Statewide Truck Volume Estimation Using Probe Vehicle Data and Machine Learning

Abstract

Network-wide truck volume information is critical for monitoring, managing, and planning highway truck systems as well as the overall transportation system. However, availability of this information is often quite limited because classification counts are only collected at a few locations each year. This paper presents a statewide truck annual average daily traffic (AADT) estimation model using widely available truck probe data that are accessible to transportation agencies. Using Kentucky as a case study, an annual average daily truck probe (AADTP) metric was derived from truck probe data and found to be strongly associated with truck AADT (Pearson’s r = 0.9). Other important variables included in the model were roadway attributes (i.e., functional class, number of lanes, lane width), network centrality, and sociodemographic characteristics of the surrounding area. The final estimation model is a random forest model as it outperformed linear regression, ridge regression, neural network, support vector machine, and extreme gradient boost algorithm in this study. Estimation results show that median and mean absolute percent errors decrease as AADTP increases. For roadways whose AADTP is greater than 53, the median and mean absolute percent errors for estimated truck AADT drop to 20% and 30%, respectively. The model’s utility is demonstrated by generating a truck volume profile for Kentucky’s statewide freight network.