Add flat start option for Newton-Raphson power flow by m-mirz · Pull Request #1607 · PowerGridModel/power-grid-model

m-mirz · 2026-09-26T17:41:07Z

The linear voltage guess that Newton-Raphson starts from can make it diverge on meshed transmission grids that a flat start solves. The new power_flow_initialization option selects the start: linear (default, unchanged) or flat. See the new Initialization section in docs/algorithms/pf-algorithms.md for details.

C API: PGM_PowerFlowInitialization and PGM_set_power_flow_initialization, plus the C++ wrapper setter.
Python: PowerFlowInitialization enum; calculate_power_flow accepts power_flow_initialization as the enum or its name.
Tests: C++ solver tests for the flat start, including a voltage regulator case; Python test for each option value.

Partially fixes #1595 and should be merged and rebased after #1610

Relates to #549

mgovers

Very clear and good suggestion!

A couple small suggestions (mostly repetitive).

Given the impact on the C API, let's also have someone else have a look at it.

TonyXiang8787 · 2026-09-28T08:15:44Z

Hi @m-mirz,

Nice that you have discovered this problem. I would like to first start with a validation case.

It would be nice if you can re-produce a minimal validation meshed case where:

network is meshed
diverges in current PGM NRPF (due to the linear power flow as starting condition)
can converge in pandapower or matpower
mark the validation case as xfail, see some xfail examples in the existing validaiton case.

Please first create a separate PR for this, we can then first merge the validation case.

Jerry-Jinfeng-Guo · 2026-09-28T14:46:57Z

Hi @m-mirz , next to what @TonyXiang8787 asked in his comment, we are also curious if the validation case could show that the linear method produces results that are indeed far away from the real solution to the extent that NR iterations diverge. We ask because by varying the starting point between said point and the real solution, we should be able to see a behavioral change if everything else is workign as they should.

m-mirz · 2026-09-28T19:17:28Z

@TonyXiang8787 I have added #1610 to show where NR does not converge from a linear start.
Of course, increasing sk also improves the results in terms of reproducing matpower since the source behaves closer to an ideal source. If sk is sufficiently large, it even fixes convergence for the matpower case118 but for none of the larger cases. Starting from case300+, only a flat start let's NR converge.

When #1610 is merged, I would rebase this one here.

TonyXiang8787 · 2026-09-29T11:33:09Z

@TonyXiang8787 I have added #1610 to show where NR does not converge from a linear start. Of course, increasing sk also improves the results in terms of reproducing matpower since the source behaves closer to an ideal source. If sk is sufficiently large, it even fixes convergence for the matpower case118 but for none of the larger cases. Starting from case300+, only a flat start let's NR converge.

When #1610 is merged, I would rebase this one here.

Hi @m-mirz, the other one is already merged and I have done a down merge for this PR.

In this PR you can remove the xfail and the test should pass.

m-mirz · 2026-09-29T17:47:30Z

@TonyXiang8787 as mentioned in #1610, it looks like the merge was not triggered yet since my GPG signature was missing on the commit. I fixed it and now I am waiting for #1610 to get merged.

The linear voltage guess that Newton-Raphson starts from can make it diverge on meshed transmission grids that a flat start solves. The new power_flow_initialization option selects the start: linear (default, unchanged) or flat. See the new Initialization section in docs/algorithms/pf-algorithms.md for details. - C API: PGM_PowerFlowInitialization and PGM_set_power_flow_initialization, plus the C++ wrapper setter. - Python: PowerFlowInitialization enum; calculate_power_flow accepts power_flow_initialization as the enum or its name. - Tests: C++ solver tests for the flat start, including a voltage regulator case; Python test for each option value. Signed-off-by: Markus Mirz <16180422+m-mirz@users.noreply.github.com>

The MATPOWER IEEE 300-bus case converges from a flat start but diverges from the linear start, also with a stiff source. The case is added twice, with power_flow_initialization flat and linear; the linear one is marked xfail with IterationDiverge. The expected output is the MATPOWER power flow solution; see the case README for source and conversion. - Python and C++ validation tests read power_flow_initialization from params.json. - IterationDiverge can be named in xfail/raises of Python validation cases. Signed-off-by: Markus Mirz <16180422+m-mirz@users.noreply.github.com>

m-mirz · 2026-09-29T20:49:31Z

Now that #1610 is merged, I have rebased and GPG signed the commits. I have also removed the xfail test so that only the passing test remains.

The flat-start case now covers the MATPOWER IEEE 300-bus system, so the xfail linear-start case is removed and the flat-start case moves up to the case directory. The README keeps the note that Newton-Raphson diverges from the linear start on this case. Signed-off-by: Markus Mirz <16180422+m-mirz@users.noreply.github.com>

TonyXiang8787 · 2026-09-30T06:53:30Z

Now that #1610 is merged, I have rebased and GPG signed the commits. I have also removed the xfail test so that only the passing test remains.

Hi, please address the comments (especially the naming decision) from @mgovers. Then ping @mgovers and myself, we can then start CI.

Generalize the Newton-Raphson start option so it can be reused by other calculation types and methods: - PowerFlowInitialization -> CalculationInitialization (C++, C API, Python) - power_flow_initialization -> calculation_initialization (options, PGM_set_calculation_initialization, Python keyword argument) - add a `default` value (0) that selects the default initialization of the calculation method (linear for Newton-Raphson); flat = 1, linear = 2 - `default` is a C++ keyword, so the C++ enumerator is named default_initialization, following CalculationMethod::default_method Signed-off-by: Markus Mirz <16180422+m-mirz@users.noreply.github.com>

m-mirz · 2026-09-30T18:28:07Z

@TonyXiang8787 and @mgovers I have accepted your changes and also applied similar changes in a few more places to make sure it is consistent.

Use the source reference angle in the flat start. Signed-off-by: Markus Mirz <16180422+m-mirz@users.noreply.github.com>

TonyXiang8787 · 2026-10-01T18:14:05Z

Hi @m-mirz,

Thanks for the update. Please fix the build errors, and then I think the PR is in good shape.

Signed-off-by: Markus Mirz <16180422+m-mirz@users.noreply.github.com>

m-mirz · 2026-10-01T18:44:19Z

@TonyXiang8787 I think that I've just fixed the CI issues but CI checks are awaiting approval.

mgovers reviewed Sep 28, 2026

View reviewed changes

mgovers added the feature New feature or request label Sep 28, 2026

mgovers assigned m-mirz Sep 28, 2026

Jerry-Jinfeng-Guo mentioned this pull request Sep 28, 2026

[FEATURE] Custom starting point for iterative calculations #1609

Open

1 task

m-mirz added 2 commits September 29, 2026 22:45

m-mirz force-pushed the feature/newton-raphson-flat-start branch from 8c150ff to 5ae220b Compare September 29, 2026 20:45

m-mirz force-pushed the feature/newton-raphson-flat-start branch from 3b2d3c1 to 43ceea9 Compare September 30, 2026 18:25

TonyXiang8787 reviewed Sep 30, 2026

View reviewed changes

Comment thread ...d_model_c/power_grid_model/include/power_grid_model/math_solver/newton_raphson_pf_solver.hpp Outdated

TonyXiang8787 reviewed Sep 30, 2026

View reviewed changes

Comment thread ...d_model_c/power_grid_model/include/power_grid_model/math_solver/newton_raphson_pf_solver.hpp Outdated

m-mirz and others added 2 commits September 30, 2026 22:00

Add average source start for Newton-Raphson power flow

0287ed1

Use the source reference angle in the flat start. Signed-off-by: Markus Mirz <16180422+m-mirz@users.noreply.github.com>

Merge branch 'main' into feature/newton-raphson-flat-start

94eabd8

Fix clang-tidy include and ruff line length issues

b01f5cc

Signed-off-by: Markus Mirz <16180422+m-mirz@users.noreply.github.com>

m-mirz force-pushed the feature/newton-raphson-flat-start branch from 88e39a1 to b01f5cc Compare October 1, 2026 18:41

Conversation

m-mirz commented Sep 26, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

mgovers left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

TonyXiang8787 commented Sep 28, 2026

Uh oh!

Jerry-Jinfeng-Guo commented Sep 28, 2026

Uh oh!

m-mirz commented Sep 28, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

TonyXiang8787 commented Sep 29, 2026

Uh oh!

m-mirz commented Sep 29, 2026

Uh oh!

m-mirz commented Sep 29, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

TonyXiang8787 commented Sep 30, 2026

Uh oh!

m-mirz commented Sep 30, 2026

Uh oh!

Uh oh!

Uh oh!

TonyXiang8787 commented Oct 1, 2026

Uh oh!

m-mirz commented Oct 1, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

This branch has not been deployed

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

m-mirz commented Sep 26, 2026 •

edited

Loading

mgovers left a comment •

edited

Loading

m-mirz commented Sep 28, 2026 •

edited

Loading

m-mirz commented Sep 29, 2026 •

edited

Loading

m-mirz commented Oct 1, 2026 •

edited

Loading