QAMPO and DTU
Signature Rail’s vehicle optimizer strategic partner wins first place in Rail Problem Solving Competition
Over 40 teams participated in this year’s Rail Problem Competition, which seeks to predict delays in the Netherlands Railway passenger rail network. The three finalists used three very diverse methods. The titles, abstracts and contributors are below. They will be presenting on Sunday, November 4, 2018 at the INFORMS Phoenix conference for first, second and third place.
First Place: Forecasting Train Delays in the Netherlands using Neural Networks
Abstract: We investigate to what extent low-maintenance and out-of-the box machine learning models can provide accurate predictions of train delays. We focus on predicting actual delay and but also the delay development. The results on real-life data from the Netherlands indicate that our models can outperform a constant prediction model.
- Jørgen Thorlund Haahr (PhD), Decision Scientist, QAMPO
- Erik Hellsten, PhD candidate, Technical University of Denmark
- Evelien van der Hurk (PhD), Assistant Professor, Technical University of Denmark
Second Place: Predicting Near-Term Train Schedule Performance and Delay Using Bi-Level Random Forests
Abstract: Near-term train delays prediction is critical for railway management. We propose a bi-level random forest approach to predict train delays. The primary level predicts the delay category, and the secondary level estimates the delay (in minutes). The proposed model is compared with several alternative approaches, validating its superior accuracy.
- Mohammad Amin Nabian, PhD Candidate, Department of Civil and Environmental Engineering, University of Illinois at Urbana-Champaign,
- Negin Alemazkoor, PhD Candidate, Department of Civil and Environmental Engineering, University of Illinois at Urbana-Champaign
- Hadi Meidani, Assistant Professor, Department of Civil and Environmental Engineering, University of Illinois at Urbana-Champaign
Third Place: A Railway Delay Prediction Model Based on Non-Homogeneous Markov Chains
Abstract: Assuming the delay on a certain station depends only on delay attained in the previous station, we model the delay evolution over stations as a non-homogenous Markov chain. By discretizing the state space and retrieving transition matrices from historical data, we accurately predict delay using a probabilistic approach.
- Gao, Zheming firstname.lastname@example.org, Department of Operations Research, NC State University
- Luo, Haochen email@example.com, ISEN Department, Texas A&M University
- Wu, Qian firstname.lastname@example.org, ISEN Department, Texas A&M University
- Xu, Jin email@example.com, ISEN Department, Texas A&M University