Multi-step-ahead traffic flow forecasting using multi-output gradient boosting regression tree

Abstract

Short-term traffic flow forecasting is an important component of Intelligent Transportation Systems (ITS). Multi-step-ahead prediction can provide more information than single-step-ahead prediction and predict the trend of traffic flow. However, there is less research for multi-step prediction in the field of traffic flow forecasting. This paper presents approaches for multi-step-ahead traffic flow prediction by improving gradient boosting regression tree (GBRT). As an ensemble learning algorithm, GBRT grows individual regression trees sequentially by the boosting method. The traditional multiple output strategies for the most algorithms are direct strategy and iterated strategy. Those two strategies for multi-step-ahead prediction share a common feature that they model the samples through a multi-input single-output mapping rather than multi-input multi-output mapping. This paper proposes multivariate GBRT to realize multiple outputs simultaneously by considering the correlations of the outputs which have not been considered in the existing strategies. For illustrative purposes, detection data were extracted at the 5-min aggregation time interval from three loop detectors in US101-N freeway between September 7 and October 20, 2016 through Performance Measurement System (PeMS). The first thirty-seven-day data are utilized as the training and validation set, and the remaining seven-day data are used as the test set. Three multi-step-ahead prediction methods for GBRT are comparatively tested using the real-world traffic data. The support vector regression (SVR) was used as a benchmark and the assessments on the three models are based on the three criteria, i.e., prediction accuracy, prediction stability, and prediction time. The results indicate that (I) Multivariate GBRT and GBRT using the direct strategy have good prediction accuracy compared with SVR; (II) GBRT using the iterated strategy have good prediction accuracy in short-step-ahead prediction and have bad prediction accuracy in long-step-ahead prediction; (III) Multivariate GBRT has the best stability which means it has more reliability in multi-step-ahead prediction while iterated GBRT has the worst stability; (IV) Multivariate GBRT has an enormous advantage in the prediction efficiency and this advantage expands with the increasing prediction horizons.