Research And Implementation Of Demand Prediction For Docked Bike Sharing Stations

Bike sharing systems become more and more popular in many major cities,which provide a good solution to the “first-and-last one mile”travel problem and has positive significance for solving urban traffic congestion and building a green ecology.In bike sharing systems,the distribution of bike rental and return demand is very imbalanced in different periods at different stations,so the system operators need to frequently rebalance the number of bikes among different stations.Accurate prediction of bike demand for each station at different times facilitates the redistribution of bikes in the system.However,the differences in the importance of each station,inevitably caused by the density of urban flow distribution,have never been considered.Also,this kind of density leads to the generation of station clusters.By dealing with multi-source urban data,this paper can mine both the central stations and station clusters in the bike sharing system,and make multi-level predictions on the bike demand.All of these can guide the bike rebalancing operation and improve the utilization rate of systems and the experience of users.This paper designs two algorithms to solve the problem of station cluster mining and demand predicting respectively,which are called CSC(Central Station based Clustering)and HP(Hierarchical Predictor).The CSC algorithm generates station clusters through three phases,namely central station mining,central station clustering,and common station partitioning.In the first stage,we first define and calculate the demand centrality of all the stations.Then a part of the stations with the largest demand centrality is selected as the central stations.The second phase starts with the station which has the largest demand centrality among the central stations.The target central station will be add to the station clusters orderly based on the geographic distance and the cluster's demand centrality we define.The third stage first defines the similarity between stations.Then each common station is divided into the station cluster where the central station with the greatest similarity is located.After clustering,multi-level demand prediction of the bike sharing system is carried out by HP algorithm.HP algorithm first obtains the bike demand prediction results of each station cluster at different time periods through TWSWK algorithm(Time and Weather Similarity Weighted K-Nearest-Neighbor).Then a regularized linear optimization model is defined and trained to predict the demand proportion of each central station in every station cluster.Finally,by combining the above two results,we can get the demand prediction result of each central station at different time periods.This thesis evaluates the effectiveness of our proposed algorithm through experiments on the two-year dataset of the Citi bike sharing system in New York.Experimental results show that the methods proposed in this thesis,compared with other benchmark methods,can effectively improve the accuracy of demand prediction results both on the scale of station-level and cluster-level.Focusing on the imbalanced demand of the central station can effectively reduce the operating cost without significantly affecting the overall service quality of the system.At the same time,the fine-grained comparison experiments shows that the proposed method can improve the accuracy of bike demand prediction at each time period.