Using Traffic Flow Data to Predict Bus Travel Time Variability Through an Enhanced Artificial Neural Network

This paper aims at predicting bus travel time and its day-to-day variability using a range of independent variables including traffic flow data. Among many factors impacting bus travel time, existing prediction approaches have not considered a traffic measure making their predictions unresponsive to the time dependent fluctuations in traffic flow and dynamic changes in traffic congestion. In addition, existing methodologies have mainly predicted the average travel time by finding the average value in a range of travel times likely to happen when a certain set of input values is considered. However, little attention has been given to predict the spread of that range created by the stochasticity of determinant factors which reflects travel time variability. This paper explains how an Artificial Neural Network (ANN) can be modified to predict the variance of a dependent variable. An integrated framework is then proposed which consists of two ANNs to predict both the average and variance of travel times. The proposed framework is developed on GPS based travel time data for a bus route in Melbourne, Australia, traffic flow data collected by the Sydney Coordinated Adaptive Traffic Systems (SCATS) loop detectors, and a measure of schedule adherence. The results demonstrate the value of traffic flow data in the prediction of bus travel time as well as the ability of the proposed method to provide fairly robust prediction intervals. INTRODUCTION Accurate predictions of travel time help operators in real time management strategies such as holding and expressing (Osuna & Newell, 1972, Fu & Yang, 2002), and off-line planning including fleet size planning and schedule design (Ceder, 2007). This information also enables passengers to better select departure times to minimize their waiting times. Information about the variability in travel times also benefits operators and passengers. It assists operators in defining optimal slack times to maximize the on-time arrival performance of buses (Kimpel et al., 2004), and in determining the reliability of systems (Turochy & Smith, 2002). A reduction in travel time variability reduces passengers’ anxiety caused by uncertainty in decision making about departure time and route choice (Bates et al., 2001, Lam & Small, 2001), which is why it is found as valuable (Sun et al., 2003), or even more valuable than a reduction in travel time (Bates et al., 2001). Mazloumi E., Currie G. and Rose G. 1 Previous studies have identified a range of determinants of bus travel time primarily through examining data from Advanced Public Transportation Systems (APTS) such as Global Position Systems (GPS), Automatic Vehicle Location (AVL) and Automatic Passenger Counting (APC) systems. They include traffic flow at intersections (Abdelfattah & Khan, 1998, Chien et al., 2002), passenger demand at stops (Shalaby & Farhan, 2004), traffic accidents (Abdelfattah & Khan, 1998), weather conditions (Hofmann & O'Mahony, 2005), different bus and driver characteristics (Mishalani et al., 2008, Strathman & Hopper, 1993), route characteristics (Abkowitz & Engelstein, 1983, Ng & Brah, 1998), and the effect of drivers’ timetable compliance on travel time (Lin & Bertini, 2004, Chen et al., 2005, Mazloumi et al., 2008). Predicting transit arrival/travel times has been the focus for many existing studies. Table 1 summarizes existing studies with respect to their adopted methodologies. As seen, existing methodologies can be generally grouped into four categories: Regression models, Kalman filter models, Artificial Neural Network (ANN) models, and Analytical approaches. Table 1 also shows the variables utilized to predict bus arrival/travel time as well as the data source used. Only two studies (Abdelfattah & Khan, 1998, Chien et al., 2002) have used traffic measures in their predictions. However, in both cases, the traffic data were sourced from simulation modelling. No previous study has adopted a real world traffic flow measure to predict bus travel time; therefore, their predictions may be unresponsive to the time dependent variations in traffic flow and dynamic changes in traffic congestion. Due to the stochasticity of bus travel time determinants, there may be a range of values corresponding to a certain set of input values. However, existing studies have primarily predicted the average in that range, and no research has been directed to quantify the spread of that range by predicting a measure of variability in travel times. This paper focuses on day-to-day variability which is the variation between the travel times of similar journeys made at the same time on different days. This paper proposes an integrated framework to predict bus travel time and its variability based on a range of independent variables including traffic flow data collected by the Sydney Coordinated Adaptive Traffic Systems (SCATS) loop detectors. The results are then used to construct a prediction interval corresponding to each input value set. Next section explains the proposed predictive framework, and its application in a case study is then presented. A closing summary, conclusions, and future research directions are included in the final section of the paper. Mazloumi E., Currie G. and Rose G. 3 Table 1: The explanatory variables used in different studies with respect to the adopted methodology. Study Route characteristics Passenger demand/dwell time Temporal variables/ scheduling Traffic measures Schedule adherence Bus progress data Historical travel times/speed/ trajectory weather Data source Regression models Abdelfattah and Khan (1998)    Simulation Patnaik et al. (2004)*    APC Artificial neural network models Kalaputapu and Demetsky (1995)   AVL Jeong and Rilett (2004)    AVL Park et al. (2004)*   GPS Chen et al. (2007)    APC Kalman filter models Chien et al. (2002)    Simulation Shalaby and Farhan (2004)   AVL-APC Chen et al. (2004)    APC Dailey et al. (2001)   AVL Chen et al. (2005)   AVL-APC