hasethi/genco-fdia-finetuned

GENCO-PF fine-tuned as a physics oracle for false-data-injection detection (IEEE 118)

Two fine-tunes of the gridfm/genco-pf-datakit IEEE 118 Base checkpoint (case118_ieee/base/seed0), produced for a GENCO Challenge entry. Code, data and the write-up: https://github.com/hasethinvd/genco-fdia.

Each folder holds best_model_state_dict.pt, the training *.yaml and normalizer_stats.pt, in the same layout as the base checkpoint, and loads with the entry's src/genco_oracle.py (GENCO_CKPT_DIR=<folder>).

folder	training data	steps / time (one GB200)	held-out per-scan max residual, p99 (sensor-noise units)
mixed	1,500 gridfm-datakit IEEE 118 load scenarios x 4 noise draws + 600 case-file load-walk trajectories (10,854 scans)	8,313 / 30 min	294 -> 6.6
case_file_only	600 case-file load-walk trajectories (5,400 scans)	7,155 / 25 min	294 -> 13.8

Training: inputs carry Gaussian sensor noise (0.5 MW on injections); labels are AC Newton-Raphson solutions of the same noisy inputs; loss is a Huber loss on Vm, Va and branch P/Q flows, each divided by the detector's sensor noise.

Intended use: research artefact for the experiment in the linked repository (a physics-consistency check of SCADA telemetry on IEEE 118). Limitations: one grid, one noise model, one training seed per model; the case_file_only model is accurate on the case file's operating points and poor on datakit scenarios (see the write-up); detections above the exact oracle's at full attacker knowledge are surrogate error, not physical detection. Not validated for operational use.