The Research On Bus Passenger Flow Forecasting Based On Machine Learning

As one of the widely used means of transportation in cities,the rational operation of public transport is directly related to the operation of social economy and people’s quality of life.Passenger flow data is the research basis of real-time bus operation planning,involving multiple decision-making stages such as departure schedule optimization and vehicle scheduling.To improve the service quality of public transport system,we should not only optimize the infrastructure construction to meet people’s growing travel demand,but also consider the operating costs of public transport companies and the existing road network conditions.In order to meet the demand for transport capacity at the minimum cost,it is necessary to constantly adjust the real-time operation plan of public transport according to the actual situation.According to the forecast results of public transport passenger flow,public transport companies can allocate resources in advance to deal with real-time situations,which will help alleviate traffic congestion and improve passengers’ travel experience.However,the current line passenger flow forecasting method mainly uses the passenger flow of a single line to predict,ignoring the improvement effect of the similarity of passenger flow between multiple lines.In addition,based on the natural geographical connection between stations and the inevitable phenomenon of passenger flow caused by urban life,the passenger flow prediction problem of stations is very suitable to be solved by graph neural network,but there is little application research in this area.Aiming at the problem of bus passenger flow prediction,based on the data of getting on and off the bus,this paper takes bus lines and bus stops as the research objects respectively.(1)Analyze and quantify the correlation between lines,determine the relevant lines of bus lines to be predicted,and extract the public characteristics of passenger flow data.A multi-task recurrent neural network model is established,and the passenger flow forecasting task of several lines with high correlation with the target forecast line is taken as an auxiliary task,and the model is trained and predicted.Multi-task learning mode can guide the model to learn the passenger flow fluctuation trend of multiple lines through the shared representation in the sharing layer,improve the universality of model parameters,reduce the sensitivity of the model to data fluctuation,and then improve the prediction effect of the model.(2)Using the location relationship between bus stops and the temporal-spatial relationship of passenger flow distribution,the correlation between stops is quantified,and the station relationship matrix is constructed.Based on the station relationship matrix,the graph neural network is used to predict the inbound passenger flow of all stations on the bus network at the same time.Short-term inbound passenger flow has the characteristics of easy to be disturbed by the outside world,large fluctuation and obvious temporal and spatial characteristics.Joint prediction of multi-stations based on network is beneficial to capture and predict the fluctuation direction of passenger flow.In this paper,the real station location information and passenger card swiping data of a city bus network are used to analyze the line passenger flow prediction method and the station passenger flow prediction method respectively.Experimental results show that the prediction accuracy of the proposed method is higher than that of the traditional sequential pre-traditional time series prediction model and the neural network prediction model which only considers the passenger flow of a single line.Compared with the usual prediction methods,the method used in this paper has higher prediction accuracy and the selected spatial features are effective.