Validation against pandapower and pandapipes¶

To trust a general multi-energy NLP, check it against established single-purpose reference tools. monee’s nonlinear models are validated against pandapower (electricity) and pandapipes (gas, water/heat, and sequential multi-energy coupling) on like-for-like networks.

The headline:

Single-sector power flow reproduces the reference solution almost always: voltages to ≤ 5e-3 pu, pressures to ~1 % of the pipe drop, temperatures to ≤ 1e-2 K. On raw power-flow speed the specialised tools win, since a dedicated Newton or hydraulic Jacobian beats a general NLP, as expected.
Optimization (AC OPF, economic dispatch) reaches pandapower’s optimum to machine precision and is frequently faster. A general NLP is no longer at a structural disadvantage once the reference tool also has to solve a nonlinear program.
Multi-energy coupling (P2G, G2P, CHP) is monee’s home turf. It solves every carrier in one simultaneous NLP, matching pandapipes’ iterative multinet to ~1e-11 pu on voltage while solving the coupled problem faster.

Electricity, against pandapower (identical model)¶

The electrical model is handed to both engines through the neutral MATPOWER mpc exchange, so any difference is the solver, not the model. monee uses its in-process CasADi/IPOPT backend; pandapower uses its native Newton-Raphson runpp and the PYPOWER AC-OPF runopp.

Regime	Voltage agreement	Power agreement	Speed vs pandapower
Power flow (CIGRE MV / MV+DER / LV)	≤ 5e-3 pu	≤ 6e-4 MW	~1.4 to 4× slower (NR wins)
OPF, economic dispatch	≤ 3e-8 pu	≤ 3e-5 MW	often faster than `runopp`
OPF + line-loading limit	≤ 8e-7 pu	≤ 2e-4 MW	often faster than `runopp`

For OPF the cost objective matches pandapower to a relative error ≤ 4e-7, so the two optimisers find the same economic optimum. The line-limited case has the cap binding on both tools, so monee’s current and loading constraints are exercised too.

Note

monee imports every generator as a fixed-PQ injection with the external grid as the voltage reference, a distribution PQ-with-slack scope. MATPOWER transmission cases with PV (voltage-controlled) generators are out of scope and are not part of this validation set. This is a modelling-scope choice, not a solver-accuracy result.

Gas, heat, and coupled MES, against pandapipes¶

Gas and heat have no neutral exchange format. monee solves a smooth Weymouth (gas) or Darcy-Weisbach plus temperature-transport (heat) NLP; pandapipes solves its own detailed hydraulic and thermal Newton system. So like-for-like networks are built independently in each tool from one spec, and the comparison is agreement within tolerance.

Suite	Agreement	Speed vs pandapipes
GAS hydraulics	pressure within 1.2 % of the pipe drop	~1.3 to 7× slower (`pipeflow` wins)
HEAT hydraulics + thermal	pressure ~0.1 %, temperature ≤ 8.7e-3 K	~1.4 to 23× slower (`pipeflow` wins)
MES P2G / G2P / CHP (coupled)	bus voltage to ~1e-11 pu, CHP heat-exchanger ΔT within 0.07 K	1.4 to 5× faster

The MES result is the point of the whole exercise. pandapipes couples a pandapower net and pandapipes gas/water nets through multinet controllers that iterate to a fixed point between sectors; monee writes all carriers into a single NLP and solves them simultaneously. The simultaneous solve agrees to machine precision on the shared electrical quantities and is faster than the iterative coupling.

Reading the story¶

The two “slower on single-sector power flow” results are not a weakness; they are the expected cost of generality. A power-flow Jacobian or a hydraulic Newton step is a specialised routine. A general interior-point NLP does strictly more work to offer the same model the freedom to become an OPF, gain a custom constraint, or couple to another carrier. Once the problem stops being a plain single-sector power flow and becomes an optimisation or a multi-energy coupling, that generality stops costing and starts paying.

Reproduce locally:

python benchmarks/pandapower_comparison.py      # electricity
python benchmarks/pandapipes_comparison.py     # gas / heat / MES
# add --plot-only to re-render the figures from the saved CSVs