PROJECT TITLE :

Long-Term Urban Traffic Speed Prediction With Deep Learning on Graphs

ABSTRACT:

The ability to predict the speed of traffic is one of the fundamentals of advanced traffic management, and the gradual deployment of sensors connected to the internet of things is empowering data-driven approaches to the prediction of traffic speeds. However, the majority of the currently conducted research studies concentrate on short-term traffic prediction, which covers up to an hour's worth of forecast into the future. In the past, methods of long-term prediction have suffered from error accumulation, exposure bias, and the generation of future data with low granularity. In this article, a novel data-driven, long-term, and high-granularity approach to predicting traffic speed is proposed. The methodology is based on recent developments in techniques for graph deep learning. The spatial-temporal data correlation of traffic dynamics is incorporated into the prediction process by the use of a predictor-regularizer architecture, which is utilized by the model that was proposed. In both sub-networks, graph convolutions have become increasingly popular as a method for the extraction and reconstruction of geometrical latent information. In order to evaluate how well the proposed method works, exhaustive case studies are carried out on actual datasets from the real world, and it is found that consistent improvements can be seen in comparison to baselines. This work is among the pioneering efforts made toward predicting the network-wide long-term traffic speed. Future transportation research that makes use of deep learning could use the design principles of the proposed approach as a point of reference.