共享单车出行OD的时空注意力残差网络预测模型  

Spatiotemporal Attention Residual Network Prediction Model for OD of Bicycle Sharing Trips

作　　者：曹弋[1] 白涵文 王艺筱 CAO Yi;BAI Hanwen;WANG Yixiao(School of Transportation Engineering,Dalian Jiaotong University,Dalian 116028,China)

机构地区：[1]大连交通大学交通运输工程学院,大连116028

出　　处：《地球信息科学学报》2024年第3期556-566,共11页Journal of Geo-information Science

基　　金：辽宁省社会科学规划基金项目(L22BSH003)。

摘　　要：为探究共享单车出行的复杂时空规律与特性,揭示城市因素对共享单车出行OD的影响,提高OD预测精度,开展本研究。结合城市计算,考虑疫情、天气、温度、风速与节假日因素,构建共享单车出行OD的时空注意力残差网络预测模型(USTARN)。USTARN先将共享单车OD数据通过时空特征切分捕捉单车流的时空依赖性,再结合注意力机制进行深度残差学习,最后根据城市因素学习结果调整预测结果。利用从政府数据开放平台获取的深圳市共享单车订单大数据及城市因素数据集,分析共享单车出行时空分布规律及其影响因素。将OD数据集按7:1:2划分为训练集、验证集与测试集,分别进行训练预测、模型参数自适应调整及模型验证对比实验。研究表明,USTARN模型的共享单车出行OD预测平均误差为7.68%,与不含城市计算的STARN模型及传统的CNN,BiLSTM模型相比,误差分别降低了5.93%、7.55%、6.07%,预测精度显著提高。USTARN模型充分反映了时间、空间、疫情、天气、温度、风速等因素对共享单车出行OD的影响。研究成果对共享单车出行OD的精准预测具有理论指导意义,对该出行模式的推广并解决居民出行“最后一公里”问题具有实际应用价值。This study aims to explore the complex spatiotemporal patterns of bicycle-sharing trips,reveal the influence of urban factors on the OD of bicycle-sharing trips,and improve the accuracy of OD prediction.Combining the theory of urban computing,urban factors such as the epidemic,months,weather conditions(minimum temperature,maximum temperature,and wind speed),and whether it is a weekday along with the length information of non-motorized lanes are selected to construct a bicycle-sharing demand prediction model(USTARN)that integrates urban computing and spatiotemporal attention residual network.USTARN first captures the spatiotemporal dependence of sharing bicycle flow through spatial area division and time series slicing,then combines the attention mechanism for deep residual learning,and finally adjusts the deep residual prediction results according to the urban factor prediction results to improve the model performance.Using the big data from bicycle orders and urban factor datasets in Shenzhen obtained from the government data open platform,this study visualizes the spatiotemporal distribution patterns of bicycle-sharing trips and analyzes their influencing factors using the Python development environment.The OD data set is divided into training set,verification set,and test set in a 7:1:2 ratio,and the model training,model parameter adaptive adjustment,and model result comparison are carried out,respectively.The results show that the average error of the USTARN model for OD prediction of bike-sharing trips is 7.68%,which is 5.93%,7.55%,and 6.07%lower than that of the STARN model without urban computing and the traditional CNN model,which is good at data feature extraction,and the BiLSTM model,which is good at dealing with bi-directional time-series data,respectively.The USTARN model fully reflects the influence of time,space,epidemic,weather,and other factors on the OD of bike-sharing trips.Our results have theoretical guiding significance for the accurate prediction of bike-sharing trip OD,which can provide a s

关 键 词：残差网络 交通数据挖掘 城市计算 时空特征 深度学习 共享单车 OD预测 注意力机制 

分 类 号：U491.225[交通运输工程—交通运输规划与管理]