Train delays are the result of a complex interplay between multiple factors, evolving in nonlinear and often unpredictable ways. While delays at one station can propagate to subsequent stops, their spread is influenced by elements such as weather conditions, network congestion, and unexpected events like strikes or technical failures. Moreover, these disruptions do not remain confined to a single railway line but can cascade across the entire network, amplifying scheduling issues on a broader scale.
This project does not primarily aim to develop the best predictive model for train delays. Instead, it focuses on evaluating how the integration of topological information - derived from railway network structure - impacts predictive performance. Traditional machine learning models rely on historical and tabular data, but this study explores whether incorporating graph-based features improves their accuracy. By comparing baseline models with those enhanced through feature engineering and network-aware methodologies, we aim to assess the contribution of structural properties to delay prediction. Additionally, the study extends to sequential modeling with LSTMs to investigate how different feature sets influence time-series forecasting.
Several challenges arise in this analysis. The first is the inherent nonlinearity and interdependence of factors affecting train delays, which require models capable of capturing complex interactions. Another key challenge is modeling the propagation of delays across interconnected railway lines, necessitating the use of network-aware approaches. Finally, incorporating temporal dependencies is crucial, as delays evolve dynamically over time, demanding methods that can effectively learn sequential patterns.