PGLib-OPF Benchmark Validation

Script: scripts/pglib_benchmark.py

Benchmarks potpourri DC-OPF and AC-OPF against the published reference objective values from the IEEE PES Power Grid Library (PGLib-OPF), solved by PowerModels.jl with IPOPT.

Purpose

PGLib-OPF is the standard benchmark suite for comparing OPF solvers and formulations. Each case ships as a MATPOWER .m file with a published reference objective value (minimum generation cost in $/h) for DC and AC OPF under three scenario groups:

Group	Abbreviation	Description
Typical Operations	TYP	Near-nominal loading, baseline scenario
Congested Operations	API	Elevated loading, active constraints
Small Angle Difference	SAD	Tight phase-angle bounds

The benchmark script runs potpourri DC and AC OPF on a configurable subset of TYP cases and reports the objective gap relative to the PowerModels.jl reference.

Setup

The PGLib-OPF case files are included as a git submodule. Initialise it once:

git submodule update --init benchmarks/pglib-opf

Or clone with submodules from the start:

git clone --recurse-submodules https://github.com/RWTH-IAEW/opf-potpourri.git

Workflow

for each selected PGLib case (.m file)
    ┌─ load via load_pglib_case()
    │    (from_mpc → pandapower net, attach ANGMIN/ANGMAX, rebalance dispatch)
    ├─ DC-OPF: DCOPF + add_poly_cost_objective + angle_limits=True
    ├─ AC-OPF: ACOPF + add_poly_cost_objective
    │    + thermal_limit='mva'
    │    + free_slack_vm=True
    │    + fix_hv_buses=False
    │    + angle_limits=True
    └─ report objective vs BASELINE.md reference, write results/

The PGLib-compatible flags are passed directly to ACOPF.add_OPF; no monkey-patching of the model is required.

Usage

conda activate potpourri_env

# Default: small and medium cases up to 300 buses
python scripts/pglib_benchmark.py

# Include larger cases (up to 1500 buses)
python scripts/pglib_benchmark.py --max-buses 1500

# Specific cases only
python scripts/pglib_benchmark.py --cases case5_pjm case14_ieee case30_ieee

# AC only (skip DC-OPF)
python scripts/pglib_benchmark.py --no-dc

Results are written to results/pglib_benchmark.csv and results/pglib_benchmark.md.

Example output

Running 6 case(s):
  - case5_pjm       (  5 buses)
  - case14_ieee     ( 14 buses)
  - case24_ieee_rts ( 24 buses)
  - case30_ieee     ( 30 buses)
  - case57_ieee     ( 57 buses)
  - case118_ieee    (118 buses)

case5_pjm         (    5 buses)  DC:   17551.89 (ref   17551.89, +0.00%, 0.4s) ✓  AC:   17551.89 (ref   17551.89, +0.00%, 1.2s) ✓
case14_ieee       (   14 buses)  DC:    7642.59 (ref    7642.59, +0.00%, 0.5s) ✓  AC:    8081.52 (ref    8081.52, +0.00%, 1.8s) ✓
case24_ieee_rts   (   24 buses)  DC:   76943.24 (ref   76943.24, +0.00%, 0.6s) ✓  AC:   63352.20 (ref   63352.20, +0.00%, 2.1s) ✓
case30_ieee       (   30 buses)  DC:     576.89 (ref     576.89, +0.00%, 0.5s) ✓  AC:     576.89 (ref     576.89, +0.00%, 2.3s) ✓
case57_ieee       (   57 buses)  DC:   41006.74 (ref   41006.74, +0.00%, 0.8s) ✓  AC:   41737.79 (ref   41737.79, +0.00%, 3.4s) ✓
case118_ieee      (  118 buses)  DC:   97212.83 (ref   97212.83, +0.00%, 1.2s) ✓  AC:  129338.15 (ref  129338.15, +0.00%, 6.1s) ✓

Key API

`potpourri.benchmarks.load_pglib_case`

from potpourri.benchmarks import load_pglib_case

net = load_pglib_case("case14_ieee")

Converts a PGLib .m file into a pandapowerNet with:

All generators flagged controllable=True
Polynomial cost coefficients in net.poly_cost
Phase-angle limits (angmin_degree / angmax_degree) on net.line and net.trafo, read from the MATPOWER ANGMIN / ANGMAX fields
Initial dispatch rebalanced so pp.runpp() converges

`potpourri.models.cost_objective.add_poly_cost_objective`

from potpourri.models.cost_objective import add_poly_cost_objective

acopf = ACOPF(net)
acopf.add_OPF(thermal_limit="mva", free_slack_vm=True, angle_limits=True)
add_poly_cost_objective(acopf)
acopf.solve(solver="ipopt")

Wires net.poly_cost coefficients (c2·P² + c1·P + c0) into a Pyomo objective over all controllable generators (ext_grid, gen, sgen).

Reference baselines

from potpourri.benchmarks import PGLIB_BASELINE_TYP, PGLIB_BASELINE_API, PGLIB_BASELINE_SAD

ref = PGLIB_BASELINE_TYP["pglib_opf_case14_ieee"]
print(ref["dc"], ref["ac"])   # 7642.59, 8081.52

Known limitations

Large transmission cases with many net.impedance rows (e.g. case200_activ, case588_sdet) show objective gaps of ±8–32 % relative to the PowerModels.jl reference. Smaller cases (case118_ieee, case197_snem) match to ≤ 0.01 %. The gap on larger cases is most plausibly explained by PowerModels.jl's input-preprocessing pipeline (topology simplification, low-impedance branch merging) that potpourri does not perform.
Only the Typical Operations (TYP) baseline is used by default. API and SAD cases can be loaded via load_pglib_case but are not included in the automated benchmark table.
DC-OPF in PGLib uses linear costs only; add_poly_cost_objective accepts quadratic terms but warns when c2 > 0 entries are present.