Abstract

The paper develops models for modeling the availability of bikes in the San Francisco Bay Area Bike Share System (BSS) applying machine learning at two levels: network and station. Investigating BSSs at the station-level is the full problem that would provide policymakers, planners, and operators with the needed level of details to make important choices and conclusions. We used Random Forest and Least-Squares Boosting as univariate regression algorithms to model the number of available bikes at the station-level. For the multivariate regression, we applied Partial Least-Squares Regression (PLSR) to reduce the needed prediction models and reproduce the spatiotemporal interactions in different stations in the system at the network-level. Although prediction errors were slightly lower in the case of univariate models, we found that the multivariate model results were promising for the network-level prediction, especially in systems where there are a relatively large number of stations that are spatially correlated. Moreover, results of the station-level analysis suggested that demographic information and other environmental variables were significant factors to model bikes in BSSs. We also demonstrated that the available bikes modeled at the station-level at time $t$ had a notable influence on the bike count models. Station neighbors and prediction horizon times were found to be significant predictors, with 15 minutes being the most effective prediction horizon time.

摘要

该论文开发了模型，用于对旧金山湾区自行车共享系统 (BSS) 中的自行车可用性进行建模，并在两个层面应用机器学习：网络和车站。在站级调查 BSS 是一个完整的问题，它将为政策制定者、规划者和运营商提供所需的详细信息，以做出重要的选择和结论。我们使用随机森林和最小二乘提升作为单变量回归算法来模拟车站级别的可用自行车数量。对于多元回归，我们应用偏最小二乘回归 (PLSR) 来减少所需的预测模型，并在网络级别重现系统中不同站点的时空交互。尽管单变量模型的预测误差略低，我们发现，多元模型结果对网络级预测很有前景，特别是在空间相关的站点数量相对较多的系统中。此外，站级分析的结果表明，人口统计信息和其他环境变量是在 BSS 中对自行车进行建模的重要因素。我们还展示了当时在车站级别建模的可用自行车$吨$对自行车数量模型产生了显着影响。站点邻居和预测范围时间被发现是重要的预测因素，15 分钟是最有效的预测范围时间。