Shared mobility refers to the shared use of a motor vehicle, bicycle, or other low-speed transportation modes, which has significant influences on transportation systems. Among different types of shared mobility, ridesplitting is developing the most rapidly because of the emergence of transportation network companies. In this study, we present exploratory evidence of on-demand ridesplitting's impacts on the vehicle use and purchase willingness using the emerging ridesourcing real-world data (i.e., app-based, on-demand ride services like Uber and DiDi Chuxing) in Hangzhou, China. We explore ridesplitting users' travel habits and their transportation modal shift behavior if the ridesplitting services are unavailable, and how ridesplitting impacts the use of public transit and private cars. In summer 2017, an online survey was conducted with respondents who had completed ridesourcing journeys within one month before the investigation. We compare the survey results with two-week ridesourcing order data provided by DiDi Chuxing, 20% of which are ridesplitting orders (including DiDi Hitch and Express ridesplitting). Considering the modal shift from public transit to ridesplitting, the findings indicate that (I). In the short term, ridesplitting services reduce the number of vehicles on road (decreased by 3,051 veh/day, nearly 2.6 parts per thousand of the vehicle ownership in the urban area of Hangzhou). (II). In the intermediate term with the development of ridesplitting, the total decreased number of vehicles is estimated to be 4,129 veh/day (nearly 3.6 parts per thousand of the vehicle ownership). Au interesting phenomenon is that Hitch's influences on the vehicle usage in the intermediate term is not obvious as the influence in the short term due to travelers' modal shift from public transit. (III). In the long term, ridesplitting will reduce the car purchase willingness and influence people's travel behavior to some extent. This paper shines light on quantitatively exploring the influence of on-demand ridesplitting on the vehicle use and purchase willingness.

交通出行是人们生活的重要组成部分,伴随着城市化水平的不断提升,城市交通问题愈发严重,城市交通治理越来越受到关注。互联网时代催生了许多出行方式,如网约出行、共享出行等,给个性化出行带来了便利,但也给交通治理带来新的问题,如何调控大规模城市路网的交通供需关系成为了关键科学问题。交通仿真作为重要的交通分析工具,具备细颗粒度、高效能、低成本等优势,可以帮助决策者对交通策略进行模拟测试,为决策优化提供技术支撑。本文面向移动互联交通大数据环境,提出网约出行多智能体交通仿真模型,自主研发了整套网约出行仿真系统,应用于多种复杂的网约出行场景模拟仿真与策略优化。本文的主要内容与研究成果如下:（1）基于多智能体的网约出行仿真框架。提出基于智能体的仿真框架,构建司机与乘客两种智能体行为模型,利用交通大数据与深度学习建立数据驱动智能体行为决策模型,依托开源地图数据构建仿真路网,实现车辆路径规划、在途导航、巡游与人车匹配等仿真流程。（2）基于开源Go语言自主研发了智能体仿真系统,实现了后端仿真、前端可视化展示等功能,从乘客、司机、平台、社会等多角度进行仿真结果统计与可视化演示。经测试,其在单机（CPU i7-4790@3.6 GHz）上仿真全杭州单日13万次出行仅需14分钟,仿真加速比达103倍。（3）基于多智能体仿真系统的网约出行平台定价研究。在多智能体仿真系统的基础上,构建模型刻画平台定价与市场供需间的关系,以进行平台定价研究。结果表明,存在特定的出行价格使得平台收益最大化,此时倾向提高订单价格,降低司机收入。而社会效用最大化的定价方案维持在一个条状区域带内。（4）有人驾驶与无人驾驶网约出行混合仿真。将网约出行拓展至有人驾驶（Human-driving Vehicle,HV）与无人驾驶（Automated Vehicle,AV）混合运营场景领域,仿真结果表明AV与HV可以进行互补,少量的AV（约10%）即可大幅降低乘客的等待时间,随着AV占比增加,网约车的总行驶里程与尾气排放也不断减少,在纯AV场景下的尾气排放比HV场景下降12.3%。综上所述,本文提出基于智能体的网约出行仿真模型框架并开发了完整的仿真系统,将其应用到网约出行定价优化与有人驾驶无人驾驶网约出行混合运营等多种仿真场景测试中,该系统具备良好的可扩展性与适应性,可以作为网约出行研究的重要工具。

随着我国社会经济的不断发展,机动车保有量逐年上升,由此产生了诸如交通拥堵等一系列交通问题,这给出行者和相关管理部门带来了极大不便和困扰。智能交通系统是缓解交通拥堵行之有效的手段,交通流预测是智能交通系统实现的基础和关键所在,利用海量交通大数据,实现城市级复杂路网的道路交通流预测,具有重要的理论和现实意义。本文提出面向大规模城市路网的短时交通流预测模型,预测未来一段时间内城市交通流演化态势,为出行者提供准确的出行信息,为管理者进行主动交通管控提供理论依据。主要研究内容和研究成果体现在以下方面:（1）短时交通流预测理论和方法分析。探讨交通流基本参数及其特性,有针对性地提出交通流预测模型所应当具备的一系列能力,并通过深度学习基础理论分析,为交通流预测模型的建立提供指导依据。（2）城市快速路交通流预测。筛选重要交通流参数,分析其时空相关性,提出基于卷积神经网络的城市快速路交通流预测模型,实现在预测过程中学习各种影响因素的时空相关性以及多种交通流参数协同预测,采用真实快速路线圈数据进行模型的训练和测试。（3）考虑时空相关性的城市路网交通流预测。构建了深度可自由变化的交通流深度时空张量,提出了用于路网交通流预测的时空深度张量神经网络预测模型（Spatial-Temporal Deep Tensor Neural Networks,ST-DTNN）,消除了 由交通流时间序列的随机堆叠引起的潜在负面影响,设计了参数更新算法供模型训练学习。采用真实大规模城市路网浮动车数据对ST-DTNN模型进行了测试,实验结果表明,ST-DTNN模型相对于传统的时间序列和机器学习等基准模型,在预测性能上有较大提升。（4）考虑路网拓扑结构的路网交通流预测模型。构建了基于图卷积网络（Graph Convolution Network,GCN）的交通流预测模型,同时考虑了路网拓扑结构和交通流时空相关性,提高了模型预测精度和现实解释意义。与ST-DTNN模型相比,GCN模型在预测精度上有一定提升,同时训练时间缩短了近35%,模型更有计算效率。图卷积网络模型在ST-DTNN基础上改进了城市级路网交通流预测框架,其高效和高精度特性为交通管控和路径规划场景应用奠定了理论基础。

随着“互联网+”战略与传统交通行业的深度融合,各种各样的即时出行需求服务平台应运而生,尤其是在交通供需矛盾突出的北京、上海、广州、深圳、杭州等城市迅速发展,由此产生了一种新的网约共享出行模式。共享出行用户能够按需获得交通工具的短期使用权,而无需拥有交通工具的所有权。网约共享出行系统分析与优化调度相关问题在最近几年引起学术界的广泛关注,然而学术界对共享出行系统运行机制的了解仍相当有限。例如,目前尚不明确共享出行对交通系统的影响,作为独立决策的参与者认知和行为如何决定共享出行意愿,以及如何建立高效的共享出行调度模型等。鉴于此,本文在研究网约共享出行行为机理及其外部影响基础上,构建了基于增强学习的网约共享出行优化调度模型,以期实现网约共享出行服务系统供需平衡,有助于优化城市交通结构和提高道路资源使用效率。本文主要包括以下三方面的研究:（1）网约拼车出行行为机理分析及其影响评估。采用SP调查和RP调查相结合的方法,获取网约共享用户的通勤出行方式分担率、出行次数、其他出行方式向网约共享出行方式的转移量等数据集,分析网约共享通勤出行方式选择的影响因素。利用从网约共享出行平台中提取的真实网约拼车出行数据和对网约拼车出行用户进行的问卷调查,定量评估了网约拼车享出行行为对城市交通系统的影响。（2）基于网约共享出行订单数据的路网行程时间可靠性分析。本文建立了路网行程时间可靠性评价指标体系,考虑了路网中OD对的行程时间率概率分布,并用中国最大的网约共享出行平台的真实数据进行实例分析。路网行程时间可靠性指标能够较好地评价区域和城市交通系统运行效率,为出行者的行程的提供意见和指导,且适用于不同层次的用户,包括出行者和交通管理人员,有助于调整个人出行策略,提高交通运营效率。（3）基于蒙特卡洛树搜索的网约共享出行平台派单优化。突破以往只从空间角度优化派单问题,本文从时空两个维度扩展可用车辆的集合,可提高司乘匹配率,缩短司乘等待时间。从多角度建立司乘调度模型,并根据司乘匹配关系,建立匹配树结构,再利用改进后的蒙特卡洛树搜索算法对模型进行求解。结果表明,所提出的派单问题优化模型和增强学习算法可以实现高效派单,特别是可以在网约共享出行平台供需不平衡的情况下提供高效可靠的派单方案。

This paper proposes a dynamic ride-hailing sharing problem with multiple vehicle types and user classes. Ride-hailing vehicles (RHVs) can be classified into express ride-hailing vehicles (ERHVs) and premier ride-hailing vehicles (PRHVs) according to service levels. PRHVs can provide the high-quality ride-hailing service with upmarket vehicles and ERHVs provide the normal ride hailing service with normal vehicles. The fare of PRHVs is higher. PRHVs can be temporarily used as ERHVs to serve the customers who order ERHVs with or without ride-sharing, which referred to as the substitution of ERHVs with PRHVs. A lexicographic multi-objective function with three-level objectives is proposed to formulate the problem, in which the first-level objective is to maximize the profit of the platform, the second-level objective is to minimize the number of requests of customers who involve ordering ERHVs matched to PRHVs, and the third-level objective is to minimize the total driving distance of all RHVs. The dynamic problem is divided into a set of continuous and small ride-hailing sharing subproblems based on equal time intervals. Each subproblem is formulated as a mixed integer nonlinear program for matching RHVs to the requests collected in the last time interval or unmatched in previous time intervals and re-scheduling the vehicle routes. To solve the subproblems, a new solution method is proposed based on the modified artificial bee colony algorithm developed by Zhan et al. (2021). Numerical examples using real request data from Didi are given to explore the problem properties, and the results gain insights into the ride-hailing market. For example, the profit of the platform and the number of matched requests are higher when the substitution of ERHVs with PRHVs is allowed while the matching percentage of requests of customers who select a mixed choice is higher when there is no substitution. Different ratios of vehicle types and user classes influence the perfor-mance of the ride-hailing sharing market (e.g., the profit of the platform, the number of matched requests, matching percentage, etc.). The value of the fare discount multiplier for the passengers who successfully share RHVs with others can affect the number of shared requests, the number of matched requests, and platform profitability.

Electric vehicles (EVs) are more environmentally friendly than gasoline vehicles (GVs). To reduce environmental pollution caused by ride-hailing gasoline vehicles (RGVs), they have been grad-ually replaced with ride-hailing electric vehicles (REVs). Like RGVs, REVs can allow passengers to share trips with others. However, REVs are plagued by charging needs in daily operations. This study develops a simulation-optimization framework for the dynamic electric ride-hailing sharing problem. This problem integrates a dynamic electric ride-hailing matching problem (with sharing) and a dynamic REV charging problem, both of which aim to match REVs to passengers willing to share their trips with others and schedule the charging events of REVs on temporal and spatial scales, respectively. The dynamic electric ride-hailing matching problem is divided into a set of electric ride-hailing matching subproblems by a rolling horizon approach without a look-ahead period, while the dynamic REV charging problem is divided into a set of REV charging sub -problems by a rolling horizon approach with look-ahead periods. Each REV charging subproblem incorporates a novel charging strategy to determine the charging schedules of REVs and relieve the charging anxiety by considering the information of requests, REVs, and charging stations. Each REV charging subproblem is formulated as a mixed integer linear program (MILP), whereas each electric ride-hailing matching subproblem is formulated as a mixed integer nonlinear pro-gram (MINLP). The MINLP and MILP are solved by the artificial bee colony algorithm and CPLEX, respectively. The proposed simulation-optimization framework includes a simulation model which is used to mimic the operations of REVs and update and track the state of passengers and the charging processes at charging stations over time using the outputs of each MILP and MINLP. The results show that the proposed charging strategy outperforms the benchmarks with a shorter waiting time for charging and a higher matching percentage in the dynamic ride-hailing matching problem. The robustness of the proposed charging strategy is tested under different scenarios with changing the initial state of charge (SOC), the number of REVs, the number of charging piles at each charging station, the time to fully charge, and the distribution of charging piles. The results show that REV drivers can charge their vehicles more flexibly without waiting too long and then pick up more passengers under all test scenarios.