Spatio-temporal Graph-based Traffic Prediction Of Shared Bikes

As a convenient and environment-friendly travel form,shared bicycles have attracted extensive attention in recent years.The flow prediction of shared bicycles is the key technology to study the dynamic evolution of shared bicycles distribution and carry out reasonable dispatching and management of bicycles.It is very important to solve the problem of uneven distribution of shared bicycles in the current city.However,it is still a difficult problem to predict the shared bicycle traffic accurately and efficiently.The traditional traffic prediction methods cannot deal with the current massive bicycle riding data and are still lacking in solving the problem of temporal and spatial correlation in large-scale areas.To solve the above problems,this paper proposes a shared bicycle traffic prediction method based on spatio-temporal graph model ST＿LGNN can accurately predict the bicycle flow of city scale.We use the weighted mathematical graph model to model the cycling data and get a high-quality sequence diagram of shared cycling stations.Through the comprehensive extraction of the spatiotemporal relationship between stations by ST＿LGNN,the time series data of shared bicycles are also accurately predicted.The test of a large number of measured data shows that the ST＿LGNN algorithm established in this paper is better than several existing optimal algorithms.The average absolute error of ST＿LGNN model is increased by 20%,which proves the correctness of this method.The important contributions of this thesis are as follows.1)Based on the aggregation of shared bicycle parking spots,the sequence model of shared bicycle site map is constructed.Using the adaptive density clustering algorithm,the bicycle parking spots in the shared bicycle data set are reasonably clustered,and a weighted mathematical graph model is established for these stations.Then,by deleting the low-quality stations with low income and utility,a high-quality shared bicycle station map is obtained,and a time dimension is added to these station maps to establish a time series model.2)A new neural network framework ST＿LGNN for spatio-temporal map of short-term and long-term memory is proposed to improve prediction performance.A location representation is established to capture the factors between each node.By learning a potential location representation,and in order to capture the correlation of events,long short-term memory is used to process the sequence information.The LSTM operation is applied to each node separately,and all node parameters are shared with each other.The model uses the converter layer to directly capture global dependencies,and the converter layer is also applied to each node separately.In order to merge spatial relations during processing,GCN operation is modified to realize the comprehensive extraction of temporal and spatial relations,so as to improve the prediction ability of the model for spatio-temporal sequence data.3)Sufficient experiments based on multiple real data sets confirm the excellent performance of the model.Using several actual bicycle trajectory data,the traffic flow of shared bicycle is predicted,and sufficient experiments are carried out,and the results are compared with many existing neural network methods such as DCRNN、Graph Wave Net、STAWnet、STGNN and so on.From the results of MAE,the performance of the model is ST＿LGNN is optimal.To sum up,the experimental results show that the proposed method ST＿LGNN in this paper has an improvement of more than 20% in average absolute error,has good prediction performance,and is better than other benchmark models.This paper uses several real shared bicycle data sets to analyze the proposed model performance of ST＿LGNN.The experimental results show that the proposed model ST＿LGNN in this paper has a very significant improvement in realizing the traffic prediction of shared bicycles in the future.