Spatiotemporal Traffic State Prediction Based on Discriminatively Pre-trained Deep Neural Networks

Abstract

The availability of traffic data and computational advances now make it possible to build data-driven models that capture the evolution of the state of traffic along modeled stretches of road. These models are used for short-time prediction so that transportation facilities can be operated in an efficient way that guarantees a high level of service. In this paper, we adopted a state-of-the-art machine learning deep neural network and the divide-and-conquer approach to model large road stretches. The proposed approach is expected to be a tool used in daily routines to enhance proactive decision support systems. The proposed approach maintains spatiotemporal correlations between contiguous road segments and is suitable for practical applications because it divides the large prediction problem into a set of smaller overlapping problems. These smaller problems can be solved in a reasonable time using a medium configuration PC. The proposed approach was used to model 21.1- and 30.7-mile stretches of highway along I-15 and I-66, respectively. The resulting predictions were better than predictions obtained using partial least squares regression.