Prediction and mitigation of nonlocal cascading failures using graph neural networks

TL;DR

This paper introduces a graph neural network approach to predict and mitigate nonlocal cascading failures in power grids, enabling efficient analysis of large networks and improving failure prevention strategies.

Contribution

It proposes an avalanche centrality measure and trains a GNN to predict failure propagation in large networks, enhancing mitigation methods.

Findings

GNN accurately predicts avalanche centrality in large networks

The framework reduces computational complexity for large-scale simulations

Effective mitigation strategies can be derived from GNN predictions

Abstract

Cascading failures (CFs) in electrical power grids propagate nonlocally; After a local disturbance, the second failure may be distant. To study the avalanche dynamics and mitigation strategy of nonlocal CFs, numerical simulation is necessary; however, computational complexity is high. Here, we first propose an avalanche centrality (AC) of each node, a measure related to avalanche size, based on the Motter and Lai model. Second, we train a graph neural network (GNN) with the AC in small networks. Next, the trained GNN predicts the AC ranking in much larger networks and real-world electrical grids. This result can be used effectively for avalanche mitigation. The framework we develop can be implemented in other complex processes that are computationally costly to simulate in large networks.