【作者】 姜山； 郭娅明； 李萌；

【Author】 Jiang Shan;Guo Yaming;Li Meng;Department of Civil Engineering,Tsinghua University;

【机构】 清华大学土木工程系；

【摘要】 随着手机移动应用与网约车平台的迅速发展,出租车出行逐渐成为城市生活中的一种重要的出行方式。当乘客想平台发送订单请求时,本文希望对本次旅程的旅行时间做初步的预测,以辅助平台进行调度决策,并给予驾驶人、乘客一个合理的出行时间预期。由于复杂的路网信息难以实时获取,因此仅依靠上下车地点等有限信息,快速进行计算的静态旅行时间预测是一种经济、合理的选择方式。本文基于著名的多头注意力机制模型,设计了适用于出租车静态旅行时间预测的一个自定义的Transformer网络模型,本文提出的模型应用多头注意力机制考虑异质的输入信息,在纽约市出租车数据集上能够给出比现存的较为流行的预测模型取得更好的预测结果,且具有更快的收敛速度。更多还原

【Abstract】 With the rapid development of mobile apps and ride-hailing platforms,taxi trip is becoming more and more popular in urban daily life.When customers send demand information to the platform,a primary trip duration prediction should be provided.One the one hand,predicted trip duration is important for the platform to make scheduling plans;one the other hand,predicted trip duration can provide a reasonable expectation for both drivers and customers.In real life,complete road information is hard to obtain,so static trip duration using only limited information,including pick-up location,destination location,etc,is a cost-efficient choice.This paper proposed a customized Transformer network designed for predicting taxi trip duration considering heterogeneity information,on the basis of the famous multi-head attention model.Experiments on2016 NYC taxi dataset illustrate that our proposed model could outperform current popular benchmark models in terms of both accuracy and rate of convergence.更多还原