AC Power Flow Validation

Script: scripts/validate_ac_model_against_pandapower.py

Validates the potpourri AC power flow solver against pandapower's Newton-Raphson power flow on six standard SimBench benchmark networks.

Purpose

The AC class in potpourri formulates AC power flow as a Pyomo optimisation problem and solves it with an NLP solver (IPOPT or NEOS). This script provides a quantitative comparison of the results against pandapower's reference solver.

Workflow

for each network in NETS
    ┌─ load simbench network
    ├─ solve with potpourri AC (NEOS / IPOPT)
    ├─ solve with pandapower pp.runpp()
    ├─ compute error metrics on res_bus and res_line
    └─ collect results
print pivot tables and error ranking

1. Load the network via simbench.get_simbench_net().
2. Build and solve the AC model — the constructor runs pp.runpp() internally to initialise voltages and angles; the NLP solver then finds the feasible AC power flow solution.
3. Run the reference power flow with pp.runpp() on the original network using voltage-independent loads.
4. Compare the following fields:

1. Report mean absolute error (MAE), maximum absolute error, RMSE, and standard deviation per field; identify the bus or line with the largest deviation.

Networks tested

Usage

Requires IPOPT (included in environment.yaml):

conda activate potpourri_env
python scripts/validate_ac_model_against_pandapower.py

To use the NEOS cloud solver instead (no local solver required), uncomment the corresponding ac.solve(solver="neos") line in the script and set:

export NEOS_EMAIL="your@email.address"

Example output

Detailed comparison:
                    net   element  variable  mean_abs   max_abs      rmse   std_abs  worst_index  solver_converged  vm_pu_close
0    1-HV-mixed--0-sw  res_bus     vm_pu  0.023850  0.042419  0.026011  0.010122            5              True        False
1    1-HV-mixed--0-sw  res_bus  va_degree  0.234032  0.739218  0.289415  0.173017           12              True        False
...

Compact pivot table (mean absolute error):
variable              pl_mw   ql_mvar  va_degree     vm_pu
net
1-HV-mixed--0-sw   0.012336  0.051528   0.234032  0.023850
1-LV-rural1--0-sw  0.002670  0.001040   7.052760  0.280162
1-MV-rural--0-sw   0.050133  0.027593   2.634818  0.035633

Networks ranked by total mean absolute error:
net
1-HV-mixed--0-sw     0.321746
1-MV-rural--0-sw     2.748177
1-LV-rural1--0-sw    7.336632

Interpreting the results

The AC model treats static generator reactive power (qsG) as a free optimisation variable, whereas pandapower fixes it to the value specified in net.sgen.q_mvar. This structural difference means the NLP solver finds a different (but equally valid) AC power flow solution in terms of reactive power dispatch, leading to measurable voltage angle and magnitude differences relative to pandapower.

Key observations:

• HV networks (larger impedances, stronger voltage coupling) show smaller relative errors than LV networks.
• vm_pu_close (within 1×10⁻³ p.u. of pandapower) is False for all networks due to the free reactive dispatch — this is expected behaviour.
• For pure power flow validation (fixed reactive dispatch), initialise the qsG variables from a prior pp.runpp() result using init_pyo_from_pp_res.

Key functions

calculate_error_metrics(reference, candidate)

compare_results(pp_net, ac_net, delta_keys)