Modelling the arrival delay of train journeys with one transfer
2024 (English)Independent thesis Advanced level (degree of Master (Two Years)), 20 credits / 30 HE credits
Student thesis
Abstract [en]
Statistical modelling of the arrival delay of train journeys can help to inform train travellers and encourage them to travel by train more often. This thesis aims to develop a model for a train journey where the final destination cannot be reached directly and one transfer of trains is necessary. The dataset consists of train arrivals and departures for a journey between the Swedish cities Linköping and Växjö with a transfer in Alvesta. Introducing a transfer of trains adds additional complexity to the problem since a delay in the first train could cause travellers to miss the second train. In such a case they would need to resort to a later connecting train, which adds additional delay and the model needs to be able to account for this possibility.
Two different models are used to model this problem. Both are based on Bayesian inference using Markov chain Monte Carlo sampling (MCMC) to generate samples from the posterior distribution. The first model is the connection model, which consists of a series of sequential binary logistic regression models to decide which connecting train will be reached. The second model is a probabilistic model for the arrival delay of the connecting train in Växjö. A mixture distribution using 2 lognormal components has shown the best results for modelling train delays. The distribution parameters µ and σ are modelled as a function of predictors and coefficients, which increases the predictive power of the model. Some predictors, like the time of the day or the train line, show a significant influence on the arrival delay.
Overall, the suggested model is capable of modelling the unique, multimodal structure of the arrival delay of a train journey with one transfer. Posterior predictive checks are showing good results and the sampled outcomes seem to be well calibrated when compared with actual outcomes of previously unseen test data samples. Some of the logistic regression models show a high uncertainty in their estimates, due to a low sample size. These problems only appear in edge cases when both the first and second connections are missed and they could be alleviated with more training data. The model developed in this thesis could be expanded to other journeys and its outcomes could be used in customer-facing applications, to provide travellers with more information about train connection reliability and the distribution of their expected arrival time.
Place, publisher, year, edition, pages
2024. , p. 49
Keywords [en]
train delay prediction, bayesian statistics, train delay, mixture model
National Category
Transport Systems and Logistics
Identifiers
URN: urn:nbn:se:liu:diva-206200ISRN: LIU-IDA/STAT-A--24/029--SEOAI: oai:DiVA.org:liu-206200DiVA, id: diva2:1888106
Subject / course
Statistics
Supervisors
Examiners
2024-08-122024-08-122024-09-06Bibliographically approved