北京市共享单车出行的时空规律与需求预测研究

基于北京市摩拜单车的工作日骑行数据，利用非负矩阵分解算法(NMF)从时间和空间两 个维度深入探究共享单车的出行规律，并构建逆序群体稳定性指标(RPSI)改善 k 值选择过程。利用得到的出行规律，运用MATLAB构建基于非负矩阵分解算法的BP神经网络预测模型，对共享单车的出行需求进行预测，并分别与BP神经网络预测模型和长短期记忆(LSTM)神经网络预测模 型的结果进行对比。研究结果表明，共享单车可分为5种基本的出行模式，每个区域的出行都可以由这5种出行模式的线性组合来表达，其中的系数代表了每一种出行模式的强度和随时间变动情况。依据这5种出行模式的时间特征和空间特征，分别确定其出行含义：通勤出行中居住地到 地铁站的出行；通勤出行中地铁站到工作地点的最后一公里连接；居民其他的非通勤出行行为， 如休闲娱乐活动等；回程通勤出行中从工作地点到地铁站；回程通勤出行中从地铁站到居住区的最后一公里连接。最后，模型预测结果的对比分析显示，本文构建的基于非负矩阵分解算法的BP神经网络预测模型不管是在预测精度还是实际操作便捷性上都优于其他两种预测模型。

Spatiotemporal Travel Patterns and Demand Prediction of Shared Bikes in Beijing

Using the Mobike data on workdays in Beijing, this study explores the patterns of bike- sharing behavior through the non-negative matrix factorization(NMF) algorithm from the spatial and temporal dimensions. A reverse population stability index was proposed to improve the selection of the k- value. Then, based on the revealed travel patterns information, the BP neural network prediction model with a non- negative matrix factorization (NMF) algorithm was built by MATLAB to predict the travel demand of shared bikes. The prediction results were compared with other two prediction models, i.e., the BP neural network model without NMF and the long and short term memory (LSTM) neural network model. Results show that the share of bicycle travel can be divided into five basic travel patterns, and the travel demand in each area can be represented by a linear combination of these five travel patterns. The coefficients represent the intensity and temporal fluctuation of each travel pattern. Based on the spatial and temporal characteristics, the five travel patterns can be regarded as cycling from residences to subway stations in commuting travel, cycling from subway stations to workplaces in commuting travel, non- commuting travel such as shopping or recreational travel, cycling from workplaces to subway stations in commuting travel and cycling fromsubway stations to residences in commuting travel. The results also show that the NMF-based BP neural network model proposed in this study is superior to the other two prediction models in both prediction accuracy and practical operational convenience.