Time Series Analysis and Models of Freeway Performance

Abstract

This paper describes a study whose primary purpose is to better understand the relationship between freeway traffic flow and speed. Incidents of recurrent and nonrecurrent congestion were encountered at six radar collection northbound locations on New Hampshire interstate I-93 in July 2010. The root cause for the onset of the recurrent congestion is explained with exploratory data analyses and a time series modeling approach. A complex combination of present and past values of traffic flow, speed, and congestion state, a congestion history of lingering effect variables, can explain the triggering and mitigation of congestion events for a highly volatile traffic environment. The approach includes two mathematical models: (1) a generalized additive binomial model to forecast the probability of congestion and (2) state-space models of speed and flow. The state-space models use a dynamic linear model (DLM) with switching structures to describe the bimodal distribution of speed and flow in the free-flow and congested states. Model selection, parameter estimation and checking are presented.