Abstract Short-term traffic speed forecasting is an important component of Intelligent Transportation Systems (ITS). Multi-step-ahead prediction can provide more information and predict the longer trend of traffic speed than single-step-ahead prediction. This paper presents a multi-step-ahead traffic speed prediction approach by improving the gradient boosting regression tree (GBRT). The traditional multiple output strategies, e.g., the direct strategy and iterated strategy, share a common feature that they model the samples through multi-input single-output mapping rather than multi-input multi-output mapping. This paper proposes multivariate GBRT to realize simultaneous multiple outputs by considering correlations of the outputs which have not been fully considered in the existing strategies. For illustrative purposes, traffic detection data are extracted at the 5-min aggregation time interval from three loop detectors in US101-N freeway through the Performance Measurement System (PeMS). The support vector regression (SVR) is used as the benchmark. 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 higher prediction accuracies compared with SVR; (II) GBRT using the iterated strategy has a good prediction accuracy in short-step-ahead prediction and the prediction accuracy decreases significantly in long-step-ahead prediction; (III) Multivariate GBRT has the best stability which means the higher reliability in multi-step-ahead prediction while iterated GBRT has the worst stability; and (IV) Multivariate GBRT has an enormous advantage in the prediction efficiency and this advantage will expand with the increasing prediction horizons.

中文翻译：

---

使用多输出梯度提升回归树的多步超前交通速度预测

摘要 短期交通速度预测是智能交通系统（ITS）的重要组成部分。与单步提前预测相比，多步提前预测可以提供更多信息并预测交通速度的更长趋势。本文提出了一种通过改进梯度提升回归树 (GBRT) 的多步超前交通速度预测方法。传统的多输出策略，例如直接策略和迭代策略，都有一个共同的特点，即它们通过多输入单输出映射而不是多输入多输出映射对样本进行建模。本文提出了多元GBRT，通过考虑现有策略中没有充分考虑的输出的相关性来实现同时多个输出。出于说明目的，通过性能测量系统 (PeMS) 从 US101-N 高速公路中的三个环路检测器以 5 分钟聚合时间间隔提取交通检测数据。支持向量回归 (SVR) 用作基准。对三个模型的评估基于三个标准，即预测准确度、预测稳定性和预测时间。结果表明(I)多变量GBRT和使用直接策略的GBRT与SVR相比具有更高的预测精度；（二）使用迭代策略的GBRT在短步预测中具有较好的预测精度，在长步预测中预测精度显着下降；(III)多变量GBRT稳定性最好，即多步超前预测的可靠性更高，而迭代GBRT的稳定性最差；