01663958

Component

Human dispatchers make freight rail dispatch decisions in real-time based on factors such as network traffic, network topology, and train characteristics. These decisions have significant influence on train delays and rail capacity, which motivates the development of tools to predict their effects. This article presents a machine learning framework to predict the occurrence of delay-inducing meets at sidings using an encoding of network state that incorporates information available to dispatchers at the time of prediction. Support vector classifiers (SVCs) are trained and predictions are compared to a simple deterministic baseline technique that uses only location information and treats trains equally. Testing is performed using historical data from a rail network in Tennessee, USA. Preliminary findings indicate that SVCs are able to exploit critical information beyond just train locations that is present in the network state to predict the occurrence of delays at sidings. The SVCs far outperform the baseline technique to which the authors compare and show that factors such as train length, train priority, and track occupancy have significant influence on delay occurrence.

This paper was sponsored by TRB committee AR030 Standing Committee on Railroad Operating Technologies.

18-05352

English

17p

2018

Transportation Research Board 97th Annual Meeting

Digital/other

Figures; References; Tables

Automatic train location; Car location (Railroads); Case studies; Delays; Dispatchers; Machine learning; Mathematical prediction; Railroad traffic; Real time information; Sidings (Railroads); Train operations

Support vectors

Chattanooga (Tennessee)

Operations and Traffic Management; Planning and Forecasting; Railroads; Vehicles and Equipment

Transportation Research Board Annual Meeting 2018 Paper #18-05352

TRIS, TRB, ATRI

Jan 8 2018 11:21AM