The travel time to proceed from one location to another in a network is an important consideration in many urban transportation settings ranging from the planning of delivery routes in freight transportation to the determination of shortest itineraries in advanced traveler information systems. Accordingly, accurate travel time predictions are of foremost importance. In an urban environment, vehicle speeds, and consequently travel times, can be highly variable due to congestion caused, for instance, by accidents or bad weather conditions. At another level, one also observes daily patterns (e.g., rush hours), weekly patterns (e.g., weekdays versus weekend), and seasonal patterns. Capturing these time-varying patterns when modeling travel speeds can provide an immediate benefit to commercial transportation companies that distribute goods, since it allows them to better optimize their routes and reduce their environmental footprint.

This paper presents the first part of a project aimed at optimizing time-dependent delivery routes in an urban setting. It focuses on the prediction of travel speeds using as input GPS traces of commercial vehicles collected over a significant period of time. The proposed algorithmic framework is made of a number of macro-steps where different machine learning and data mining methods are applied. Computational results are reported on real data to empirically demonstrate the accuracy of the obtained predictions.

在许多城市交通设置中，从计划从货运地点到另一地点的行进时间是一个重要的考虑因素，从货运运输的规划路线到高级旅客信息系统中最短行程的确定。因此，准确的行进时间预测至关重要。在城市环境中，由于拥堵（例如，事故或恶劣的天气状况）导致的车辆速度以及因此的行驶时间可能会发生很大变化。在另一个层次上，人们还观察每日模式（例如高峰时间），每周模式（例如工作日与周末）和季节性模式。在对行驶速度进行建模时捕获这些时变模式可以为分销货物的商业运输公司带来直接的收益，

本文介绍了该项目的第一部分，旨在优化城市环境中随时间变化的配送路线。它着重于使用在相当长的一段时间内收集到的商用车辆的GPS轨迹作为输入来预测行驶速度。所提出的算法框架由许多宏步骤组成，其中应用了不同的机器学习和数据挖掘方法。计算结果报告在真实数据上，以经验证明所获得预测的准确性。