TLDR: This research explores adapting a machine learning framework, originally for predicting link removals in fast-changing networks like US air transport, to predict delayed train trajectories in the Dutch railway network. While the framework performed well on air data, it showed limited success in predicting Dutch railway delays, even with added network features and different classifiers. The study highlights that the unique, dense, and ripple-effect-prone nature of the Dutch railway system requires more context-specific adaptations and a broader range of features beyond just network topology for accurate delay prediction.
The Dutch railway network, known for being one of the busiest globally, faces a persistent challenge: train delays. These delays are a major concern for the principal passenger railway operator, NS, with reports indicating that one in ten journeys experienced delays in 2023. Addressing this critical issue, recent research explores the application of machine learning to predict these delays, moving beyond traditional short-term forecasts to consider broader network-wide patterns.
Understanding the Delay Prediction Gap
Previous studies on train delays have largely concentrated on short-term predictions, often focusing on the next station within a 20-minute timeframe. While valuable, these approaches tend to overlook the larger ripple effects that a single delay can trigger across an entire transportation system. This new research aims to bridge this gap by investigating how network-wide topological features can be used to predict delayed trajectories, rather than just individual short-term delays.
Adapting a Predictive Framework
The study builds upon an existing machine learning framework developed by Lei et al., which was originally designed to forecast changes, specifically ‘removed links,’ in fast-evolving networks like the US air transport system. The core idea was to adapt this framework to the Dutch railway network, where ‘removed links’ are redefined as ‘delayed links.’ This redefinition allows the problem of predicting delays to be treated as a binary classification problem: whether a trajectory is significantly delayed or not.
The researchers utilized publicly accessible data from rijdendetreinen.nl, which includes real-time departure and arrival times, and service updates for all passenger train services in the Netherlands since 2019. This extensive dataset underwent significant transformation to fit the model, focusing on ‘trajectories’ (source-target station pairs) rather than individual stops. A key step involved defining a ‘significant delay’ for a trajectory: if more than 21% of trains on that trajectory in a given month experienced a final arrival delay, it was classified as significantly delayed.
Features and Models Explored
The models primarily leveraged ‘topological features,’ which describe the structural properties of the network, such as common neighbors between stations or the strength of connections. An important improvement introduced in this research was the inclusion of ‘Node Centrality Measures’ like Degree Centrality (number of direct connections), Closeness Centrality (how close a station is to all others), and Node Strength (total traffic through a station). These measures were added to capture the importance and influence of individual stations within the network.
The study employed the XGBoost Classifier, identified as a top performer in the baseline work, and also evaluated other common machine learning classifiers such as RandomForest, DecisionTree, GradientBoosting, AdaBoost, and LogisticRegression. Models were tested using two methods: ‘simultaneous testing’ (training and testing on data from the same period) and ‘non-simultaneous testing’ (training on past data and testing on future data to assess generalization over time).
Results and Challenges
While the implemented baseline model successfully replicated its original performance in predicting removed links in the US air network, its application to the Dutch railway data yielded less favorable results. The balanced accuracy scores for predicting delayed trajectories in the Dutch network were significantly lower, hovering just above 50%—meaning the models were not much better than random guessing, especially in non-simultaneous testing scenarios.
Even with the addition of node centrality measures and the exploration of various other machine learning classifiers, the predictive performance remained poor. This suggests that the framework, while effective for fast-changing networks like air transport, may not be directly applicable to the unique dynamics of the Dutch railway system.
Why the Discrepancy?
The researchers point to several factors that might explain the limited performance. The Dutch railway network is exceptionally dense and highly sensitive to minor timetable changes, leading to a significant ‘ripple effect’ where a small congestion can cause widespread delays. It’s possible that topological and network features alone are insufficient to capture the complexity of these delays, which can be influenced by operational factors, environmental conditions, or staffing issues.
Another limitation identified was the amount of available data. The Dutch railway dataset had significantly fewer data points per month compared to the US air data, potentially hindering the models’ ability to identify relevant patterns. The study concludes that while the effort to adapt the framework was substantial, it highlights the need for more context-specific adaptations and the incorporation of a wider range of features for accurate delay prediction in complex railway networks.
Also Read:
- Advancing Flight Delay Prediction with a Transferable Deep Learning Framework
- Smart Traffic Forecasting: A New Approach to Federated Learning
Future Outlook
This research, detailed further in the full paper available at arXiv.org, contributes to understanding the challenges of applying machine learning to transportation networks. Future work should explore integrating operational, environmental, and temporal variables, alongside more advanced machine learning techniques, to develop more robust and accurate models tailored to the multifaceted nature of railway operations. The goal remains to enhance the efficiency and reliability of railway systems for passengers.
