Non-parametric Bayesian network to forecast railway disruption lengths

More Info
expand_more

Abstract

The length of a disruption in a railway network is highly uncertain which complicates the incident management of traffic controllers. This paper proposes a probabilistic model based on historical data to provide the prediction of the disruption length to the Dutch Operational Control Centre Rail (OCCR). A good prediction of disruption length is believed to help the OCCR to implement an appropriate response that minimizes the impact of the disruption for the railway users. The model that is proposed in this paper is a Non-Parametric Bayesian Network (NPBN) which represents the joint distribution between variables that describe the nature of the disruption. To obtain the prediction of the disruption length, this joint distribution is conditionalized on the particular values of variables in the model that are describing the situation at hand. The NPBN allows rapid conditionalization/inference which is attractive for the real-time decision making process of the OCCR. This paper presents the first attempt to model disruption length with NPBNs. A case study concerning a specific type of railway disruption, namely malfunctioning train detection, is considered as an example of the application of the method.