Study Of The Short-term Passenger Flow Prediction In Urban Rail Transit Networks

With the development of urban rail transit(URT)networks,the spatiotemporal distributions of passengers are increasingly sophisticated.Therefore,there exists significant challenges for the operation and emergency management of URT systems.Nowadays,the technologies such as big data mining and artificial intelligence prevail.Therefore,the operation and management of urban rail transit systems become more and more intelligent.Short-term passenger flow prediction is a basic and principal task for URT systems.However,many related methods cannot meet some requirements such as the high prediction accuracy.Given the aforementioned background,this study seeks to solve the problems of how to obtain the inflow volume,origin-destination(OD)flow volume,and sectional flow volume in URT systems via data mining and artificial intelligence techniques.A framework including short-term inflow prediction,short-term OD flow prediction,and short-term sectional flow prediction is proposed.Results can give critical insights for improving real-time,intelligent,and efficient operations in URT systems.Several sections are listed as follows.(1)A station-level short-term inflow prediction model in URT is formulated.First,an innovative two-step k-means clustering model to cluster subway stations is constructed.Based on the clustering results,their spatial distribution characteristics are analyzed.Second,from the temporal perspectives,the similarities and stationarities of passenger flow series in different time granularities are analyzed.Subsequently,a predictability assessment model is proposed to analyze the temporal characteristics.Finally,a cluster-based LSTM model to conduct station-level short-term inflow prediction in URT is constructed.The model is tested on real-world datasets from the Beijing subway.Results show that the proposed cluster-based LSTM model outperforms most state-of-the-art existing traffic prediction models.(2)Two network-level short-term inflow prediction models in URT are formulated.First,problems of network-level short-term inflow prediction in URT are summarized in detail.Second,a deep learning architecture called Res LSTM to perform network-level short-term inflow prediction in URT is formulated.The Res LSTM comprises Res Net,GCN,and attention LSTM network,etc.The model takes inflows,outflows,network topological information,weather conditions,and air quality information as inputs.Third,a deep learning architecture called Conv-GCN to perform network-level short-term inflow prediction in URT is constructed.The Conv-GCN comprises multi-GCN and 3D CNN network,etc.The model takes inflows and outflows as inputs.Real-time patterns,daily patterns,and weekly patterns of passenger flow series are considered both in the two deep learning modes.Results show that two models solved the problem of how to obtain the inflow volumes and can make predictions for all subway stations in a subway network.The two models are tested on real-world datasets from the Beijing subway.Results show their superiority.(3)A short-term origin-destination flow prediction model in URT is formulated.First,an indicator called origin-destination attraction degree(ODAD)to capture the degree of attracting passengers between the origin and destination is introduced.All OD pairs are classified into five ODAD levels.Second,a station-level LSTM model under different time intervals and ODAD levels to conduct short-term OD flow prediction for each of the subway stations is formulated.An optimal combination of the time interval and ODAD level is subsequently recommended based on results.Third,a deep learning architecture called CAS-CNN to perform network-level short-term OD flow prediction in URT is proposed.The CAS-CNN is composed of split CNN,channel-wise attentive mechanism,inflow/outflow-gated mechanism,and masked loss function,dealing with specific problems in OD prediction tasks,respectively.It elaborately fuses inflows,outflows,and OD matrices in the model architecture.Results show that the model solves the problem of how to obtain the OD flows and could make predictions for all OD pairs in a subway network.The model is tested on real-world datasets from the Beijing subway.Results indicate that the CAS-CNN performs better than many deep learning models.(4)A short-term sectional flow prediction model in URT based on the computational graph(CG)model is formulated.First,the CG model is introduced and the superiority to use the CG-based traffic assignment over the conventional traffic assignment model is presented.Second,the link travel time and station waiting time using the CG-based traffic assignment model are estimated,including the route choice model,k-shortest paths and effective paths selection,the construction of mathematical optimization model,optimization model vectorization,CG modeling.Using the estimated link travel time and station waiting time,the sectional flows are obtained using the passenger flow assignment and agent-based simulation.Finally,two case studies are conducted.One is to verify the proposed model using a virtual subway network.The other is to obtain the sectional flow in the real-world Beijing subway network.Results show that the model solves the problem of how to obtain the sectional flows and can obtain all sectional flows in a subway network.To the best of our knowledge,this is the first time that the CG model is applied to the URT.