OpenGridFlex

Trustworthy AI benchmarking for active distribution grids under distribution shift.

OpenGridFlex is being built as reusable research infrastructure for GridShiftBench rather than a one-off manuscript repository. Its central scientific question is whether forecasting and uncertainty methods that look strong under IID evaluation remain accurate, calibrated and transferable under realistic distribution-grid shifts.

Paper 1: GridShiftBench

Planned shift families:

• seasonal/weather regime shift;
• DER penetration shift;
• topology/reconfiguration shift;
• sensor dropout and structured outage shift;
• cross-feeder and cross-dataset transfer.

Planned model ladder:

1. persistence / seasonal naive;
2. classical linear baseline where appropriate;
3. LightGBM point and quantile models;
4. modern temporal deep baseline;
5. graph-temporal baseline;
6. ShiftGuard-GT candidate.

Quality contract

Every manuscript table/figure must originate from a versioned config and result artifact. Raw data are immutable, split/calibration leakage is prohibited, simulator compatibility is validated explicitly, and every run records provenance (config hash, Git state, seed, packages and platform).

See:

• docs/MILESTONES.md for the gated build plan;
• docs/QUALITY_POLICY.md for non-negotiable scientific/software rules.

Core checks

PYTHONPATH=src python -m opengridflex.cli validate-config configs/paper1/mvp.yaml --repo-root .
PYTHONPATH=src python -m opengridflex.cli self-check configs/paper1/mvp.yaml --repo-root .
PYTHONPATH=src pytest -q

Current status

M3.2 — global LightGBM deterministic + quantile baseline: GREEN locally.

Implemented and gated so far:

• M0: engineering and reproducibility foundation;
• M1: SimBench adapter and electrical integrity;
• M2: leak-proof forecasting dataset and benchmark contract;
• M3.1: persistence / seasonal-naive sanity baselines;
• M3.2: global LightGBM point + quantile baseline on the sealed-validation primary task.

The full M3 baseline suite is not yet complete: the next model gate is a modern deep temporal baseline. No novel model work starts until the baseline ladder remains reproducible and CI is green.