基于多尺度时空聚类的共享单车潮汐特征挖掘与需求预测研究

当前,我国政府和单车企业多以划定电子围栏停车点的方式进行共享单车的规范化管理,由于单个电子围栏内部单车流入流出的随机性和不确定性较大,以单个围栏为单位进行单车管理的工作量大且不具现实意义。因此,有必要对电子围栏停车点进行聚类划分,实行区域化的管理与调度。基于此,本文提出一种基于时空约束的网络图聚类算法,该算法综合考虑空间因素（地理位置、地理环境特征）和时间因素（历史订单）,只需通过距离阈值设定即可实现电子围栏的多尺度聚类划分,实验分别在3000 m和700 m距离阈值条件下对厦门岛和乌石浦地区电子围栏进行聚类,结果显示该算法不仅能够将具有相似时空特征的电子围栏聚到同一社区簇内,而且能够使得单车流动主要集中在划分后的社区内部;随后,在社区划分基础上进行单车潮汐特征挖掘,能够有效识别和定位单车使用的热点地区;最后,利用长短时记忆神经网络（Long-Short Time Memory network, LSTM）进行单车订单需求预测,结果显示有84%以上社区的预测准确率在85%以上,平均预测准确率为91.301%,预测效果较好,可有效满足单车调度需求。本文研究成果可服务于电子围栏停车点规划与共享单车的区域化管理与调度工作。

Usage Patterns Identification and Flow Prediction of Bike-sharing System based on Multiscale Spatiotemporal Clustering

At present, China government and bike-sharing companies mostly use electronic fence parking stations to manage the shared bicycles normatively. Electric fence parking stations for free-floating bike-sharing are predetermined 'virtual fences' to guide users to park bikes in designated zones and regulate inappropriate parking behaviors. However, due to the randomness and uncertainty of the inflow and outflow of bicycles at a single parking station, the scheduling of bicycles based on an independent parking station is hard to realize. Therefore, it is necessary to group fence stations into clusters and implement regional management. In this paper, we proposed a network clustering algorithm based on spatiotemporal constraints, which comprehensively considered spatial factors (location and geographical environment of the parking stations) and temporal factors (historical bike-sharing system orders) as the clustering partition basis, and this algorithm can realize the multi-scale groups division of parking stations only by setting a distance threshold. We chose Xiamen Island as the research region. Using the distance thresholds of 3000 m and 700 m respectively, we carried out clustering experiments on the electronic fence parking stations in the whole Xiamen Island and its Wushipu block. The results showed that this algorithm can not only gather the parking stations with similar temporal and spatial characteristics into the same group, but also make the shared bike flow mainly concentrated in the streets within each group, which is convenient for regional management. Then, we mined the characteristics of shared bikes among the partitioned groups, which can effectively identify and locate hot areas for shared bikes. The results showed that subway stations, office buildings, parks, hospitals, shopping malls, and residential areas had a greater impact on the usage pattern of shared bikes. In particular, it is necessary to focus on the accumulation of shared bikes near office buildings, shopping malls, hospitals, and subway stations, and the shortage of bicycles near the residential areas, parks, and factories during the morning rush hours. Finally, we used the Long Short Time Memory network (LSTM) to predict the orders of shared bikes. The results showed that 84% of the groups had a prediction accuracy of more than 85%, and the average of the overall prediction accuracy was 91.301%, which can meet the needs of bike-sharing system scheduling. Our research provides scientific suggestions for relevant departments to arrange electronic fence parking stations, and the LSTM model has high accuracy in predicting bicycle flow, which is effective in reducing the scheduling cost of bike-sharing system and improve the management efficiency.