Abstract: Ridesplitting can effectively improve the transportation efficiency of vehicles and has great potential for emission reduction compared with regular ridesourcing. However, in reality, whether a ridesplitting trip reduces CO2 emissions, is determined by many factors with heterogeneity and uncertainty. To identify the ridesplitting trips with greater carbon emission reduction potential, this study proposes a machine learning-based model for predicting the CO2 emission reduction state and interpreting the CO2 emission reduction mechanism of ridesplitting. First, the CO2 emissions of shared rides (ridesplitting) and their substituted single rides (regular ridesourcing) are calculated based on the COPERT (COmputer Program to calculate Emissions from Road Transport) model using the real-world order data and trajectory data of ridesplitting in Chengdu City. Then, the actual CO2 emission reduction of each ridesplitting trip compared with regular ridesourcing trips is quantified. Given the CO2 emission reduction and order attributes of ridesplitting trips, the XGBoost (eXtreme Gradient Boosting) model is trained to predict the CO2 emission reduction states of potential ridesplitting trips in the future. Finally, the ALE (Accumulated Local Effects) analysis method is used to analyze the mechanism of the prediction model to identify the key factors influencing the CO2 emission reduction state of ridesplitting trips. The results showed that the average CO2 emission of each ridesplitting trip is 307.23 g in the study area. However, there are still 15% of ridesplitting trips even increasing CO2 emissions. The XGBoost model can effectively predict the CO2 emission reduction state of ridesplitting trips. In addition, the detour rate, the number of shared rides, and the overlap rate are identified to be the three key factors that determine the CO2 emission reduction state of ridesplitting trips. This study provides a theoretical basis for the ridesourcing platform to optimize the matching algorithms of shared rides. It can also realize more efficient and low-carbon ridesplitting and further improve the environmental benefits of ridesplitting.

Key words: urban traffic, CO2 emission reduction prediction, machine learning, ridesplitting, ridesourcing

摘要： 网约车合乘出行可有效提高车辆运输效率，与常规网约车出行相比具有显著的碳减排潜力。然而，现实中网约车合乘出行能否真正减少碳排放受多方面因素影响，往往存在较大差异与不确定性。为识别碳减排潜力较大的网约车合乘订单，提出一种基于机器学习的网约车合乘出行碳减排状态预测模型，并解析其碳减排机理。首先，基于成都市真实的网约车合乘订单与轨迹数据，应用COPERT(COmputer Program to calculate Emissions from Road Transport)排放模型分别计算合乘出行碳排放量及其替代的独乘出行碳排放量，进而得到合乘出行相比独乘出行的碳减排量。然后，基于历史的合乘行程碳减排及其订单特征数据，训练 XGBoost(eXtreme Gradient Boosting)模型以预测未来潜在合乘出行的碳减排状态。最后，采用 ALE (Accumulated Local Effects)分析方法对预测模型进行特征变量解析，以识别影响合乘出行碳减排状态的关键因素。 结果显示：研究区域内平均每次网约车合乘出行可减少碳排放307.23g，但仍有15%的网约车合乘行程未能实现减碳；XGBoost模型可以有效预测网约车合乘出行的碳减排状态，并识别出绕路率、合乘数、重叠率是决定网约车合乘出行碳减排状态的三大关键指标。研究结论可为网约车平台优化合乘订单匹配算法提供理论依据，以实现更高效、更低碳的合乘出行，进一步提高网约车合乘的环境效益。

关键词: 城市交通, 碳减排预测, 机器学习, 合乘出行, 网约车