Benchmarks (grid size)

In this paragraph we will expose some brief benchmarks about the use of lightsim2grid in the grid2op settings. The code to run these benchmarks are given with this package int the [benchmark](./benchmarks) folder.

TODO DOC in progress

If you are interested in other type of benchmarks, let us know !

TL;DR

In summary, lightsim2grid (when using KLU linear solver) perfomances are:

grid

size (nb bus)

time (recycling)

time (no recycling)

time (TimeSerie)

time (ContingencyAnalysis)

case14

14

0.014566

0.0360995

0.00546858

0.0109907

case118

118

0.0835847

0.221692

0.038204

0.0535954

case_illinois200

200

0.167528

0.393047

0.0718412

0.112911

case300

300

0.277824

0.637875

0.1498

0.201345

case1354pegase

1354

1.58167

2.97318

0.909525

1.14373

case1888rte

1888

2.46013

4.16834

1.19669

1.4888

case2848rte

2848

3.82885

6.47039

1.81877

2.34153

case2869pegase

2869

3.73823

6.84596

2.1224

2.49936

case3120sp

3120

4.26664

7.29163

1.65787

2.47875

case6495rte

6495

11.8336

18.459

5.33566

6.15428

case6515rte

6515

13.6057

20.0196

5.22459

6.27699

case9241pegase

9241

17.3575

28.5716

8.76378

9.95424

All timings reported above are in milliseconds (ms) for one powerflow (in all cases lots of powerflow are carried out, up to a thousands and the timings here are averaged accross all the powerflows performed)

For detailed explanation about each column as well as the hardware used, please refer to the section below, but in summary:

  • benchmark were run on python 3.12 with a laptop (see section Using a grid2op environment and page Deep dive into the benchmarking of lightsim2grid for more information about the exact definition of the timers ):

  • time (recycling) indicates the average time it took to run 1 powerflow (with consecutive run of 288 powerflows) while allowing lighsim2grid to re use some basic previous computation from one powerflow to another. This is the most consommations usecase in grid2op for example (default behaviour). See Computation time using grid2op for more information

  • time (no recycling) indicates the same average time as aboved but lightsim2grid is forced to restart the computation from scratch each time, as if it was a completely different grid on a completely different computers. See Computation time using grid2op for more information.

  • time (TimeSerie) reports the time it takes to run one powerflow using the lightsim2grid TimeSerie module, were everything is in c++ and some care has been taken to improve the performance (reuse of as many things as possible, carefull memory allocation, etc.). See Computation time using the lightsim2grid TimeSerie module for more information.

  • time (ContingencyAnalysis) reports the time it takes to run one powerflow using the lightsim2grid ContingencyAnalysis module, were everything is in c++ and some care has been taken to improve the performance (reuse of as many things as possible, carefull memory allocation, etc.). See Computation time using the lightsim2grid ContingencyAnalysis module for more information. NB on this settings, as opposed to the others, the grid production / generations stay the same, but the grid topology changes by the connection and disconnection of powerlines.

Using a grid2op environment

In this section we perform some benchmark of a do nothing agent to test the raw performance of lightsim2grid on different grid sizes varying from the ieee case 14 grid (14 buses) up to the pegase 9241 grid (case9241 from pandapower counting 9241 buses).

All of them has been run on a computer with a the following characteristics:

  • date: 2026-04-21 09:51 CEST

  • system: Linux 6.8.0-60-generic

  • OS: ubuntu 22.04

  • processor: 13th Gen Intel(R) Core(TM) i7-13700H

  • python version: 3.13.5.final.0 (64 bit)

  • numpy version: 2.3.5

  • pandas version: 2.3.3

  • pandapower version: 3.4.0

  • grid2op version: 1.12.4.dev0

  • lightsim2grid version: 0.13.1

  • lightsim2grid extra information:

    • klu_solver_available: True

    • nicslu_solver_available: True

    • cktso_solver_available: True

    • compiled_march_native: False

    • compiled_o3_optim: True

Solver used for linear algebra: NR single (KLU)

To run the benchmark cd in the [benchmark](./benchmarks) folder and type:

python benchmark_grid_size.py

(results may vary depending on the hard drive, the ram etc. and are presented here for illustration only)

(we remind that these simulations correspond to simulation on one core of the CPU. Of course it is possible to make use of all the available cores, which would increase the number of steps that can be performed)

Computation time using grid2op

This benchmark in doing repeat calls to env.step(do_nothing) (usually 288 or 1000) for a given environment build on a grid coming from data available in pandapower.

Then we compare different measurments:

  • avg step duration (ms) is the average time it takes to perform the grid2op.step. It is given in milliseconds (ms). It takes into account the time to read the data, to feed the data to the underlying c++ model, to run the powerflow and to read back the data from the c++ model.

  • time [DC + AC] (ms / pf) is the time it takes to perform the entire powerflow, which consists in first providing an initial guess (DC approximation) and then to compute the powerflow. As compared to the above timings, it only take into account the time to run the powerflow. This “time to run the powerflow” can be at this stage decomposed in:

    • converting the provided data into valid matrix / vector to run a DC powerflow

    • computing a DC powerflow (used to initialize the AC powerflow)

    • converting again the provided data into valid matrix / vector to run an AC powerflow

    • computint the AC Powerflow

    • post processing the internal data (which includes eg the flows on the lines in amps, the reactive value produced / absorbed by each generator etc.)

  • time in ‘solver’ (ms / pf) gives the time it takes to only perform the AC powerflow:

    • converting the provided data into valid matrix / vector to run an AC powerflow

    • computing the AC Powerflow

    • post processing the internal data (which includes eg the flows on the lines in amps, the reactive value produced / absorbed by each generator etc.)

  • time in ‘algo’ (ms / pf) gives the time spent in the algorithm that computes the AC powerflow only

Warning

For more information about what is actually done and the wordings used in this section, you can consult the page Deep dive into the benchmarking of lightsim2grid

The results are given in two tables:

  • the first one corresponds to the default settings were lightsim2grid is allowed to “recycle” previous results, which is the default in grid2op and lightsim2grid. This corresponds to a generic grid2op usecase.

  • the second one is the same run for the same environment, but this time lightsim2grid recreate everything from scratch each time, the “recycling” is deactivated.

The main impact on “recycling” is that, when activated (default), lightsim2grid can skip some of its internal computation, especially in the steps:

  • “converting the provided data into valid matrix / vector to run a DC powerflow”

  • “converting again the provided data into valid matrix / vector to run an AC powerflow”

  • also the computation of the DC and AC powerflows can be a little bit faster (depending on the linear solver used)

The “no recycling” strategy is closer to a situation were you would simulate different powerflows on different cores or even on different computers and cannot share the internal state of the solvers (for example). It can also represent a situation were you would run powerflows for vastly different grids one after the other.

Results using grid2op.steps (288 consecutive steps, only measuring ‘dc pf [init] + ac pf’) (recyling allowed, default)

grid

size (nb bus)

avg step duration (ms)

time [DC + AC] (ms / pf)

speed (pf / s)

time in ‘solver’ (ms / pf)

time in ‘algo’ (ms / pf)

case14

14

0.317006

0.0236869

42217.4

0.014566

0.0110552

case118

118

0.439654

0.0977845

10226.6

0.0835847

0.0714621

case_illinois200

200

0.52879

0.184574

5417.89

0.167528

0.152389

case300

300

0.687898

0.300009

3333.23

0.277824

0.255286

case1354pegase

1354

2.44809

1.64378

608.353

1.58167

1.48944

case1888rte

1888

3.22291

2.54091

393.56

2.46013

2.34192

case2848rte

2848

4.75012

3.94042

253.78

3.82885

3.65101

case2869pegase

2869

5.20068

3.86489

258.74

3.73823

3.52338

case3120sp

3120

5.25079

4.3943

227.568

4.26664

4.08523

case6495rte

6495

13.6749

12.1174

82.5257

11.8336

11.3858

case6515rte

6515

15.4622

13.8893

71.9976

13.6057

13.1395

case9241pegase

9241

21.9036

17.8509

56.0197

17.3575

16.5355

Results using grid2op.steps (288 consecutive steps, only measuring ‘dc pf [init] + ac pf’) (no recycling allowed, non default)

grid name

size (nb bus)

avg step duration (ms)

time [DC + AC] (ms / pf)

speed (pf / s)

time in ‘solver’ (ms / pf)

time in ‘algo’ (ms / pf)

case14

14

0.368089

0.0590322

16939.9

0.0360995

0.0290161

case118

118

0.6709

0.305026

3278.4

0.221692

0.194384

case_illinois200

200

0.87917

0.509671

1962.05

0.393047

0.357242

case300

300

1.21275

0.808751

1236.47

0.637875

0.584015

case1354pegase

1354

4.48071

3.65709

273.441

2.97318

2.71566

case1888rte

1888

5.67926

4.99325

200.27

4.16834

3.88297

case2848rte

2848

8.57918

7.73595

129.267

6.47039

6.03774

case2869pegase

2869

9.76003

8.38388

119.277

6.84596

6.29586

case3120sp

3120

9.6056

8.71158

114.79

7.29163

6.85212

case6495rte

6495

23.2027

21.5196

46.4693

18.459

17.391

case6515rte

6515

24.7994

23.0958

43.2979

20.0196

18.9434

case9241pegase

9241

38.38

34.1234

29.3054

28.5716

26.5461

Computation time using the lightsim2grid TimeSerie module

As opposed to the experiment above, the TimeSerie lightsim2grid module allows to perform sequential computation of varying productions and loads with the exact same grid topology.

This does not rely on grid2op and is coded in “pure c++” still using one single CPU core. It should be faster than the timings reported on the above sequence because:

  • the loop is made in c++ instead of python

  • the code has been optimize to run faster and “recycle” as many things as possible: the matrices representing the grid is computed only once, it is factorized only once, conversion from the internal solver representation to MW, MVAr and A is done in a vectorized way etc.

This rapidity has a cost, it is much less flexible. With the grid2op framework an “agent” can do a lot of different actions (even though “do nothing” was used for the benchmark). Here on the other hand, only a “do nothing” action can be performed (and without emulation of any kind of protections).

The column time (ms / pf) can be compared with the column time [DC + AC] (ms / pf) of the table in the previous benchmark.

grid

size (nb bus)

time (ms / pf)

speed (pf / s)

case14

14

0.00546858

182863

case118

118

0.038204

26175.2

case_illinois200

200

0.0718412

13919.6

case300

300

0.1498

6675.56

case1354pegase

1354

0.909525

1099.48

case1888rte

1888

1.19669

835.64

case2848rte

2848

1.81877

549.823

case2869pegase

2869

2.1224

471.164

case3120sp

3120

1.65787

603.184

case6495rte

6495

5.33566

187.418

case6515rte

6515

5.22459

191.402

case9241pegase

9241

8.76378

114.106

Computation time using the lightsim2grid ContingencyAnalysis module

As opposed to the benchmarks reported in the previous two sections, this benchmark is focused on the ContingencyAnalysis lightsim2grid module.

A “contingency analysis” is often carried out in power system. The objective is to assess whether or not the current grid state is safe if one (or more) powerline would be disconnected. It uses the same productions / consommations for each computation. Each time it disconnects one or more powerlines, run the powerflow and then stores the results.

For this benchmark we focus on disconnecting only one powerline (though lightsim2grid offers the possibility to disconnect as many as you want) with a limit on 1000 contingency simulated (even for grid were there would be more than 1000 powerlines / trafos to disconnect we limit the computation to only 1000).

grid

size (nb bus)

time (ms / cont.)

speed (cont. / s)

case14

14

0.0109907

90986.1

case118

118

0.0535954

18658.3

case_illinois200

200

0.112911

8856.57

case300

300

0.201345

4966.6

case1354pegase

1354

1.14373

874.333

case1888rte

1888

1.4888

671.68

case2848rte

2848

2.34153

427.072

case2869pegase

2869

2.49936

400.102

case3120sp

3120

2.47875

403.429

case6495rte

6495

6.15428

162.488

case6515rte

6515

6.27699

159.312

case9241pegase

9241

9.95424

100.46