Detecting Demand Not Served (DNS) and Cascading Failures (CF) in Power Grids using Graph Neural Networks

Authors: Michael Zipperle and Min Wang

This project aims to develop an artificial intelligence-enhanced detection tool for predicting Demand Not Served (DNS) amd Cascading Failures (CF) in power grids. A deep learning model based on graph neural networks is designed for this purpose. The model predicts whether a DNS or CF will occur based on the input power grid status and topology. The dataset used for training the model was collected by the Reliability and Risk Engineering Group in ETH Zurich through a physics-based simulation model built on the Cascades platform [1]. For information about the dataset, please refer to the repository in reference [2]. This repository contains some testing subset which has been derived from the original dataset.

[1] https://rre.ethz.ch/research/research-pillars/power-systems/cascading-failure-in-power-systems.html

[2] https://github.com/k-amara/PowerGraph

Problem Definition

The aim of the detection tool is to distinglish between normal and abnormal power grid states. A normal powergrids state is characterized by the absence of DNS and CF.

We have defined the learning task of the detection tool as a binary or multiclass classification problem.

The multiclass labels are:

0: DNS and CF
1: DNS and no CF
2: no DNS and CF
3: no DNS and no CF

The binary labels are:

1: DNS and CF or DNS and no CF or no DNS and CF
0: no DNS and no CF

Preparation

Install all required python packages:

pip install -r requirements.txt

Download the testing subset and store in the datasets folder:

https://drive.google.com/drive/folders/1-jLgaCfTE7v3VBzfSUIfx3mK5Yli9yu1?usp=share_link

Run experiments on the testing subset

The experiments are highly configureable:

hidden_dims = [<int>, ...]
num_conv_layers = [<int>, ...]
num_linear_layers = [<int>, ...]
conv_types = [<ConvType>,...]
task_types = [<TaskType>,...]
dataset_types = [<DatasetType>,...]
batch_sizes = [<int>, ...]
poolings = [<PoolingType>,...]

The following options are available for the costum data_types:

ConvType: TRANSFORMER, GAT, GATv2
TaskType: BINARY, MULTICLASS
DatasetType: IEEE39, IEEE118
PoolingType: MEAN, MAX, SUM, ALL

We applied grid search to find the optimal parameters and included the corresponding models in this repository

hidden_dims = 32
num_conv_layers = 3
num_linear_layers = 2
conv_types = TRANSFORMER
task_types = N/A
dataset_types = N/A
batch_sizes = 128
poolings = MEAN

The results can be replicated by running the following code either directly in python or in a notebook:

import test
from itertools import product
from data_structures import *

# Define parameters
number_of_iterations = 1
hidden_dims = [32]
num_conv_layers = [3]
num_linear_layers = [2]
conv_types = [ConvType.TRANSFORMER]
task_types = [TaskType.MULTICLASS, TaskType.BINARY]
dataset_types = [DatasetType.IEEE118, DatasetType.IEEE39]
batch_sizes = [128]
poolings = [PoolingType.MEAN]


# Create a grid of all parameter combinations
param_grid = product(hidden_dims, num_conv_layers, num_linear_layers, conv_types, poolings, task_types, dataset_types, batch_sizes)

test.run_experiment(param_grid, number_of_iterations, debug=False)

The results are automatically stored in the experiments folder in the root directory.

Here is an overview of the results:

task_type	dataset	accuracy	balanced_accuracy	sensitivity	specificity	precision	recall	f1	TP	TN	FP	FN
multiclass	ieee118	0.967	0.828	0.967	0.000	0.977	0.967	0.970	NaN	NaN	NaN	NaN
multiclass	ieee39	0.968	0.969	0.968	0.000	0.981	0.968	0.972	NaN	NaN	NaN	NaN
binary	ieee118	0.995	0.994	0.993	0.995	0.949	0.993	0.970	1179.0	13749.0	64.0	8.0
binary	ieee39	0.991	0.995	1.000	0.990	0.883	1.000	0.938	235.0	3034.0	31.0	0.0

Train own models and run experiments

One can create a dataset of powergrid states with binary and multiclass labels and store them in respective folders. One must give the dataset a name and extend the DatasetType data structure in data_structures.py to support it.

Similar to running experiments on the test subsets, one can train and evaluate the model by running the following code either directly in python or in a notebook: