Graph Neural Networks for Real-Time Rail Disruption Prediction and Passenger-Centric Re-Routing

Rail networks frequently experience disruptions that can lead to widespread delays and missed connections. This review examines how graph neural networks have emerged as practical tools for real-time disruption prediction and management in railway operations from 2020 to the present. Unlike traditional forecasting methods that treat stations in isolation, graph-based models can capture the network’s interconnected nature by learning which upstream delays matter and when downstream impacts will arrive. Early deployments demonstrate the tangible benefits: British rail systems achieved improved short-horizon predictions for connections, and Dutch railways discovered that headway dominates delay cascades. France’s national network now runs minute-by-minute forecasts for thousands of trains. The most successful implementations combine several key elements: joint modeling of spatial network structures and temporal patterns, adaptive connectivity, and integration of railway knowledge, such as headways and buffer times, to produce interpretable outcomes that explain why specific connections are at risk. Operational deployment requires matching prediction horizons to decision windows, mixed with robust handling of degraded data quality and careful monitoring for bias against low-traffic routes. The challenges remain, particularly the lack of shared disruption datasets and the limited number of field trials. However, the convergence of improved data feeds, faster inference techniques, and mature graph architecture can make real-time network-aware disruption management a reality. Proper implementation, focusing on operator-relevant metrics rather than abstract accuracy advances, can deliver measurable reductions in passenger minutes lost and missed connections.