Traffic Speed Prediction for Urban Arterial Roads Using Deep Neural Networks

There has been a growing interest in the development of traffic speed prediction engines in recent decades, especially for urban arterial roads to aid in either the early detection of traffic incidents, augmentation of sparse probe data, or flagging of a malfunctioning infrastructure mounted sensor. Inspired by recent developments in data-driven analytics and machine learning, this paper presents a novel approach for predicting traffic speeds on urban arterial roads via pattern recognition and deep neural network modelling. The methodology developed in the current study is as follows: first, a pattern recognition system adaptively extracts inherent data trends and volatilities from historical traffic flow data. The resulting trend data together with historical temporal traffic flow data are used to prepare a representative training set for the deep learning model. In the current study, we adopt a modified empirical mode decomposition for multiscale pattern recognition. Improvements are also made in the deep learning architecture used to model the underlying relationships which influences traffic speed on urban arterials. A series of experiments are designed on U.S. 50 corridor in northern Virginia to evaluate the strengths and limitations of the methodology developed based on mean squared error (MSE) criterion.