Empirical study of daily link traffic volume forecasting based on a deep neural network

Abstract

Forecasting the daily link traffic volume is critical in transportation demand analysis in feasibility studies for planning transportation facilities. The high purchase and maintenance cost of commercial transport planning software poses a challenge for several underdeveloped and developing countries. Therefore, there is a need for cost-effective methodology to forecast link traffic volume. This study proposes a data-driven approach for modeling traffic assignment and employs a deep neural network to forecast daily link volume derived from transport planning software. The main idea is that link traffic volume is significantly associated with traffic network attributes (i.e., number of lanes, travel speed, lane capacity, and roadway type) and network flow attributes (i.e., number of shortest paths on the corresponding link and origin-destination travel demand). Therefore, a multi-layer perception model is developed to effectively capture the nonlinear relationship among the link traffic volume, traffic network attributes, and network flow attributes. A case study demonstrated that the proposed method achieves comparable performance to commercial software in forecasting long-term link traffic volume. The obtained results indicated that the proposed method has the potential to serve as an alternative to commercialized software, although further studies are required to validate and enhance its application.