An Experimental Study on Real Time Bus Arrival Time Prediction with GPS Data

485152801999PDFResearch PaperBus transit operationsTimed transferArrivals and departuresMathematical predictionGlobal Positioning SystemTravel timeReal time informationAdvanced public transportation systemsAVL techniquesAdvanced Traveler Information SystemBusesMathematical modelsPublic transportationReal timeTravel timeVirginiaVirginia Polytechnic Institute and State UniversityLin, W. H.Zeng, J.Virginia Polytechnic Institute and State UniversityNTL-INTELLIGENT TRANSPORTATION SYSTEMS-INTELLIGENT TRANSPORTATION SYSTEMSNTL-INTELLIGENT TRANSPORTATION SYSTEMS-Traveler InformationNTL-INTELLIGENT TRANSPORTATION SYSTEMS-Transit ManagementNTL-GEOGRAPHIC INFORMATION SYSTEMS-GEOGRAPHIC INFORMATION SYSTEMSIntelligent Transportation Systems Joint Program OfficeBus headway in a rural area is usually much larger than that in an urban area. Providing real-time bus arrival information could make the public transit system more user-friendly and thus enhance its competitiveness among various transportation modes. As part of an operational test for rural traveler information systems currently ongoing in Blacksburg, Virginia, an experimental study has been conducted on forecasting the arrival time of the next bus with AVL techniques. This paper discusses the process of developing arrival time estimation algorithms, including route representation, GPS data screening for identifying data quality and delay patterns, algorithm formulation, and the development of measures of performance. Whereas GPS-based bus location data are adopted in all four algorithms presented in the paper, the extent to which other information is used in these algorithms varies. In addition to bus location data, information relevant to the performance of an algorithm also includes scheduled arrival time, delay correlation, and waiting time at time-check stops. The performance of an algorithm using different levels of information is compared against three criteria: overall precision, robustness, and stability. Our results show that at the site where the study is being conducted, the dwell time at time-check stops is most relevant to the performance of an algorithm.93