PREDICTING DEMAND-RESPONSIVE TRANSIT SCHEDULE VARIATIONS USING NEURAL NETWORKS

Abstract

Scheduling and dispatching procedures in transit systems have been increasingly automated, resulting in higher efficiencies and reliability. However, automated systems often lack the flexibility to adapt to external changes. The problem becomes more acute with transit systems that provide demand responsive service. The project described in this thesis considered the use of an artificial neural network to assist in determining the impact of external factors on the demand service scheduling system operated by the Saskatchewan Abilities Council. The study focused on two primary external variables, the effect of weather, and the time and date. The neural network was designed to provide an indication of expected schedule deviation for the transit system. The components of an acquisition system that would acquire and process data from a weather station and from either real time or stored bus data were determined. The acquisition system's main components were developed, consisting primarily of processing software that extracted and formatted weather and bus data for a neural network. An initial simulation showed that the neural network approach was viable. Additional simulations, using training data based on representative conditions, indicated that the neural network learned well and predicted schedule impacts on previously unseen data with a mean absolute error of less than 2%. The neural network approach has been shown to be very promising for the application of determining weather and time impacts on transit schedules.