[feature] Improve Python-Based OPF Speed

[determ1ne]

Feature Checklist

• Searched the issues page for similar feature requests

• Read the relevant sections of the documentation

• Browse the repository, tutorials and tests for already existing functionalities

Feature Type

• Adding new functionality to pandapower

• Changing existing functionality in pandapower

• Removing existing functionality in pandapower

Problem Description

Hi, pandapower community! I am currently working on a yet another MATPOWER port to Python called pwrs.

Compared to pypower, it was targeted at MATPOWER 7.1.
The OPF related code was updated to the corresponding MATPOWER/MIPS version, and the speed was really great.
Here is a benchmark:

import pwrs as mp
import numpy as np
from pandapower.pypower.opf import opf
from pandapower.pypower.ppoption import ppoption
import time

mpc = mp.case9241pegase()
mp.runpf(mpc)

ppc = mpc.to_dict()

def enlarge(m, col):
    b = np.zeros((m.shape[0], col))
    b[:, :m.shape[1]] = m
    return b

ppc["bus"][:, 0] -= 1
ppc["gen"][:, 0] -= 1
ppc["branch"][:, 0:2] -= 1
ppc["bus"] = enlarge(ppc["bus"], 18)
ppc["gen"] = enlarge(ppc["gen"], 26)
ppc["branch"] = enlarge(ppc["branch"], 27)

# currently pandapower runopp is broken, see #2932 
t = time.perf_counter()
result = opf(ppc, ppoption(VERBOSE=3, PF_DC=0, INIT="flat"))
t2 = time.perf_counter()
print(f"pypower opf took {t2 - t:.4f} seconds")

mpc = mp.case9241pegase()
mp.runpf(mpc)
t = time.perf_counter()
mp.runopf(mpc)
t2 = time.perf_counter()
print(f"pwrs opf took {t2 - t:.4f} seconds")

Output:

...
pypower opf took 319.8075 seconds
...
pwrs opf took 7.4259 seconds

As a comparison, MATPOWER 8.0 (MATLAB R2024a, Windows 11, Intel Core i7 13700K) took 11.69 seconds with runopf(case9241pegase).

The code was oracle tested with MATPOWER, for every data case came bundled with MATPOWER.

Feature Description

Add: pwrs as a new backend for pandapower OPF

Additional Context

No response

Label

• Relevant labels are selected

[vogt31337]

@determ1ne , thanks for the headsup! Looks interesting what you are building there. Do you already have some documentation?
In pp we are able to make the "solver" (lack of better word) exchangable. So we could interface with yours. But since you know your pwrs the best maybe think about an PR?
Our version of ppc is heavily modified to support a lot of additional features. Internally we use ppci, which would be the best starting point, since it represents a grid which is reduced and fused (out-of-service elements removed, closed bus-bus switches fused to a single bus bar, etc.).

[determ1ne]

@determ1ne , thanks for the headsup! Looks interesting what you are building there. Do you already have some documentation? In pp we are able to make the "solver" (lack of better word) exchangable. So we could interface with yours. But since you know your pwrs the best maybe think about an PR? Our version of ppc is heavily modified to support a lot of additional features. Internally we use ppci, which would be the best starting point, since it represents a grid which is reduced and fused (out-of-service elements removed, closed bus-bus switches fused to a single bus bar, etc.).

I'd like to make a PR for testing. However, I'm having trouble running cases (like #2932 ). Is there any alternative OPF examples that I could start with?

[vogt31337]

@determ1ne , I think resolving these problems will take some time. As far as I understood the modelling has changed internally. For example a 3W Trafo will be split into three 2W Trafos and this has to be reversed. Also we introduced a calculation which changes a Trafo from Wye to Delta configuration. I think these are the culprits why #2932 and #2620 are still open.

For your case, like I wrote you can use the internal ppc structure, ppc = pypowercasefile, which is what you want to create with the to_mpc method (=to matpowercasefile). It should be compatible with your code, except that it models a lot more things (for example VSCs and other FACT devices).

There is even a tutorial under "tutorials/internal_datastructure.ipynb" which shows how these internal datastructures can be used.