mlpf 0.0.7

Machine learning for power flow

MLPF is a python library for (optimal) power flow calculations with machine learning. It offers:

• efficient loss functions compatible with both PyTorch and scikit-learn
• utilities such as data structures and loading pipelines that make it easy to go from pandapower nets or PYPOWER case files to arrays and tensors in just one line of code
• visualization and description tools to take a quick look at your data

Contributions welcome!

Installation

pip install mlpf

The previous command will install all the dependencies for working with numpy and scikit-learn. It will not, however, install all the dependencies needed for working in torch. To use the torch functionalities, please install PyTorch, PyTorch Geometric with its dependencies(torch-scatter etc.) and optionally TorchMetrics.

Usage

1. Load/create data and turn it into the PYPOWER case format
2. From then on we provide functionality to express the data as numpy arrays or torch tensors.
3. Feed that data into your ML (scikit-learn or torch) models and use our tried and tested loss functions to train, validate or monitor your model development.

import pandapower as pp
import pandapower.networks as pn

from pypower.ppoption import ppoption
from pypower.runpf import runpf

net = pn.case118()

ppc = pp.converter.to_ppc(net, init="flat")

ppopt = ppoption(OUT_ALL=0, VERBOSE=0)
ppc, converged = runpf(ppc, ppopt=ppopt)

Loss

numpy / scikit-learn

import numpy as np

from mlpf.data.conversion.numpy.power_flow import ppc2power_flow_arrays
from mlpf.loss.numpy.power_flow import active_power_errors, reactive_power_errors

edge_index, active_powers_pu, reactive_powers_pu, voltages_pu, angles_rad, conductances_pu, susceptances_pu = ppc2power_flow_arrays(ppc)

active_errors = active_power_errors(edge_index, active_powers_pu, voltages_pu, angles_rad, conductances_pu, susceptances_pu)
reactive_errors = reactive_power_errors(edge_index, reactive_powers_pu, voltages_pu, angles_rad, conductances_pu, susceptances_pu)

print(f"Total P loss = {np.sum(active_errors):.3e} p.u.")
print(f"Total Q loss = {np.sum(reactive_errors):.3e} p.u.")

torch

import torch

from mlpf.data.conversion.torch.power_flow import ppc2power_flow_tensors
from mlpf.loss.torch.power_flow import active_power_errors, reactive_power_errors

# note: going from float64(standard in PYPOWER) to float32(standard in torch) will increase the PF loss significantly
edge_index, active_powers_pu, reactive_powers_pu, voltages_pu, angles_rad, conductances_pu, susceptances_pu = ppc2power_flow_tensors(ppc, dtype=torch.float64)

active_errors = active_power_errors(edge_index, active_powers_pu, voltages_pu, angles_rad, conductances_pu, susceptances_pu)
reactive_errors = reactive_power_errors(edge_index, reactive_powers_pu, voltages_pu, angles_rad, conductances_pu, susceptances_pu)

print(f"Total P loss = {torch.sum(active_errors):.3e} p.u.")
print(f"Total Q loss = {torch.sum(reactive_errors):.3e} p.u.")

Data loading

• TODO

Indepth examples

General

• NumPy/scikit-learn loss
• Torch loss

Supervised learning

Power flow

• scikit-learn linear regression
• torch MLP(multilayer perceptron)
• torch GCN(graph convolutional network)

Supervised learning

Power flow

• torch MLP(multilayer perceptron
• torch GCN(graph convolutional network)

Development

git clone https://github.com/viktor-ktorvi/mlpf.git
cd mlpf

conda env create -f environment.yml
conda activate mlpfenv