Short-term Prediction of Freeway Travel Times Using Data from Bluetooth Detectors
There is increasing recognition among travelers, transportation professionals, and decision makers of the importance of the reliability of transportation facilities. An important step towards improving system reliability is developing methods that can be used in practice to predict freeway travel times for the near future (e.g. 5 – 15 minutes). Reliable and accurate predictions of future travel times can be used by travelers to make better decisions and by system operators to engage in pre-active rather than reactive system management. Recent advances in wireless communications and the proliferation of personal devices that communicate wirelessly using the Bluetooth protocol have resulted in the development of a Bluetooth traffic monitoring system. This system is becoming increasingly popular for collecting vehicle travel time data in real-time, mainly because it has the following advantages over other technologies: (1) measuring travel time directly; (2) anonymous detection; (3) weatherproof; and (4) cost-effectiveness. The data collected from Bluetooth detectors are similar to data collected from Automatic Vehicle Identification (AVI) systems using dedicated transponders (e.g. such as electronic toll tags), and therefore using these data for travel time prediction faces some of the same challenges as using AVI measurements, namely: (1) determining the optimal spacing between detectors; (2) dynamic outlier detection and travel time estimation must be able to respond quickly to rapid travel time changes; and (3) a time lag exists between the time when vehicles enter the segment and the time that their travel time can be measured (i.e. when the vehicle exits the monitored segment). In this thesis, a generalized model was proposed to determine the optimal average spacing of Bluetooth detector deployments on urban freeways as a function of the length of the route for which travel times are to be estimated; a traffic flow filtering model was proposed to be applied as an enhancement to existing data-driven outlier detection algorithms as a mechanism to improve outlier detection performance; a short-term prediction model combining outlier filtering algorithm with Kalman filter was proposed for predicting near future freeway travel times using Bluetooth data with special attention to the time lag problem. The results of this thesis indicate that the optimal detector spacing ranges from 2km for routes of 4km in length to 5km for routes of 20km in length; the proposed filtering model is able to solve the problem of tracking sudden changes in travel times and enhance the performance of the data-driven outlier detection algorithms; the proposed short-term prediction model significantly improves the accuracy of travel time prediction for 5, 10 and 15 minutes prediction horizon under both free flow and non-free flow traffic states. The mean absolute relative errors (MARE) are improved by 8.8% to 30.6% under free flow traffic conditions, and 7.5% to 49.9% under non-free flow traffic conditions. The 90th percentile errors and standard deviation of the prediction errors are also improved.