TrainChaiNN - A Train Delay Prediction Model Using Chained Adversarial Neural Networks

In addition to operational impacts, delays in transportation networks like railways adversely impact the journey plans of passengers. These have proven economic impact on society at large. To manage and mitigate delays, it is crucial for transportation operators to predict delays accurately in advance. Existing works for train delay prediction attempt to improve accuracy by using complex models, performing extensive feature engineering and leveraging domain experts. They also tend to focus on aggregate behaviour like station-wise and segment-wise delays. Predictions are most useful when done at trip-level and when cascade effects of delay are captured. In this work, we present TrainChaiNN, a novel chained adversarial neural network model. TrainChaiNN uses fewer features and is capable of predicting delays and the subsequent cascaded delays with high accuracy. We have benchmarked the model against three published models using nation-wide railway datasets from Belgium and the UK. The model predicts arrival and departure delays with low Mean Absolute Error (MAE) of 23 and 39 seconds for Belgium and UK datasets respectively. Applying predictions of TrainChainNN to multi-hop journey-planning more than halves delay-induced failures when compared to other models. It also outperforms other models for dealing with out-of-distribution data and for predicting rarely occurring major delays.