Cascading Failure Prediction in Power Grid Using Node and Edge Attributed Graph Neural Networks

In recent years, probabilistic data-driven methods have been gaining popularity for the study of cascading failures in power systems. These data-driven models are well-suited for analysis of large exploration spaces, historical event data, and online cascade prediction compared to computationally expensive physics-based methods. Moreover, propagation of failure is commonly observed in physical systems, preceding a cascading failure that occasionally results in widespread failure events such as blackouts in power grids. Considering such failure trends, we aim to investigate the use of Graph Neural Networks (GNNs) for the analysis and prediction of cascading failures in power systems by means of information propagation. We focus on data-driven modeling and formulate end-to-end mechanisms to train models capable of predicting the cascading effects of various failure-inducing events introduced to power grid profiles. Unlike physics-based models that require iterative power flow solutions, we predict post-event grid component vulnerability based only on the initiating events and network operational status. We evaluate several representative GNN models in this framework. We also formulate a graph convolution mechanism to perform message-passing between power grid components using both bus and branch features. Our experimental evaluation demonstrates the efficiency of GNN-based methods for end-to-end cascading failure prediction on the IEEE 39-bus and 118-bus test systems.