Graph Neural Networks for Supply Disruption Forecasting in Logistics Partner Networks

Abstract

Global logistics networks have evolved into highly complex, interconnected systems that are increasingly vulnerable to cascading disruptions caused by natural disasters, geopolitical tensions, and unforeseen economic shifts. Accurate prediction of supply chain disruptions is a critical objective for organizations seeking to maintain operational resilience and minimize financial losses. Traditional forecasting methods often fail to capture the non-linear, spatial-temporal dependencies inherent in modern multi-tier supply networks. This paper presents a comprehensive investigation into the application of graph neural networks for predicting supply disruption forecasting within logistics partner networks through an extensive simulation study. By representing the supply chain as a heterogeneous graph where nodes represent operational entities and edges denote relational dependencies such as material flow and communication, we demonstrate how advanced representation learning can capture complex topological characteristics. The study utilizes a highly detailed discrete-event simulation to generate realistic, synthetic supply chain datasets encompassing various disruption scenarios, including supplier bankruptcies, transportation delays, and sudden demand surges. Through rigorous comparative analysis against conventional machine learning baselines, the graph-based approach exhibits superior predictive accuracy and resilience in identifying latent vulnerabilities before they propagate through the network. The findings suggest that integrating structural network data with temporal operational metrics significantly enhances predictive forecasting capabilities. This research contributes to the growing body of knowledge in supply chain risk management and provides a scalable, data-driven framework for practitioners to proactively mitigate structural vulnerabilities in their logistics operations.