PowerPINN

Physics-Informed Neural Networks (PINNs) for Power System Components

Overview

This repository provides a framework for generating and training Physics-Informed Neural Networks (PINNs) for power system components. It allows users to define Ordinary Differential Equations (ODEs), generate datasets, and train PINNs to approximate system dynamics efficiently.

Features

• Define and integrate new sets of ODEs for different power system components.
• Configure initial conditions and variable ranges.
• Automatically generate datasets using numerical solvers.
• Train and test PINNs using WandB for tracking.
• Fully parameterized using YAML configuration files.
• Modular design for easy extension.

Installation

Prerequisites

Ensure you have Python installed (>=3.8). Install required dependencies:

pip install -r requirements.txt

Usage

1. Define ODEs

ODEs are stored in src/ode/sm_models_d.py. You can add any new ODE model in this directory.

2. Configure Variables

The independent variables should be defined in modellings_guide.yaml, ensuring they are in the same order as in the ODEs.

3. Set Initial Conditions

Initial condition values and ranges should be specified in respective YAML files, located in the src/conf/initial_conditions/ folder under the corresponding ODE name (e.g., SM_AVR_GOV/init_cond.yaml).

4. Define Machine Parameters

Different machine parameters can be configured in the src/conf/params/ folder.

5. Generate Dataset

To generate the dataset for PINN training, use:

python create_dataset_d.py

Configuration file: setup_dataset.yaml

• time: Total simulation time.
• num_of_points: Number of data points per trajectory.
• modelling_method: Defines how state variables evolve.
• model: Specifies which ODE model to use (e.g., SM_AVR_GOV).
• sampling: Type of sampling for initial conditions (Lhs, Linear, Random).
• dirs: Paths for storing parameters, initial conditions, and dataset.

6. Train & Test a PINN

Train the PINN model with:

python test_sweep.py

Configuration file: setup_dataset_nn.yaml

• time, num_of_points, modelling_method: Same as dataset setup.
• seed: Ensures reproducibility.
• dataset: Defines data usage (shuffle, split_ratio, transform_input/output).
• nn (Neural Network Configuration):
  • type: Network architecture (DynamicNN, StaticNN, etc.).
  • hidden_layers, hidden_dim: Network depth and width.
  • optimizer: (Adam, LBFGS, etc.).
  • weighting: Adjusts loss function balance.
• dirs: Paths for dataset, model, and training parameters.

Configuration Files

File	Purpose
setup_dataset.yaml	Defines parameters for dataset generation
setup_dataset_nn.yaml	Defines parameters for neural network training
modellings_guide.yaml	Lists different ODE models and their variables
initial_conditions/	Folder containing initial conditions for each ODE
params/	Folder containing different synchronous machine parameters

Citation

If you use this repository in your research, please cite the following paper:

Ioannis Karampinis, Petros Ellinas, Ignasi Ventura Nadal, Rahul Nellikkath, Spyros Chatzivasileiadis, Toolbox for Developing Physics-Informed Neural Networks for Power System Components, DTU.

License

This project is licensed under the MIT License.