Research On Passenger Flow Prediction Model Of Rail Stations In Special Events By Integrating Social Network Data

With the acceleration of urbanization in China,urban transportation is confronted with a range of issues,including road congestion and suboptimal transportation systems.Urban rail transit,as the focal point of new infrastructure construction,has emerged as a primary driver supporting the advancement of urban public transportation systems due to its numerous advantages,including superior efficiency and convenience,enhanced safety and reliability,as well as environmental sustainability and energy conservation.The holding of special events,represented by cultural and sports activities,has the potential to generate large-scale crowd dispersal within rail transit stations,exacerbating sudden surges and accumulation of short-term passenger flows.This can significantly impact the travel demand patterns of rail transit stations surrounding event venues and result in operational safety issues,including passenger crowding and stampede congestion.In order to effectively manage the spatial and temporal effects of rail traffic during special events,it is crucial to accurately forecast and control passenger flows.This is a necessary prerequisite for developing capacity plans and implementing control strategies to ensure optimal operational efficiency and safety,as well as enhance the quality of passenger services within the rail transit system during such events.Currently,short-term prediction of urban rail transit passenger flow is primarily based on typical operational conditions,failing to account for distinct passenger flow demand patterns during special events.To overcome this challenge,the present study introduces social network data as a non-traffic data set that captures the psychological and behavioral characteristics of potential travelers.By analyzing historical rail transit passenger flow patterns,this thesis investigates the prediction of passenger flow entering and exiting rail transit stations during special events.The key research findings and conclusions are outlined as follows:1.Analysis of the evolution of rail transit passenger flow during special events.The thesis defines the concept,characteristics,and classification of special events,and sorts out the concept of rail traffic flow under special events and the passenger flow description indicators.The mathematical and statistical method is utilized to quantitatively describe the distribution characteristics of rail transit passenger flow under special events,aiming to clarify the process of passenger flow changes at stations during special events and the scope of influence,and to explore the differences between the distribution of passenger flow and normal passenger flow.Additionally,the changing characteristics of rail transit passenger flow are summarized,and the factors that influence rail traffic flow under special events are summarized from the dimensions of time,space,and external factors.2.Validation of the correlation between social network data and rail transit passenger flows.The thesis analyzes and classifies the current social network data collection methods,examines the content composition of social networks based on user behavior from static attributes and dynamic behaviors,designs social network data collection rules based on the Scrapy crawler framework.Text preprocessing techniques such as redundant data removal,text segmentation,and stop word removal are applied.Latent Dirichlet Allocation(LDA)topic model and natural language sentiment analysis methods are used to extract semantic features of text data from the perspectives of topic attributes and emotional attributes.The content of social network data is characterized by the frequency of posts and sentiment scores,and the time distribution characteristics of the indicators are described.The correlation between social network data and passenger flows at rail transit stations is quantitatively analyzed using the relevant coefficient.3.Constructing inbound and outbound rail transit passenger flow prediction models based on ensemble learning theory.From the perspective of normal and activity features,the factors affecting passenger flow in urban rail transit under special events are constructed,and the selection basis and construction method of indicators are explained.Based on the analysis of the difficulties in predicting urban rail transit passenger flow under special events,an ensemble learning framework for short-term passenger flow prediction is proposed from the perspective of solving ideas and methods.e Xtreme Gradient Boosting(XGBoost),Random Forest(RF),and K-Nearest Neighbor(KNN)are selected as the primary learners,and the Linear Regression model is the meta learner to design a Stacking integrated learning model suitable for predicting rail transit passenger flow under special events.The cross-validation method and evaluation indicators are determined for model hyperparameter selection and validation evaluation.4.Empirical research and result analysis.To validate the effectiveness of the features and models proposed in this thesis,an empirical study is conducted on the passenger flow at stations during special events in Chongqing rail transit.The feature set without social network data and the full feature set are used as inputs for the base model,and the prediction results are compared among different feature sets.The prediction accuracy of the base model is also compared with that of other commonly used passenger flow prediction models using the Stacking integrated learning model to predict the inbound and outbound passenger flow of the rail transit system.The thesis demonstrates that the inclusion of social network data can typically decrease the prediction error rate of individual models,and the average prediction accuracy of the 15-minute granularity inbound and outbound passenger flow prediction with the full feature set as input is increased by over 4%.Furthermore,the error rates of the fusion model’s inbound and outbound passenger flow predictions are 8.55% and 9.64%,respectively,in comparison to the XGBoost,RF,and KNN models.The highest prediction accuracy of the Stacking integrated learning model is nearly 3.5%,while the lowest is nearly 0.7%,which verifies its precision in predicting inbound and outbound passenger flow during special events.