Prediction On The Demand Of Public Bike System Based On The Station Clustering

Public bike sharing systems have been widely deployed in many cities worldwide,providing a flexible and “green” transportation mode for citizens.As the rents/returns of bikes at different places are unbalanced at times,the bikes in systems need to be rebalanced frequently.The general solution to this problem is that the system should often use trucks or bike trailers to monitor and redistribute bikes between stations.However,monitoring the current number of bikes at each station cannot solve this problem very well,as it is too late to reallocate bikes after an imbalance has occurred,and the execute time is too long,especially during rush hours.It is desirable to accurately predict the demand and stock levels at each station.This article proposes a demand forecasting method for public bicycle systems based on clusters,because the regularity of the demand of a single station is difficult to capture,which can predict the demand for bicycles of a cluster in a certain period of time in the future.We firstly preprocess the historical data of the New York City public bike system,analyze the general laws and travel characteristics of the bike system,and use bipartite clustering algorithm to cluster stations with near geographic locations and similar transfer patterns to obtain correlations.Stations clustered into one cluster have similar functional properties,and the regularity is more easily captured,and the accuracy of predictions will be improved.Then,according to the time sequence(periodic,trend)and dependency of the rent and return of bike system,the model is divided into two parts: a time series model capturing the periodic pattern and trend pattern based on SARIMA;a residual model capturing the instantaneous changes.After that,we use LASSO regression algorithm combined with meteorological data,transit feature(inter-cluster dependence)and historical residuals(intra-cluster dependence)to get the residual model.Finally,we use the New York City public bike data to conduct an analysis of the model proposed in this paper by comparing with some commonly used methods.Besides,we analyze the impact of each part on the model,and two cases that deviate from the daily situation.The result shows that the cluster-based demand forecasting model proposed in this paper has a high accuracy and can predict unexpected events very well.