.pytorch.org/whl/cu128
python -m pip install torch-geometric==2.8.0
python -m pip install --no-deps torch-scatter==2.1.2 -f https://data.pyg.org/whl/torch-2.10.0+cu128.html

CPU in a fresh Python 3.12 environment:

python -m pip install torch==2.10.0 torchvision==0.25.0 torchaudio==2.10.0 --index-url https://download.pytorch.org/whl/cpu
python -m pip install torch-geometric==2.8.0
python -m pip install --no-deps torch-scatter==2.1.2 -f https://data.pyg.org/whl/torch-2.10.0+cpu.html

For either fresh environment, install the existing GENCO runtime prerequisites and Graphkit without allowing it to replace the chosen torch stack:

python -m pip install lightning==2.6.6 opt-einsum==3.4.0 seaborn==0.13.2 plotly==7.1.0
python -m pip install --no-deps "gridfm-graphkit @ git+https://github.com/gridfm/gridfm-graphkit.git@f104900245cba26d1380037eea562cf633a97484"
python -m pip install -e .

Aurora uses the existing frameworks modules and environment. Do not install a PyPI torch wheel or upgrade vendor torch, Transformers or their dependencies. The installed PyG is a vendor build, not a request to fetch that version from PyPI. For a new Aurora environment, follow the recorded installation and preserve its base-constraints.txt/PIP_CONSTRAINT policy:

# Only if absent, use the recorded vendor PyG source/build, not a CUDA wheel.
# Scatter is built against the active vendor torch:
MAX_JOBS=2 python -m pip install --no-deps --no-build-isolation --no-binary=torch-scatter torch-scatter==2.1.2
# Graphkit needs newer isolated build tools; the runtime environment stays constrained.
env -u PIP_CONSTRAINT python -m pip install --no-deps "gridfm-graphkit @ git+https://github.com/gridfm/gridfm-graphkit.git@f104900245cba26d1380037eea562cf633a97484"
python -m pip install --no-deps --no-build-isolation -e .

The Aurora module setup is recorded in experiments/package/verify.pbs. These commands are instructions for provisioning, not permission to replace packages in a working vendor environment. Existing runtime dependencies, including OmegaConf, Datakit, matpowercaseframes and requests, are declared in pyproject.toml. Graphkit's optional CLI/evaluator extras remain separate. Ipopt uses the packaged pinned Julia project and an explicitly configured Julia executable/depot; installing the wheel does not install Julia or download assets.

Packaged recipes use PFF_ROOT as the bulk storage root and PFF_JULIA and PFF_JULIA_DEPOT for the solver. Explicit legacy configs retain their original ${base} interpolation and paths. With a wheel, the default remains the v3 recipe; choose the packaged v4 recipe explicitly:

from importlib.resources import files
from power_fm_foresight.data import load_config
from power_fm_foresight.grids import grid_config, registry

recipe = files("power_fm_foresight").joinpath("resources/v4.yaml")
config = load_config(recipe)
case = registry()["case_ids"][30]
config = grid_config(case, config)

Set PFF_ROOT, PFF_JULIA and PFF_JULIA_DEPOT before resolving a packaged recipe, or supply explicit path overrides. Package import itself does not need these variables. New manifests hash the relocated source; historical manifests are never rewritten and incompatible training resumes remain rejected.

pff CLI (0.3.0.dev0)

pff and power-fm-foresight invoke the same five argparse commands: prepare, train, evaluate, report, and verify. Run pff VERB --help for the supported flags. Configuration precedence is command flags, then --config YAML, then the recipe. Each command records its effective configuration, package/source identity, device, seeds and paths. The package has no PBS submission backend.

Weekend reproduction commands cover the fleet, deeper ensembles, A0, lambda sweep and pilots, plus the external Aurora wrapper. Reports use persisted results only. CLI verification is documented in the Phase 3 report. This development phase does not publish a wheel to TestPyPI or PyPI.