Demand Forecast Of Shared Bicycle Based On Deep Learning Method

Currently,sharing-bicycle mode generally serve as a potential travel mode for both the“first mile” and “last mile” transportation solutions and also act as connections among different urban transit systems.To meet the challenges of rapid population and motor vehicle growth,interests in development and planning for non-motorized transport systems are growing in many countries.At the same time,the lack of public parking spots,the damage of shared bicycle vehicles,and the imbalance of scheduling in time and space reduce the service quality of shared bicycle system to a certain extent,which restricts the development of sharing bicycle system.Therefore,it’s of great importance for operators and planners to acquire quantitative analysis(including behavioral feature extraction,market segmentation,demand forecasting,etc.)to support facility planning,system redevelopment,and vehicle scheduling.The prediction of traffic demand at a network level is a challenging task in the existing works of literature,not only due to spatial dependencies and time-varying traffic patterns but also complicated external factors(such as weather and events).With the continuous development of big data technology,people’s travel information is conveniently recorded and stored,which provide good conditions for developing data-driven technologies in real-world non-motorized traffic problems.Deep learning methods have been widely used in short-term traffic prediction because of the abilities to better understand and analyze complex nonlinear relationships among traffic data,compared with traditional statistical methods,resulting in more accurate quantitative analyses.Specifically,three sub-topics are conducted to analyze the pedestrian(riders)travel characteristics,forecast the traffic demandIn the first section,data preprocessing and travel characteristics analysis are carried out.The temporal and spatial characteristics of different types of individuals were quantitatively analyzed.In addition,the land-use intensity of residents with different activities(commuting,consumption,transferring,etc.)are identified based on non-motorized data and deep learning methods.This pre-learning work can quantify the attributes and categories of nodes in the network,which provide medium-and long-term information for the networked traffic prediction model.Secondly,Long Short Term Memory Neural Network are used to predict sharing bicycle traffic flow of a single bicycle spot(station).By adjusting the super parameter combination,the model is compared with a series of traditional methods.Finally,this chapter combined with land use data and riding data to predict the demand of shared bicycle parking spots on the network level.In order to integrate land use information effectively,this chapter chooses Spatio-Temporal Graph Convolutional Neural Network as prediction model.Chebyshev polynomials are used to reduce the complexity of the network in the convolution of spectrum domain,and Gate Graph Neural Network is used as the information transfer model of spatial dimension in the convolution of spatial domain.Then,this chapter optimizes the training accuracy of the prediction network by adjusting the parameters.The case study shows that the error term(MAPE)of 1 hour,3 hour and 6 hour are4.8%,5.6%,7.3% respectively,which shows that the proposed model has performance in shared bicycle demand forecasting.The results show that can effectively predict the demand of parking points in the region.In addition,the input information with land use characteristics performs much better than the physical distance.The proposed model lays the foundation for the research of shared bicycle scheduling and parking location optimization.