Negative Binomial Additive Models for Short-Term Traffic Flow Forecasting in Urban Areas

Abstract

Parallel, coordinated, and network-wide traffic management requires accurate and efficient traffic forecasting models to support online, real-time, and proactive dynamic control. Forecast accuracy is impacted by a critical characteristic of traffic flow, i.e., overdispersion. Efficiency depends on the time complexity of forecasting algorithms. Therefore, this paper proposes a novel spatiotemporal multivariate forecasting model that is based on the negative binomial additive models (NBAMs). Negative binomial is utilized to handle overdispersion, and additive models are used to efficiently smooth nonlinear spatial and temporal variables. To evaluate the model, it is applied to real-world data collected from Taipei City and compared with other forecasting models. The results indicate that the proposed model is an accurate and efficient approach in forecasting traffic flow in urban context where flow is overdispersed, autocorrelated, and influenced by upstream and downstream roads as well as the daily seasonal patterns, namely, low-, moderate-, and high-traffic seasons.