Research On Forecasting Method Of Urban Rail Transit Passenger Flow Based On Big Data

Passenger flow forecast plays an essential role in the operation and management of urban rail transit system,which is an important prerequisite for rational allocation of transportation resources,optimization of train operation plan,guidance of daily transportation organization and evaluation of economic benefits.With the advent of the information age,technologies such as mobile Internet technology,information communication technology and cloud computing have developed rapidly,and the amount of data generated during passenger travel and rail transit operations has shown a rapid growth.The original passenger flow forecasting methods based on sampled data faces great challenges when fusing massive,multi-source,heterogeneous traffic big data and meeting personalized and diversified travel demands.This paper studies the big-data-driven urban rail transit passenger flow forecasting method guided by traffic planning theory,traffic behavior theory,statistical theory and deep learning theory,to process comprehensive traffic information with different levels and different structures timely,efficiently and accurately.The research on passenger flow forecasting method has important theoretical significance and engineering practice value for promoting urban traffic more intelligent.This paper proposes a prediction method for medium and long term passenger flow of urban rail transit based on computational graph,which integrates the basic framework of deep learning—computational graph and four-stage method in traffic planning theory.A passenger flow prediction nonlinear optimization model is constructed with the social economic data and AFC data as the input information.The relevant variables and parameters in the passenger flow prediction model are corrected and the OD demands can be predicted using computational graph.In addition,in view of the statistical characteristics of regularity,volatility and randomness of passenger flow,the Kalman filter algorithm is used as a framework to study the short-term passenger flow prediction both under normal and unnormal conditions.The major research contents involved in this thesis is listed as follows:(1)The multi-source traffic big data generated in passenger travel and urban rail transit operations are discussed systematically,and the key technologies used in the big data process of collecting,storing,mining and visualizing are introduced.The research analyzes the multisource data fusion mechanism based on computational graphs by adopting distributed expression and back propagation technologies.(2)The process from demand generation and OD formation of urban rail transit passenger flow is divided into three stages: traffic generation,traffic distribution and mode selection.Due to the nonlinear complex relationship between traffic generation and its influencing factors,a traffic generation prediction model based on RBF neural network is established,taking into account conventional and special factors including land use,family characteristics,individual characteristics,holidays,severe weather and big events etc.the kmeans clustering algorithm is designed to learn the basic function center of RBF neural network.(3)The nonlinear optimization model of the urban rail transit passenger flow forecasting is established to minimize the total error between the prediction and measurement results using the traffic generation and attraction quantity obtained in the traffic generation prediction stage and the historical OD demand obtained by the AFC data.The working mechanism of the model is represented using computational graph.The forward and backward propagation algorithm framework of computational graph is designed,and the gradient descent algorithm is embedded to update traffic parameters.The partial derivative between loss function and traffic parameters of computational graph is obtained using the calculus chain rule,and marginal effect of influence factors on passenger flow is analyzed quantitatively.(4)The short-term passenger flow forecasting system architecture based on Kalman filter is proposed.The state transition equation is used to predict the future state using the current passenger flow information.The measurement equation is established to correct the predicted value and gradually reduce the total error between predicted and actrual passenger volume.The passenger flow is divided into three parts: regular passenger flow,disturbance passenger flow and random passenger flow.The regular passenger flow is updated by analyzing historical passenger flow data,and the disturbance passenger flow is none when no special event occurs.The short-term prediction of normal passenger flow under normal conditions is realized based on the ordinary Kalman filtering method.The polynomial trend model is constructed to analyze the variation of the disturbance passenger flow in time series using,and the unnormal passenger flow is predicted using the Kalman filtering prediction framework.