Short-Term Traffic Flow Prediction with ACD and Particle Filter

Urban road traffic is a non-linear, non-stationary, and uncertain process; its uncertainty increases rapidly when making short-term (five minute-long) traffic flow prediction. The state-of-the-art in traffic flow prediction is reviewed here, including Time-Series, Artificial Neural Network, Ada-Boost, and Support Vector Machine based algorithms. A novel prediction method is proposed for short-term (five minute-long) traffic flow prediction. The time interval between vehicles is treated as a stochastic variable and described with the Marked Point Process in the ARCH (autoregressive conditional heteroskedasticity) framework. Different point processes may generate a corresponding ACD (autoregressive conditional duration) model for the prediction of time intervals between vehicles in the traffic's flow. A particle filter is applied with measured vehicle speed data for traffic flow speed and density prediction. We apply this algorithm to the Traffic Guidance System of the Cross-River Tunnel, Lujiazui, Pudong, Shanghai. Nine congestion levels are proposed and the prediction error for short time (five minute-long) is within three levels.