Case

The grid after the outage

On 870 contingencies no model had seen, verification never passed a case it could not solve.

The protocol was fixed before the run: 174 untouched operating points, five outage families, the same cases for every workflow. GENCO alone gets one or more decisions wrong in a third of the cases that have a solution.

Outcome of every case, by outage family

Each bar is 174 cases, one per operating point. The number on the right counts the cases in which GENCO alone gets at least one decision wrong.

Verified safe Verified violation Unresolved (alarm)

Checking cases in warning order

Sort the queue by GENCO's warning score and work down it. The curve shows how many of the 122 unresolved cases you have reached after checking a given share of all cases.

GENCO proposes a state. Newton-Raphson decides.

GENCO predicts every voltage and angle of a contingency in one forward pass. On its own it returns a full state for every case, including severe cases for which none of the tested Newton starts converges. The verifier turns that prediction into a starting point and lets the AC equations make the call.

The workflow

1. GENCO

One forward pass on a T4 GPU (about 70 ms here) gives the state of all 2,000 buses.

2. Warning

A logistic regression on GENCO's outputs and embeddings scores how likely the solver is to fail. It orders the queue; it decides nothing.

3. Newton-Raphson

Starts from GENCO's state; if that fails, from the pre-contingency, DC and flat states. Converged means a power mismatch below 10⁻⁸ p.u.

4. Outcome

Verified safe: an AC solution with no limit violated

Verified violation: an overload, or a bus voltage outside its limits

Unresolved: no start converged; the case goes to an engineer

What the outcomes mean, and what they do not

Unresolved is a statement about the solver and its four starts, not a proof that no AC solution exists. In the storms studied, continuation placed the fold of the AC solution below the actual load for 32 of the 33 such cases; for the remaining case, no start converged anywhere in the tested load-factor range, so no fold was identified.
The equations can have two solutions. In 2 of the 748 verified cases GENCO's start reached a high-voltage solution and the classical starts a low-voltage one. A static power flow cannot say which the grid would settle on.
The warning catches 95 % of unresolved cases at its preset threshold, with 16 % false alarms, and is close to chance within 5 % of the fold. The solver still runs on every case.
On this grid GENCO does not make screening faster: its start saves about one Newton iteration, but inference costs more than the classical solve.
This explorer shows the Texas grid; the repository also repeats the workflow test on IEEE 118. Research code, not for operational use.