Physics-Informed Graph Neural Jump ODEs for Cascading Failure Prediction in Power Grids

TL;DR

This paper introduces a physics-informed graph neural jump ODE model for real-time cascading failure prediction in power grids, integrating physical laws and temporal dynamics to outperform existing methods.

Contribution

The paper presents a novel neural network architecture combining physics-based regularization, continuous-time modeling, and jump processes for accurate, real-time cascade prediction.

Findings

Achieves high AUC scores for failure detection on IEEE systems

Outperforms standard graph neural networks in accuracy

Physics-informed loss significantly improves demand regression

Abstract

Cascading failures in power grids pose severe risks to infrastructure reliability, yet real-time prediction of their progression remains an open challenge. Physics-based simulators require minutes to hours per scenario, while existing graph neural network approaches treat cascading failures as static classification tasks, ignoring temporal evolution and physical laws. This paper proposes Physics-Informed Graph Neural Jump ODEs (PI-GN-JODE), combining an edge-conditioned graph neural network encoder, a Neural ODE for continuous power redistribution, a jump process handler for discrete relay trips, and Kirchhoff-based physics regularization. The model simultaneously predicts edge and node failure probabilities, severity classification, and demand not served, while an autoregressive extension enables round-by-round temporal cascade prediction. Evaluated on the IEEE 24-bus and 118-bus systems with 20,000 scenarios each, PI-GN-JODE achieves a Precision--Recall Area Under the Curve of 0.991 for edge failure detection, 0.973 for node failure detection, and a coefficient of determination of 0.951 for demand-not-served regression on the 118-bus system, outperforming a standard graph convolutional network baseline (0.948, 0.925, and 0.912, respectively). Ablation studies reveal that the four components function synergistically, with the physics-informed loss alone contributing +9.2 points to demand-not-served regression. Performance improves when scaling to larger grids, and the architecture achieves the highest balanced accuracy (0.996) on the PowerGraph benchmark using data from a different simulation framework.