Short-Term travel speed prediction for urban expressways using convolutional neural network and tensor decomposition

Abstract

Prediction is an important part of the traffic management system (TMS) which supports route planning, dynamic traffic control, and information provision. We developed a multi-dimensional learning machine for predicting the traffic speed. Proposed methodology considered both historical experience and near past observation of traffic data by combining a convolutional neural network (CNN) with tensor decomposition (TD) predictor. TD based method is treated as an effective traffic predictor considering temporal-spatial neighbourhood data. However, limited by learning mechanism, such a method cannot extract historical traffic pattern from large datasets. Our main contribution is converting the traffic prediction into a tensor imputation problem, which considers the historical pattern information from CNN by constructing input tensor. Besides, multiple low-rank choosing and weighted optimization are introduced to improve the accuracy of TD-based prediction. We validated our methodology using empirical detector data of urban expressway in Shanghai. Compared with single algorithms, our method has smaller absolute and relative error.