网约车已经成为城市多样化出行服务体系的组成部分，在缓解公众出行压力，提升城市公共交通运力和服务品质方面，发挥着越来越重要的作用。然而，由于乘客出行动态时变特性以及网约车系统自身的开放性，导致网约车运营优化与合理调度成为亟需解决的技术难题。本文围绕网约车系统中乘客出行需求预测、车辆电池优化、路网状态预测和系统运行调度等人-车-路-系统的关键技术环节展开研究，探究网约车系统的需求变化及相关因素对系统的影响，从而触发主动性的匹配和调度，实现系统运行的优化。首先，针对网约车系统运行中时变特性引发的短时需求预测精度低的问题，基于支持向量回归、决策树回归等7种经典预测模型构建需求预测模型，从预测精度，预测时间和空间复杂度3个指标进行分析评价，数值实验表明Lasso线性回归模型3个指标均最优。在此基础上，提出了一种基于LS-SVM的网约车需求预测模型，研究了该模型的核函数和超参数标定，探讨预测步长、地域特征、周期特征对其预测精度的影响，并利用网约车出行实际数据验证了模型的有效性，与Lasso线性回归、神经网络等经典预测模型做了对比，精度平均提高了6.57%。其次，针对传统的单分布模型难以刻画网约车司机异质性出行需求多维特征问题，本文提出了混合分布模型来描述不同日期不同司机日出行里程的异质性，采用风险价值和均值方差两种优化模型求得电池容量。数值实验表明，混合分布模型相对于单分布模型设计的电池容量更接近于网约车司机的需求。然后，针对系统中车辆调度路径选择问题道路状态预测的难题，本文采用贝叶斯网络建立短时交通流预测模型，提出了基于相关系数的传感器选择方法。数值实验表明该方法相对于单传感器选择方法，预测精度平均优化了6.08%。提出了基于供需的空载网约车优化调度模型，并设计了空载网约车引导方法，能够有效优化运营指标，仿真实验表明无响应订单比例平均下降45.69%。最后，将短时网约车需求预测、电动网约车电池容量优化、短时交通流预测、空载网约车等待区域优化调度4项关键技术集成，构建了一个功能完善的网约车系统。该系统具有基于预测的主动性匹配和优化调度功能，能提高网约车系统的运营效率。
Online car-hailing service has become an integral part of the city's diversified travel service system. It plays an increasingly important role in easing the pressure of public travel and improving the capacity and service quality of urban public transportation. However, due to the dynamic and time-varying characteristics of passengers' travel and the openness of the online car-hailing system itself, the optimization and reasonable scheduling of online car-hailing operation have become technical problems that need to be solved urgently. This paper focuses on the research on the key technical links of the human-vehicle-road-system, such as passenger travel demand prediction, vehicle battery optimization, road network state prediction and system operation scheduling in the online car-hailing system, and explores the changes in demand of the online car-hailing system and the influence of related factors on the system, thus trigger active matching and scheduling and realizing the optimization of system operation. Firstly, in order to solve the problem of low precision of short-term demand prediction caused by time-varying characteristics in the operation of the online car-hailing system, a demand prediction model is built based on seven classic prediction models such as support vector regression and decision tree regression. The three indexes of the model are analyzed and evaluated from the prediction precision, prediction time and space complexity. Numerical experiments show that the three indexes of Lasso linear regression model are all optimal. On this basis, a demand prediction model for online car-hailing based on LS-SVM is proposed. The kernel function and hyper-parameter calibration of the model are studied. The influence of prediction step size, regional characteristics and period characteristics on the prediction accuracy is discussed. The actual data of online car-hailing trip are used to verify the validity of the model. Compared with classical prediction models such as Lasso linear regression and neural network, the accuracy is improved by 6.57% on average. Secondly, in order to solve the problem that the traditional single distribution model is difficult to describe the multi-dimensional characteristics of diversity trip demand of drivers, this paper proposes a mixture distribution model to describe the diversity of daily trip mileage of different drivers on different days, and uses two optimization models of value at risk and mean variance to obtain battery capacity. Numerical experiments show that the battery capacity designed by the mixture distribution model is closer to the demand of the driver than that designed by the single distribution model.Then, aiming at the difficult problem of road state prediction in vehicle scheduling route selection, this paper uses Bayesian network to establish a short-term traffic flow prediction model and proposes a sensor selection method based on correlation coefficient. Numerical experiments show that compared with the single sensor selection method, the prediction accuracy of the method is optimized by 6.08% on average. An optimal scheduling model based on supply and demand for no-load online car-hailing dispatching is proposed, and a guiding method for no-load online car-hailing dispatching is designed, which can effectively optimize operation indexes. Simulation experiments show that the proportion of no-response orders decreases by 45.69% on average. Finally, a fully functional online car-hailing system is constructed by integrating four key technologies: short-term online car-hailing demand prediction, battery capacity optimization of electric online car-hailing, short-term traffic flow prediction, and optimal scheduling of empty online car-hailing waiting areas. The system has the function of proactive matching and optimal scheduling based on prediction, which can improve the operation efficiency of the online car-hailing system.
