基于集成算法的路段短时行驶时间预测

Short term prediction of road travel time based on an ensemble algorithm

为了更好地解决路段行驶时间的短时预测问题,提出并改善了一种基于树的集成算法。针对小时间尺度下交通时变性强这一特性,构建更加鲁棒的梯度提升树(GBDT)以减少突变点的干扰。为了克服偏差-方差窘境,将随机树(RF)与GBDT进行融合,提出RF-GBDT的集成算法,并考虑各种历史旅行时间数据的相关变量以提高模型的可解释性。预测结果表明,与单独的RF或GBDT相比,RF-GBDT具有更好的预测准确度与算法稳定性。

In order to better solve the short term prediction problem of travel time on links,a tree based ensemble method was proposed and improved. First,a more robust GBDT was established to reduce the disturbance caused by bursts aiming at the strong upheavals of traffic in a small time scale. Then,the RF and GBDT were fused and a new method for RF-GBDT was proposed to overcome the problem of bias-variance dilemma. In addition,various relevant variables derived from historical travel time data were considered to improve the interpretability of the model. The results of predictions show that compared with the single RF or GBDT,the RF-GBDT method is preferable in the accuracy and the stability of algorithms.