Deep learning-based traffic congestion forecasting research

Abstract

The explosive economic development of the world, the rise in urban population, the steady increase in the number of private cars on the road, the unequal distribution of traffic flow, and the localized congestion of the road network are all recent drivers to traffic congestion. As cities grow, traffic congestion has emerged as an unavoidable issue that threatens both the development of cities and resident mobility. In order to increase the utilization rate of urban infrastructure road facilities and more effectively relieve traffic congestion, management of traffic congestion first entails the accurate completion of the identification of road traffic status and the requirement to forecast traffic congestion in the city. This study suggests applying a deep spatial and temporal network model (DSGCN) to forecast how crowded the highways will be. The traffic network is initially divided into grids for the purposes of our study, with each grid standing for a different independent region. The grid region centroids are referred to as nodes in this study, and the dynamic correlations between the nodes are expressed by an adjacency matrix. Then, a two-layer long and short-term feature model (DSTM) is employed to capture the temporal correlation between regions, and a graph convolution-al neural network is used to capture the spatial correlation (Akhatov et al. 2021a; Rashidov et al. 2021). Studies on actual PeMS datasets show that the DSGCN outperforms other baseline models and is more accurate for predicting traffic congestion.