When the QP solver says infeasible: auditing 347 optimal power flow repair steps against a second solver

Every QP of an AC OPF repair on 34 PGLib-OPF cases re-solved with Clarabel: PIPS was right when it answered, but failed on 26 solvable QPs.

The grid-linopf research programme asks whether a market's linear (DC) dispatch can be repaired to satisfy the real AC power-flow limits cheaply. Its current method takes the plain DC dispatch, linearises the AC power flow, solves a quadratic programme (QP) for the next dispatch and voltage setpoints, and repeats for up to ten rounds. When the QP has no point inside every limit, the method falls back to an elastic QP, where each limit can be broken at a price, and then halves its step box.

The morning's milestone had found that the QP, solved by PYPOWER's PIPS interior-point solver, takes 98 to 99 % of the run time. The obvious next step was to make the QP smaller by keeping only the limits near their bounds, adding any others the solution breaks and solving again. For a convex QP this constraint-generation loop gives exactly the full QP's answer.

The screened QP gave a different answer, and that was the finding. On the 73-bus angle-limited case, PIPS had reported the full round-1 QP as failed. The repair then ran on the elastic fallback every round and never got feasible. The screened version solved the same QP to a point that breaks no limit. Since the screened QP is a relaxation, the full one cannot have been infeasible. So we wrote the predictions down and audited every QP the repair poses, 347 across the 34 benchmark cases, against Clarabel, an independent interior-point solver (Goulart and Chen, 2024).

What came back (research/infra-dcopf/experiments/audit_qp.txt):

  • When PIPS reports a solution, it is the right one. Its objective matches Clarabel's within one part in a million on all 236 QPs both solve. The costs of every repair that finished are unchanged.
  • Its failures on hard QPs are mostly real. Clarabel solves only 5 of the 75 hard QPs PIPS gave up on. Usually the linear model really has no point inside every limit. (We had predicted at least half would be solvable; refuted.)
  • Its failures on elastic QPs mostly are not. An elastic QP has a slack on every limit, so it is almost always feasible. PIPS failed 36 of 99, and Clarabel solved 26 of those. The other 10 all come from one case (the 300-bus congested one), where every solver failed.

Rerun with Clarabel on every QP, the repair is fully AC-feasible on 21 of 34 cases (19 with PIPS), and on 31 of 34 with the voltage-only stage after it (29). No case is lost, and the summed run time falls from 1,353 s to 260 to 277 s. That is 4.9x in one run and 5.2x in a rerun, so our 5x prediction holds only within timing noise. The screened PIPS solve also reaches 21. It is 3.0x faster but inherits PIPS's elastic failures.

This is a solver result, not a better method, and it is one benchmark library at one load point per case. Still, it moves two of the programme's misses from "the method fails" to "the solver failed". It also gives a cheap check for anyone running successive linearisation on top of an interior-point QP: re-solve the steps that failed with a second solver before calling the linear model infeasible. The next steps use Clarabel: the LPAC and optimised-linearisation baselines, and the held-out 500 to 2,000-bus cases.

Code, data and the pre-registered plan are in research/infra-dcopf (LOG.md, 2026-11-13). Produced with AI assistance (Claude).

All writing · More in Research findings · More on grid-linopf · All tools and research