Golchehreh, Masoomeh (2026) Developing ensemble methods for predicting train delays in real-time. Doctoral thesis, University of East Anglia.
Preview |
PDF
Download (28MB) | Preview |
Abstract
Train delays are a common problem for railway systems in the United Kingdom and many other countries. They affect passengers, railway services, and daily operations. More accurate predictions can help railway operators prevent or reduce delays. Although many studies have sought to improve prediction accuracy, most have focused on individual models or simple ensemble methods and have relied on a limited number of evaluation metrics. Consequently, the accuracy and reliability of these models remain unsatisfactory under different conditions. In addition, despite the importance of real-time prediction, research in this area remains impractical.
To address these limitations, this research develops two ensemble methods for train delay prediction: the Multi-Metric Weighted Ensemble Method (MWEM) and the Multi-Metric Bayesian-Optimised Ensemble Method (BOEM). MWEM assigns model weights based on performance across multiple evaluation metrics and validation rounds, whereas BOEM uses Bayesian optimisation during each validation round to identify an effective combination of model weights. BOEM is then integrated into a software prototype for real-time delay prediction.
To evaluate the proposed methods, historical and live data from several UK railway routes were used. Statistical analysis showed that the proposed ensemble methods performed significantly better than the individual machine learning and deep learning models and the ensemble baselines. In the historical five-fold cross-validation experiments across both railway routes, BOEM reduced MAE by approximately 7% compared with MWEM and by approximately 8% compared with the strongest individual model. In the real-time evaluation, BOEM reduced MAE by approximately 11% compared with the available railway estimates and by approximately 9% in recursive downstream prediction. Further experiments showed that the proposed method remained more reliable under missing-value and cross-dataset conditions.
Overall, the findings show that the proposed multi-metric ensemble methods can improve train delay prediction across historical, real-time, and additional evaluation settings. This research, therefore, provides a practical framework for combining different prediction models and applying them to railway delay prediction.
| Item Type: | Thesis (Doctoral) |
|---|---|
| Faculty \ School: | Faculty of Science > School of Computing Sciences |
| Depositing User: | Chris White |
| Date Deposited: | 18 Nov 2026 09:19 |
| Last Modified: | 18 Nov 2026 09:19 |
| URI: | https://ueaeprints.uea.ac.uk/id/eprint/105137 |
| DOI: |
Downloads
Downloads per month over past year
Actions (login required)
![]() |
View Item |
Tools
Tools