Study On Traffic Forecasting And Travel Pattern Analysis Of Public Bike System

Bike-sharing systems have become an important part of urban intelligent transportation system with their flexibility and convenience.With the outbreak of the Corona Virus Disease 2019(COVID-19),people need to maintain social distance,and cycling has become an important option for people to travel during the epidemic because it reduces the risk of virus infection compared to other travel modes,so research on bikesharing system has become increasingly important.Existing studies related to traffic forecasting for bike-sharing system often rely only on a single spatio-temporal characteristic of traffic data for traffic forecasting.In addition,current research on travel pattern analysis of bike-sharing systems lacks systematic analysis of long span of time before and after the epidemic and difficulty in unifying the spatio-temporal representation of bike networks.In order to accurately predict the traffic flow of bike stations and effectively analyze bike-sharing system travel mode,the following studies are conducted in this thesis:(1)In terms of the bike-sharing system traffic prediction problem,this thesis proposes a bike flow prediction model based on Multi-Graph convolution network and Temporal Attention Neural Networks(MGTANN).The model uses multiple homogeneous Multi-Graph Spatio-Temporal Block(MGSTBLOCK)to handle the spatio-temporal characteristics of data at different temporal granularities,and each network blocks use Multi-Graph Convolutional Networks(MGCN)and Temporal Self-attention Mechanisms(TSM)to obtain the spatio-temporal representations of multi-source data.Then,the spatio-temporal representations of data at different temporal granularities are fused and further combined with external information for bike flow prediction.The performance of the model was compared with nine other baseline models based on real data sets.The experiments show that MGTANN achieves optimal results,decreasing 7.20%and 7.64% in Root Mean Squared Error(RMSE)metrics and 10.26% and 9.68% in Mean Absolute Error(MAE)metrics for station outbound and inbound bike flow prediction,respectively,compared to the suboptimal model.The results of the ablation experiments show that each component can improve the model prediction.The case study results show that the model is able to learn the structure characteristics of the bike network and has good performance.(2)Aiming at the problem of analyzing travel patterns of bike-sharing systems over long time spans,this thesis proposes a model to analyze the difference of public bike travel mode under long-term span.Firstly,the model collects multi-source data,designs an incremental clustering algorithm to unify the spatial representation of bike networks in different periods,and uses the spatio-temporal clustering algorithm to mine potential spatio-temporal features.Then,the visualization and graph network indicators are used to analyze the different characteristics of the clustering network and the travel characteristics of different periods.Finally,the Page Rank algorithm is used to mark important areas in different periods and analyze their characteristics.The analysis results on real data sets reveal people’s travel characteristics in different periods: people’s travel in normal times shows the classic "weekday-weekend" pattern;During the outbreak,people’s travel was severely affected,and the travel time,distance and frequency of bike travel decreased significantly;Over time,people have become accustomed to the impact of the pandemic on their lifestyle,and even when the new wave of the pandemic hits,they have maintained their travel habits without being significantly affected by the pandemic.