Spaces:

---

gridfm

/

genco-challenge

Like 25

Running

Example submission
This illustrative entry shows the suggested structure and level of detail for a GENCO Challenge submission.

Participant: John Doe, The Safest of Grids Team
Contact: john.doe@example.com

Summary

We use GENCO to accelerate contingency analysis: screening thousands of possible grid failures to identify the scenarios most likely to require operator attention. On the GOC 500 system, GENCO is approximately twice as fast as a conventional AC power-flow solver while producing substantially more accurate overload classifications than a DC approximation.

Our workflow combines GENCO's speed with the reliability of classical solvers. GENCO first screens all candidate contingencies; a high-accuracy AC solver can then be reserved for the cases classified as potentially critical.

Motivation

Contingency analysis evaluates how a power system responds to failures such as the loss of a transmission line or generator. Operators use it to detect overloads, voltage violations, and other security risks before they threaten reliable operation.

The challenge is scale: each candidate failure normally requires a separate power-flow simulation, and a large grid may have thousands of credible contingencies. Fast, parallel inference makes GENCO a promising screening model for this workload.

Figure 1 — GENCO screens candidate contingencies so that high-accuracy solvers can focus on the most critical cases.

Method

1. Generate training data. We used gridfm-datakit to create power-flow solutions for the GOC 500 system. Random topological perturbations removed up to five branches (k = 5) from each grid state.
2. Train GENCO. We trained a GENCO Tiny model with the default gridfm-graphkit configuration.
3. Build a held-out test set. We generated approximately 15,000 additional scenarios with the same perturbation process and evaluated the trained model with gridfm-graphkit.
4. Establish a baseline. We compared GENCO against a DC power-flow approximation available through gridfm-datakit.
5. Classify overloads. We converted predicted voltage magnitudes and angles into transmission-line flows, then divided each flow by its line limit to obtain loading. A scenario is critical if any line is overloaded and safe otherwise.

Safety-aware decision threshold

A false negative—classifying an overloaded line as normal—is more consequential than a false positive. We therefore classify a line as overloaded when predicted_loading > threshold, with threshold < 1.0. The threshold is selected on held-out data by maximizing the F-beta score with beta = 3, placing greater weight on recall.

Results

Line-level loading prediction

Figure 2 — Predicted versus true line loading for the DC approximation (left) and GENCO (right). The vertical dashed line marks the true overload boundary; the horizontal dashed line marks each model's tuned classification threshold.

GENCO's predictions remain much closer to the y = x line and produce notably fewer false negatives than the DC approximation.

Scenario-level screening

Figure 3 — Maximum predicted versus maximum true line loading for each scenario: DC approximation (left) and GENCO (right).

At the scenario level, GENCO's predictions cluster tightly around the diagonal. The DC approximation produces a broad band of false positives: approximately 98.5% of its predictions are positive, while only about 11% are true positives.

Figure 4 — Scenario-level confusion matrices for the DC approximation (left) and GENCO (right), each using its independently tuned threshold.

Runtime

For a grid of this size, representative inference times are approximately:

• AC power flow: 200 ms
• GENCO: 100 ms
• DC power flow: 40 ms

GENCO is therefore about 2× faster than AC power flow. Its inference takes approximately 2.5× as long as DC power flow, but it provides substantially better classification accuracy than the DC approximation. Runtime values are based on Figure 8 of the GENCO paper.

Conclusion

GENCO offers a strong speed–accuracy trade-off for approximate contingency analysis. It can rapidly screen a large set of candidate failures, reduce unnecessary calls to expensive classical solvers, and retain the accuracy needed to prioritize genuinely critical scenarios.

Next steps

We plan to add conformal prediction so the classifier can quantify uncertainty and abstain when it cannot make a sufficiently reliable safety decision. This would make the screening stage more suitable for risk-sensitive operational workflows.