Data-Driven Analysis Modeling And Optimization For Bike-Sharing System

Recent years have witnessed the blooming of bike-sharing system(BSS)worldwide.BSS provides a green,convenient and low-cost transportation means,which becomes a critical solution to the“last mile”problem in urban transportation.There are mainly two kinds of bike-sharing systems,the docked one and the dockless one.This thesis focuses on the docked bike-sharing system.Compared with its dockless counterpart though bikes can only be checked out/in at the stations,docked system reduces a large amount of maintenance cost by the centralized management.However,the urban tidal traffic flow and limited station capacity usually lead to unbalanced usage in BSS:some stations are filled with bicycles while others are in lack of bicycles.This undermines the efficiency of BSS.To improve the service largely depends on the accurate knowledge and modeling of user be?havior.However,the complicated interaction among the users,bicycles and stations and system randomness make it infeasible to build a theoretical model.Thanks to the development of Internet of things,big data and data mining techniques,massive amount of BSS trip data is collected via sensors equipped in the stations,which opens the door for analysis and modeling the user mobility patterns from a data-driven perspective.Combined with machine learning algorithms,we can pre-dict the bicycle usage demand in the near future,which lays the fundamental for system optim iza-tion applications such as effective bicycle rebalance,accurate user navigation.However,due to the mutual impact between stations and users as well as between checkin and checkout,data-driven modeling remains a challenging problems.The shortcomings of existing works are three-fold:1)How to design a fine-grained per station prediction model remains an unsolved problem;2)There is no rebalancing algorithm designed for large-scale BSS;3)In lack of a data-driven evaluation sys-tem.For this end,in this thesis,a fine grained prediction model and related optimization algorithms are proposed.The primary work and contribution are summaries as follows:1.A brief introduction to bike-sharing system and related work are provided,and the deficien-cies of existing works are also analyzed.2.Over 100 million trip records from Hangzhou BSS of year 2013 are visualized from station,user and bicycle perspective,which provides a vivid and systematic view of how bicycles are used.The usage patterns of users and the unbalanced problem of BSS are highlighted.The different characteristics of docked and dockless bike-sharing systems are also presented.3.For the usage prediction problem,started from users'mobility model,a probabilistic transfer model is built to capture the relationship among stations.Combined with random forest model,a station-level fine-grained prediction model is proposed,which lays the f-undamental for efficient system optimization.4.For the effective rebalancing problem,based on the prediction results,this thesis introduces the concept of rebalancing interval and then,the rebalancing problem is formulated to min-imize the operational cost.This problem is an NP-hard problem so an asymptotic algorithm and an approximate algorithm based on deep sequential learning are proposed.A dedicate event-driven simulation tool is designed for evaluating the proposed approaches.5.To improve the successful rate of checkin/checkout and keep the station self-balanced,a navigation system for user is devised.The aim of this design is to take advantage of the crowd.By directing users from concentrated stations to surrounding ones,the system tries to migrate the usage imbalance problem.Based on the predicted checkout/checkin demand,the system first models the interaction between users and stations and then according to this model,a cost function is established to determine the optimal checkout/checkin stations.Last but not least,the thesis is concluded and some future works are discussed.