GitHub

Calculation parity record

The pre-refactor source was frozen at commit 28553eb6e2cb36df5f86c6d5194dbd4232ea1812. The numerical comparison uses the IEEE 39-bus AC/DC fixture, including generators, loads, transformers, transmission branches, a DC source, and the STATCOM model. The power-flow options are identical in both calculations: a 267.94 kV line-to-line voltage base and bounded bus-voltage variables.

Power flow and operating point

The direct NetworkState parameterization and the preserved Classic network calculation both terminated with LOCALLY_SOLVED. Their integer bus numbers differ because the two representations construct buses in different orders. Values were therefore compared through element-to-component mappings and through the complete set of physical node names assigned to each AC or DC bus.

The frozen reference contains every numeric field returned in the bus, gen, and branch solution groups, together with every numeric field of the calculated STATCOM Setpoint, for:

  • 49 AC buses and 2 DC buses;
  • 10 AC generators and 1 DC generator;
  • 55 AC branches and 1 DC branch;
  • the AC/DC converter;
  • the calculated STATCOM Setpoint.

Passive shunts occur in the PowerModelsACDC input data but have no entries in the solved solution dictionary. Their input counts and mappings are checked separately. The largest absolute difference among solved component fields was 1.990074771640593e-14, in the reactive flow of Zg31. The largest absolute difference among bus fields was 1.1102230246251565e-15, in the voltage magnitude of the bus containing Bus7d and Bus7q. All six STATCOM setpoint fields agree within an absolute tolerance of 1e-8.

The regression is implemented in test/NetworkBuilder_test.jl. It also checks termination status, solution-group presence, input-data counts, node mappings, element mappings, and the compatibility accessors of PowerFlowResult.

Frequency-domain calculations

The same IEEE 39-bus fixture compares the Classic and NetworkModel routes after operating-point calculation. The tests retain the same frequency axis, node order, grounded-node treatment, and eliminated STATCOM selection. They compare:

  • linearized converter state matrices;
  • nodal and edge admittance matrices;
  • the reduced harmonic impedance tensor.

The impedance comparison uses relative and absolute tolerances of 1e-12. The nodal-admittance comparison uses relative and absolute tolerances of 1e-7, matching the existing fixture treatment of the DC-side dummy admittance.

The separate network-model performance comparison records warmed execution time and allocated bytes on the same frozen source and refactored working tree.