Spatial-temporal traffic performance collaborative forecast in urban road network based on dynamic factor model

Many urban road networks today are experiencing increasing congestion that threatens not only transport efficiency but also living environment. To solve these problems, providing proactive knowledge of traffic performance is of significant importance. However, due to the inherent uncertainties of the signalized urban road network, it is a challenging work and some gaps still exist. First, the existing approaches are usually limited to a single location or region. Second, traffic flow parameters such as volume are used as a proxy of traffic state. Aiming to fill these gaps, this study developed a dynamic factor model-based approach to forecast the multi-step network traffic states of a group of regions in urban road network collaboratively. The novel model decomposes a set of traffic state time series into two orthogonal components: the common latent factor and the idiosyncratic disturbance. The common latent factor drives the co-movement dynamics of network traffic states, while the idiosyncratic disturbance captures the region-specific distinctions of traffic states in different regions. By extracting the principal variations of traffic states in a group of regions, the proposed model reduces the dimensionality of the traffic state variable from high-dimensional original space to low-dimensional latent factor space. By means of the common latent factor and its evolution over time, spatial-temporal correlations of traffic states in different regions and different time slots are seamlessly incorporated. The proposed model exhibits four distinct advantages, (1) it collaboratively produces forecasts for a group of regions; (2) it considers both comovement dynamics and region-specific distinctions of traffic state; (3) the model incorporates both spatial and temporal correlations seamlessly, and (4) it reduces dimensionality from a network-wide high-dimensional space to a low-dimensional latent factors space. The proposed model is applied to the real urban road network in Shanghai, China, based on the large-scale traffic performance index data released by the Shanghai Transportation Big Data Joint Innovation Laboratory. Empirical results from extensive experiments demonstrate the proposed dynamic factor model provides a promising approach for multi-step network traffic state forecast in urban road network, and outperforms the competing models considered in this study.