Free-floating bike-sharing systems have rapidly gained popularity as a viable short-distance transportation mode. As users play a significant role in the movement of the bikes in such a system, the basis for evaluating bike usage is to predict the gap between the locking and unlocking behaviour in a specified grid, which is defined as the inflow and outflow gap. The Spatial-Temporal grid size, including its time and space dimensions, has a significant impact on the microscopic forecasting of the inflow and outflow gap. In this study, a flexible framework for testing the impact of the grid on inflow and outflow gap predictions is proposed. The performance of four algorithms, i.e., linear regression, support vector regression, random forest, and gradient boost machine, was compared for different grid sizes based on the dataset provided by the Shanghai Big Data Joint Innovation Laboratory. The results show that the proposed framework can be used to evaluate the impact of grid size on microscopic forecasts. A smaller grid helps to achieve better model prediction results, but it also involves longer calculation time. Comparisons of the four algorithms show that the larger the spatial dimensions of the grid, the worse the predicted results. Excluding random forest, the other algorithms tend to achieve better results when the temporal dimension of the grid is larger. The gradient boost machine algorithm provides the best results in most scenarios; the optimal spatial-temporal grid side length is 200 m, 24 h with 5 by 5 grid combination. Dispatchers or analysts in bike-sharing companies can refer to the proposed framework to select the bike delivery scale and management scope. In addition, this study will help researchers and practitioners quickly select the appropriate grid size and machine learning algorithms for bike-sharing analysis.

自由浮动自行车共享系统作为一种可行的短途交通方式迅速流行起来。由于用户在此类系统中对自行车的移动起着重要作用，因此评估自行车使用情况的基础是预测指定网格中锁定和解锁行为之间的差距，即流入和流出差距。时空网格大小，包括其时间和空间维度，对流入流出缺口的微观预测具有显着影响。在这项研究中，提出了一个灵活的框架，用于测试网格对流入和流出缺口预测的影响。四种算法的性能，即线性回归、支持向量回归、随机森林和梯度提升机，基于上海大数据联合创新实验室提供的数据集，对不同网格大小进行了比较。结果表明，所提出的框架可用于评估网格大小对微观预测的影响。较小的网格有助于获得更好的模型预测结果，但也涉及更长的计算时间。四种算法的比较表明，网格的空间维度越大，预测结果越差。除了随机森林，当网格的时间维度较大时，其他算法往往会取得更好的结果。梯度提升机算法在大多数场景下提供了最好的结果；最佳时空网格边长为 200 m，24 h，5 x 5 网格组合。共享单车公司的调度员或分析师可以参考提议的框架来选择自行车交付规模和管理范围。此外，这项研究将帮助研究人员和从业者快速选择合适的网格大小和机器学习算法进行共享单车分析。