Graph Neural Network Framework for Line Status Prediction in Power Systems Cascading Failures

Abstract

Cascading failures in power systems are complex, high-impact phenomena that can result in significant financial and societal impacts. Accurately predicting the severity of these cascades and the postcascade status of transmission lines early in the event can greatly aid in risk assessment and the formulation of effective mitigation strategies. Graph neural network (GNN)-based models are emerging as powerful tools for such predictive analyses in power systems due to their ability to capture topological relationships among system measurements. In this work, various GNN-based models are proposed to improve the prediction of transmission line statuses and overall cascade size using the initial cascade data. Performance evaluations on the IEEE 118-bus and 300-bus systems demonstrate that GNN-based models outperform the traditional machine learning models that do not incorporate topological information. In addition, the sensitivity of the proposed GNN-based models to noise in the initial cascade data is evaluated, highlighting the critical importance of accurate initial trigger information.