Distributional prediction of short-term traffic using neural networks

Abstract

Neural network (NN)-based models have recently achieved outstanding results in short-term traffic prediction. However, most of these are based on the regression approach and trained to generate a single data point as a predicted value for future timesteps, which does not provide information on prediction uncertainty and limits its performance under different traffic conditions. To solve this problem, this study proposes a novel, high-dimensional distributional prediction (HDP) framework. This method has been validated by a series of experiments using the Caltrans Performance Measurement System dataset and four widely used NN models. The results suggest that the proposed HDP scheme can help existing NN structures to (1) generate adaptive distributional predictions for quantifying the uncertainty of multiple targets, and (2) gain better point prediction in terms of accuracy and robustness. Furthermore, we demonstrate that predicted speed distributions can be used for travel time estimation, outperforming other traditional methods in unexpected traffic conditions such as traffic incidents.