A Reliable Travel Time Prediction System With Sparsely Distributed Detectors

Thumbnail Image
umi-umd-4528.pdf(979.8 KB)
No. of downloads: 776
Publication or External Link
Zou, Nan
Chang, Gang-Len
This study aims to develop a travel time prediction system that needs only a small number of reliable traffic detectors to perform accurate real-time travel time predictions under recurrent traffic conditions. To ensure its effectiveness, the proposed system consists of three principle modules: travel time estimation module, travel time prediction module, and the missing data estimation module. The travel time estimation module with its specially designed hybrid structure is responsible for estimating travel times for traffic scenarios with or without sufficient field observations, and for supplying the estimated results to support the prediction module. The travel time prediction module is developed to take full advantage of various available information, including historical travel times, geometric features, and daily/weekly traffic patterns. It can effectively deal with various traffic patterns with its multiple embedded models, including the primary module of a multi-topology Neural Network model with a rule-based clustering function and the supplemental module of an enhanced k-Nearest Neighbor model. To contend with the missing data issue, which occurs frequently in any real-world system, this study incorporates a missing data estimation module in the travel time prediction system, which is based on the multiple imputation technique to estimate both the short- and long-term missing traffic data so as to avoid interrupting the operations. The system developed in this study has been implemented with data from 10 roadside detectors on a 25-mile stretch of I-70 eastbound, and its performance has been tested against actual travel time data collected by an independent evaluation team. Results of extensive evaluation have indicated that the developed system is capable of generating reliable prediction of travel times under various types of traffic conditions and outperforms both state-of-the-practice and state-of-the-art models in the literature. Its embedded missing data estimation models also top existing methods and are able to maintain the prediction system under a reliable state when one of its detectors at a key location experience the data missing rate from 20% to 100% during uncongested, congested and transition periods.