Formulate branch-level Jabr inequality in SOCWR

AI4OPT / OPFGenerator

Instance generator for OPF problems

MIT License

2 stars 3 forks source link

Formulate branch-level Jabr inequality in SOCWR #66

Closed mtanneau closed 6 months ago

mtanneau commented 6 months ago

closes #60

This PR makes the following changes:

It formulates Jabr's inequality at the branch-level, as opposed to bus-pair before. This means that wr, wi variables are also created for each branch
It fixes a bug in the JSON -> H5 converter that prevented dual variables associated to wr/wi upper bounds were not exported properly
It adds a generic dual feasibility test for the dual constraints associated to w, wr, wi variables.

TODO:

[x] Fix unit tests that error because of discrepancy between us and PowerModels (which uses a bus-pair formulation)

mtanneau commented 6 months ago

@klamike @ivanightingale our unit tests currently check the objective value of our formulation against that found by PowerModels. Since this PR breaks the equivalence with PowerModels, that check is no longer valid.

https://github.com/AI4OPT/OPFGenerator/blob/e66221e661e8dfe7c2e06eeb5e2275d907b1bf87/test/opf/opf.jl#L36

I can do two things to fix the broken tests:

relax the relative tolerance to 1e-3 in the test. This seems to work for our setting, but may not be valid in general
remove this test altogether (which would require me to update things so that we deactivate it only for SOCWR problems)

klamike commented 6 months ago

I think the latter is worth doing since other custom formulations will probably also need custom tests.

ivanightingale commented 6 months ago

I agree with number 2. I think the test for feasibility in this PR is sufficient.

mtanneau commented 6 months ago

OK, I'm putting the update to unit tests in #67