Hybrid Traffic Speed Modeling and Prediction Using Real-World Data

Traffic speed modeling and prediction is of great importance for both individuals and government authorities due to the increasing number of traffic congestion and corresponding social and economic impacts. Various approaches have been proposed to predict traffic speed. However, the long-term prediction accuracy is still unsatisfactory especially when occasional events such as extreme weather conditions occur. Based on over 880,000 traces of taxicab GPS collected in Hangzhou, China, as well as the dataset of weather conditions and holidays, we demonstrate a multi-time-scale correlation of traffic speed and the effects of various related events. We propose a hybrid traffic speed modeling and prediction framework which takes multi-time-scale historical traffic speed data as well as related events as inputs. For all segments of major roads in Hangzhou, we establish corresponding traffic speed models through a recursive model identification algorithm. We validate the effectiveness of our approach under various conditions through extensive trace-driven simulations.