Graph Attention Networks for 15-Min Traffic Volume Estimation on Urban Freeways Using Connected Vehicle Data

Abstract

Traffic flow is a critical factor that influences decision making in traffic operations and transportation planning. Current methods to collect traffic volume data predominantly involve the use of stationary sensors. However, owing to the high costs associated with the installation, operation, and maintenance of the sensors, they can only be installed on selected road segments. Consequently, estimating traffic volume in road segments without stationary sensors poses a significant challenge. This paper presents a graph attention network (GAT)-based algorithm to estimate the 15-min traffic volume on urban freeway segments without a stationary sensor using connected vehicle data. Specifically, a two-layer GAT was used to learn spatial relationships based on connected vehicle data and the difference in annual average daily traffic (AADT) between neighboring segments. Under identical input conditions, the proposed GAT method outperformed the baseline methods, including Graph Sample and Aggregate (GraphSAGE), AADT adjustment factor, linear regression (LR), K-nearest neighbors (KNN), and artificial neural networks (ANNs). All baseline models, except for the AADT adjustment factor method, incorporated connected vehicle data in their volume estimation processes. When tested across six urban areas in Iowa over 1 year, the mean absolute percentage error values for GAT, GraphSAGE, AADT adjustment factor, LR, KNN, and ANN were 13.2%, 15.7%, 22.3%, 24.6%, 24.0%, and 34.3%, respectively. Thus, the proposed method demonstrates the feasibility and reliability of using connected vehicle data to estimate traffic volume on urban freeway segments without stationary sensors.