A joint temporal-spatial ensemble model for short-term traffic prediction.

Abstract

In this paper, we address the problem of short-term traffic flow prediction since accurate prediction of short-term traffic flow facilitates timely traffic manage- ment and rapid response. We advocate deep machine learning approach and propose a novel ensemble model, named ALLSCP, that considers both tempo- ral and spatial characteristics of traffic conditions. Specifically, we consider (1) short-, medium- and long-term temporal traffic evolution, (2) global and local spatial traffic patterns and (3) the correlation of temporal-spatial features in our predictions. We use real-world traffic data from two locations (i.e., Los Angeles and London) with frequent fluctuations (due to proneness to traffic accidents and/or congestion) to train and test our model. For each location, we consider road segments with and without junctions (i.e., linear vs intersection). We com- pare our model against well-known existing machine/deep learning prediction models. Our results indicate that our ALLSCP model consistently achieves the most accurate predictions (≈ 96% accuracy both on linear and intersec- tion roadways) when compared against existing models in the literature. In addition, we conducted ablation experiments to further gain insights into the contributions of individual constituent models of our ensemble ALLSCP model. Our results indicate that ALLSCP achieves the best results and is also robust against emergent traffic situations.