Deep Learning to Identify the Spatio-Temporal Cascading Effects of Train Delays in a High-Density Network

TL;DR

This work tackles the challenge of predicting and explaining delay cascades in dense rail networks by introducing XGeoAI, a live, explainable, multi-step forecasting framework based on a two-stage autoregressive GATv2. The model operates on a spatio-temporal graph of train events with Run, Dwell, and Headway edges, and incorporates novel congestion features to improve robustness under sequential rollout. A key contribution is the Edge Propagation Error (EPE) metric and an analysis showing the model leverages local congestion signals to anticipate propagation across headway edges, yielding higher precision even when raw MAE lags behind simple baselines. The results offer a practical blueprint for deployable decision support in real-world rail systems, highlighting the trade-offs between short-term accuracy and reliable, interpretable predictions, and pointing to avenues for richer features and more sophisticated graph architectures to further enhance live operational utility.

Abstract

The operational efficiency of railway networks, a cornerstone of modern economies, is persistently undermined by the cascading effects of train delays. Accurately forecasting this delay propagation is a critical challenge for real-time traffic management. While recent research has leveraged Graph Neural Networks (GNNs) to model the network structure of railways, a significant gap remains in developing frameworks that provide multi-step autoregressive forecasts at a network-wide scale, while simultaneously offering the live, interpretable explanations needed for decision support. This paper addresses this gap by developing and evaluating a novel XGeoAI framework for live, explainable, multi-step train delay forecasting. The core of this work is a two-stage, autoregressive Graph Attention Network (GAT) model, trained on a real-world dataset covering over 40% of the Dutch railway network. The model represents the system as a spatio-temporal graph of operational events (arrivals and departures) and is enriched with granular features, including platform and station congestion. To test its viability for live deployment, the model is rigorously evaluated using a sequential, k-step-ahead forecasting protocol that simulates real-world conditions where prediction errors can compound. The results demonstrate that while the proposed GATv2 model is challenged on pure error metrics (MAE) by a simpler Persistence baseline, it achieves consistently higher precision in classifying delay events -- a crucial advantage for a reliable decision support tool.

Figures (8)

• Figure 1: The complete two-stage GATv2 model architecture. Stage 1 & 2 (Input & Embedding Layer) processes raw node and edge features into fused vectors. Stage 3 (GATv2 Body) uses a stack of GATv2 layers with residual connections to learn graph representations. Stage 4 (Two-Stage Hurdle Prediction) splits the output into a Classifier and a Regressor and combines the output into a final delay forecast.
• Figure 2: Model performance degradation over the k-step forecast horizon. Top: MAE. Bottom: Precision.
• Figure 3: Comparative analysis of feature distributions for low vs. high-attention edges.
• Figure 4: Analysis of GATv2 attention scores. Top: Proportion of Edge Types in High-Attention Group. Bottom: Distribution of Attention Scores by Edge Type.
• Figure 5: Comparative error distribution by true delay magnitude.