stvsd1314 / PPGN-Physics-Preserved-Graph-Networks

The increasing number of variable renewable energy (solar and wind power) causes power grids to have more abnormal conditions or faults. Faults may further trigger power blackouts or wildfires without timely monitoring and control strategy. Machine learning is a promising technology to accelerate the automation and intelligence of power grid monitoring systems. Unfortunately, the black-box machine learning methods are weak to the realistic challenges in power grids: low observation, insufficient labels, and stochastic environments. To overcome the vulnerability of black-box machine learning, we preserve the physics of power grids through graph networks to efficiently and accurately locate the faults even with limited observability and low label rates. We first calculate the graph embedding of power grid infrastructure by establishing a reduced graph network with the observed nodes, then efficiently locate the fault on the node level using the low-dimensional graph embedding. To augment the location accuracy at low label rates, we build another graph network representing the physical similarity of labeled and unlabeled data samples. Importantly, we provide the physical interpretations of the benefits of the graph design through a random walk equivalence. We conduct comprehensive numerical experiments in the IEEE 123-node. Our proposed method shows superior performance than three baseline classifiers for different fault types, label rates, and robustness to out-of-distribution (OOD) data. Additionally, we extend the proposed method to the IEEE 37-node benchmark system and validate the effectiveness of the proposed training strategy.

PPGN:Physics-Preserved-Graph-Networks

============================================================================== This software is to locate faults in distribution systems with limited observations and labels through PPGN. PPGN performs better than the baselines when labeled data are insufficient and the distribution of data vary randomly. Code accompanying the paper ["PPGN: Physics-Preserved Graph Networks for Fault Location with Limited Observation and Labels"]

Prerequisites

The proposed method is implemented through Jupyter Notebook. The required packages include:

• Jupyter Notebook
• Python 3
• Python packages: Numpy, torch, time, os, scipy, matplotlib

Getting started

1. You can train the proposed model with "training_123nodes.ipynb" and the test the well-trained model through "Testing_123nodes.ipynb" for the IEEE 123-node test case.
2. data: Check the "readme.md" file to find the training and testing datasets in various situations. Others are the mediate information used in the codes.
3. trained: this folder has the pre-trained models.