Modelling traffic management decisions using a hybrid machine learning and simulation approach

Abstract

Railway simulation models can be used to assess the robustness of timetables by subjecting the simulated traffic to minor disruptions and analysing their impact. For the output of such models to be meaningful, signallers' actions in the event of disruptions need to be represented with a reasonable level of accuracy. This paper presents a hybrid modelling approach that combines simulation software with a model for making traffic management decisions. The construction of the traffic management model is flexible, and this paper considers different approaches. Each approach takes a pair of trains as input and predicts which one will have priority at a conflict location. Traffic management models created using conditional logic are compared with machine learning models built using years of historical data. Results are presented using a case study of six conflict
locations: the models make predictions for a dataset of pairs of conflicting trains gathered over 90 days. The machine learning models demonstrate a higher level of agreement with the data than the programmatic models, although the gains for some conflict locations were more significant than others, indicating that each conflict location has its unique characteristics. The traffic management models were then integrated with the simulation software, and a week's worth of historical data was simulated. The machine learning approach for predicting
traffic management actions again showed better agreement with the real data.

More information

Identifiers

ORCID for Joanna Cameron Knight: orcid.org/0000-0002-2558-2815

ORCID for Andy Keane: orcid.org/0000-0001-7993-1569

ORCID for Ondrej Hovorka: orcid.org/0000-0002-6707-4325

Catalogue record

Export record