Multi-station Passenger Flow Prediction For Urban Rail Transit Based On Graph Convolutional Network

This paper aimed to explore the problem of metro passenger flow prediction.With the continuous expansion of the urban scale,the metro has become an important part of urban rail transportation and a preferred mode of travel for people.However,due to the growing number of urban residents and the accelerating pace of urbanization,traffic congestion has become increasingly severe,particularly during peak periods when metro stations experience high passenger flow,which can lead to congestion and safety hazards.Therefore,predicting passenger flow is crucial for efficient metro operation and intelligent transportation systems.However,existing methods for predicting passenger flow often need help accurately forecasting changes in passenger numbers.Therefore,accurately predicting metro station passenger flow in advance is of great practical importance,as it enables the implementation of effective measures to alleviate congestion promptly.This paper proposes three deep-learning models to address the prediction problem of urban rail transit passenger flow and the relationship between multiple stations.In-depth research and experimental validation were conducted to validate the effectiveness of these models.Firstly,a Spatial-temporal Gated Recurrent Graph Convolutional Network(STGRGCN)was proposed to capture local and global correlations dynamically.The model accomplished this by constructing three spatiotemporal graphs,namely the spatial distance graph,temporal correlation graph,and temporal connectivity graph,to capture spatiotemporal correlations between adjacent time steps.The spatiotemporal graph convolutional module was employed to model spatial and temporal correlations.Additionally,the model incorporated a gated recursive embedded graph convolution unit to capture dynamic long-range spatiotemporal dependencies.Finally,a recurrent neural network structure was utilized to output the predicted multi-site features.Subsequently,a Multi-Graph Convolution and Transformer Neural Network(MGCTNN)was proposed,considering the complex relationships between stations and time-varying characteristics.The model introduced three graph relations: connectivity graph,association graph,and interaction graph,based on the passenger flow characteristics and the topology between rail transit stations.The model effectively captured local spatiotemporal dependencies by constructing a multi-graph convolutional unit that combined these three graph relations with GRU.Additionally,the output state of each multi-graph convolution unit was combined with the input and fed into the Transformer model to extract global features from the long-range temporal dimension.Then,a new model called the Multi-scale Spatio-Temporal Dynamic Hypergraph Convolution Network(MSSTDHCN)was proposed to capture the multi-site spatiotemporal relationship of passenger flow from higher-order dimensions.The main structure of the model consisted of four components: the Multi-scale hypergraph convolution module(MSHCM),the Dynamic graph convolution module(DGCM),the Gated temporal convolutional network(Gated TCN),and the external factor feature fusion(EFFF).First,the MSHCM extracted high-dimensional spatial features using hypergraph convolution.Then,it processed the OD pair information in the data by constructing three different hypergraph structures and employing clustering methods to form hyperedges.Next,the DGCM utilized an embedding matrix instead of a predefined adjacency matrix.Finally,it combined the embedding matrix with the output of MSHCM and integrated it with GRU to learn the implicit spatiotemporal relationships between sites.Finally,the Gated TCN employed various sense fields to extract spatial features from the original long-series data and the upper GRU output states.It generated predicted values for future passenger flow data.Finally,a prediction and analysis system was presented and constructed based on the proposed passenger flow prediction method.The system utilized data processing algorithms built on the Flink tool,enabling real-time or offline analysis of metro station passenger flow data.The system leveraged the passenger flow prediction model mentioned in the previous section to forecast passenger flow.Once the prediction process was completed,the system intuitively presented the prediction results to users.Additionally,the system offered a passenger flow data analysis function,facilitating a better understanding of the changing patterns of passenger flow through visual analysis.This feature aimed to assist metro companies in formulating more scientifically grounded operation strategies.In this paper,the AFC data of Hangzhou metro stations in 2019 were selected to verify the effectiveness of the three models upon completion,and the results were obtained through repeated experiments.The experimental verification demonstrated that the method proposed in this paper exhibited excellent performance in predicting metro passenger flow.In the experiments,the proposed method achieved significant improvements in all evaluation indexes compared to the traditional method.The method holds crucial practical application value in enhancing the accuracy of metro passenger flow prediction and improving the efficiency of urban rail transportation.