PowerGraph

We have created a graph dataset that models cascading failure events, which are the main cause of blackouts in power grids. To generate a comprehensive dataset, we used a physics-based cascading failure model called Cascades. This model simulates how failures propagate in a power grid, resulting in the final demand not being served (DNS). Our dataset consists of a large set of power grid states, representing the operating conditions before an outage, and is linked to an initial triggering outage (one or more failed elements). Each power grid state is represented as a graph, with a graph-level label assigned based on the results of the physics-based model. The dataset is designed for various graph-level tasks, such as multi-class classification, binary classification, and regression. Bus and branches are the elements of a power grid, buses include loads and generators which represent the nodes of the graph, while branches include transmission lines and transformers which represent the edges of the graph. We provide three features per node: net active power, net apparent power and voltage magnitude. While the features per edge are four: active power flow, reactive power flow, line reactance and line rating. Finally, we assign ground-truth explanations on the edges, allowing the dataset to be used to benchmark GNN explainability models for graph level tasks.