Incorporating congestion patterns into spatio-temporal deep learning algorithms

Abstract

Efficient machine learning algorithms capturing both spatial and temporal dependencies offer promising results for many fields including traffic prediction. However, their performance is still limited, particularly in congested areas where the speed on a road varies significantly. This paper exploits spatial (congestion) dependencies in road networks to improve the performance of spatio-temporal deep learning algorithms. Essentially, we extract spatial information relevant to how congestion develops and evolves in the network and enhance deep learning algorithms with it by developing custom prediction models for various network components. This research identifies traffic patterns through graph theory and traffic flow fundamentals, integrates them with deep learning algorithms for traffic prediction purposes, and enables better predictions in critically congested areas of road networks. The case studies with New York and Amsterdam networks show promising results; the proposed enhanced models significantly outperform the original deep learning models that consider the whole network as the prediction domain.