Demand Prediction For Stations Of Public Bikesharing Based On Spatial-Temporal Data Mining

Public bikesharing(PBS)is an integral part of the urban green transportation system and an important way to serve the citizen for short-distance travel and public transport connection and transfer.However,the challenge of bike unavailability and dock shortages has always been the main drawback of restricting its service upgrading.In the context of big data driving transportation service innovation and upgrading,how to establish a close connection between data resources and travel demand prediction is the key to solving the current challenge.Consequently,the primary objective of this study is to explore temporal-spatial characteristics and propose a novel demand prediction method using IC datasets in Ningbo,China,which can be utilized to enhance the PBS service level.The research work is as follows:Firstly,an analysis of the temporal-spatial demand characteristics of PBS was conducted thoroughly on the premise of IC dataset structure explanation and preprocessing.The usage mode was extracted using time series analysis,K-Means clustering,and spatial statistics.The results showed that:(1)The usage behavior of PBS had obvious morning and evening peak characteristics in the main urban area of Ningbo.(2)Some stations in residential and commercial areas showed obvious tidal usage characteristics.(3)The behavior of pick-up and drop-off showed the characteristics of positive spatial autocorrelation.After that,a spatial lag model(SLM)was established to explore the correlation of station usage demand and surrounding built environment.The model results showed that:(1)Population density,length of main and second road,bus stops,the capacity bike station and the POI types of residential,public services,and road facilities hold a significantly positive correlation with station demand.(2)Some effect variances of pick-up and drop-off in weekday and weekend,could be created from the different types of POI in the buffer zone.The study deepened the understanding of the heterogeneous characteristics of PBS travel demand in the study area,and provided a basis for the feature selection in the subsequent demand prediction.Finally,this study aimed at the actual demand for the prediction accuracy of pick-up and drop-off.A combination model based K-nearest neighbor(KNN)and LightGBM algorithm,which was called as KNN-LightGBM,was proposed to predict the PBS trip production and attraction.The method inherited the advantages of LightGBM algorithm in terms of accuracy and interpretability,and could measure the similarity of time series benefit from KNN algorithm.To validate the effective extraction of relevant series from model,a delicate control experiment was designed considering algorithm selection rationality and combined model effectiveness.The results showed that:(1)The LightGBM algorithm could effectively and efficiently predict the production and attraction in daily and hour dimensions.(2)The prediction accuracy of the knn-LightGBM model considering the spatiotemporal correlation characteristics was superior to the single LightGBM prediction method.The average reduction of mean square error(MAE)was 11.93%,and the goodness of fit(R~2)increased by an average of 4.67%.The feasibility and validity of KNN-LightGBM combination prediction model was verified.In the future,it can play an application value in the big data management system of PBS.