Model development for predicting short-term travel time based on GPS data

Abstract

This research aimed to develop models for short-term travel time prediction based on data from a GPS device. The experiment was conducted using a GPS device to track a bus from suburban Bangkok to Central Business District (CBD). GPS data were extracted and subsequently analyzed. Kalman filter algorithm was developed as a dynamic travel time prediction tool by utilizing inputs from GPS data from the three previous consecutive days. The prediction was carried out over subsections rather than over time periods. The mean absolute percentage error (MAPE) of seven-day travel time prediction between successive bus stops was varied from 7.14% to 104.07% for inbound trip and from 9.98% to 177.29% for outbound trip. It was noticed that the algorithm produced high prediction error at bus stops located near signalized intersections: Taling Chan area (merging traffic from other roads and road maintenance) as well as CBD. Finally, the web-based system adopted the current model to provide an estimated arrival time of buses. Based on the empirical analysis of prediction accuracy, the average error of arrival time prediction was -16 seconds with a standard deviation of 4 minutes for inbound and 1.27 minute with standard deviation of 7 minutes for outbound. The results of both travel time and arrival time prediction seemed to produce higher error among some particular bus stops. This is because the use of data from the three consecutive days as inputs cannot reflect the current traffic condition. It is expected that the accuracy of prediction will be improved when the real time information of the preceding buses are available and incorporated.