公共自行车使用时空特性挖掘及租还需求预测

基于宁波市公共自行车刷卡数据、POI(Point of Interest)数据、气象和空气质量等数据，从数据驱动视角，深入挖掘公共自行车使用的时空特征及站点租还车需求预测。在时间上，采用KMeans算法，将站点聚为5类，探讨各类站点的时变需求规律及影响因素；在空间上，提出基于POI 数据的站点用地类型识别方法，将站点分为居住类、交通设施类、办公类和商业休闲类。构建以 15，30，60 min 为间隔，以租还车需求为目标变量的随机森林预测模型，并与常用的 BP (Back Propagation)神经网络、K最近邻方法进行比较。结果表明，随机森林模型的精度更高，适用性更强。以30 min为间隔的站点租还车需求预测精度最高，考虑站点土地利用类型后能有效提高模型的预测精度。本文结果可作为未来站点平衡调度的依据并推广应用于共享单车系统，为改善服务水平提供技术和理论支撑。

Spatiotemporal Characteristics Mining and Demand Forecasting of Shared Bicycle Borrow and Return

This study collected the shared bicycle usage data, POI (Point of Interest) data, the weather, and air quality data in Ningbo, China. With data mining techniques, this study analyzed the spatiotemporal characteristics of shared bicycle usage and predicted the demand for borrowing and returning the shared bicycles. The K-Means algorithm was used to cluster the study sites into five categories to explore the time-varying demand rules and influencing factors. To analyze the space characteristics, this study proposed a method to identify the land use type of study sites based on POI data, which divides the sites into residential, transportation facilities, office, and business leisure. The study developed a random forest prediction model using the time intervals of 15, 30, and 60 minutes and the demand of borrowing and returning bicycles as the target variable. The method was also compared with the commonly used BP (Back Propagation) neural network and K-nearest neighbor method. The results show that the random forest model has higher accuracy and better applicability. An interval of 30 minutes produced the best accuracy in forecasting the station borrowing and returning bicycle demand. The prediction accuracy of the model was improved because of the consideration of the station's land- use type. The result from this study can be used as the basis for scheduling and balancing of future stations and can also be applied to the shared bicycle system as a theoretical support for the service quality improvement.