Relating short-term traffic forecasting to current system state using nonparametric regression

Traffic management systems have been developed as an intelligent transportation systems application in major metropolitan areas to improve the safety and efficiency of the highway network. Monitoring of traffic conditions is a key function of these systems; short-term forecasting of traffic conditions is a potentially valuable function still under development. These systems typically archive extensive amounts of traffic data that can be mined to provide bases for monitoring and forecasting applications. Multivariate statistical quality control provides a means to describe the extent to which an observation deviates from a definition of normal, with respect to historical means and variances. A promising set of non-parametric regression techniques that use the nearest neighbor concept to locate past observations that are similar to current conditions for use in short-term forecasting has been developed. Information on deviation from normal of current and historical observations can be used to enhance these forecasting procedures. This paper documents several approaches to linking condition monitoring with forecasting, such as in the nearest neighbor selection process, in the determination of proximity of neighbors to the current observation, and in their relative contributions to the forecasts.