Forecasting Network Data

Abstract

Annual average daily traffic (AADT) values have long played an important role in transportation design, operations, planning, and policy making. However, AADT values are almost always rough estimates that are based on the closest short-period traffic counts and are factored by using permanent automatic traffic recorder data. This study develops Kriging-based methods for mining network and count data over time and space. With the use of Texas highway count data, the method forecasts AADT values at locations where no traffic detectors are present. While low-volume road counts remain difficult to predict, available explanatory variables are few, and extremely high-count outlier sites skew predictions in the data set used here, overall AADT-weighted median prediction error is 31% (across all Texas network sites). Here, Kriging performed far better than other options for spatial extrapolation, such as assigning AADT on the basis of a point's nearest sampling site, which yields errors of 80%. Beyond AADT estimation, Kriging is a promising way to explore spatial relationships across a wide variety of data sets, including, for example, pavement conditions, traffic speeds, population densities, land values, household incomes, and trip generation rates. Further refinements, including spatial autocorrelation functions based on network (rather than Euclidean) distances and inclusion of far more explanatory variables are possible, and will further enhance estimation.