Research On Rail Transit Passenger Flow Prediction Based On Deep Learning

With the rapid growth of the economy,the acceleration of urban modernization,and the steady growth of the population,the problem of urban traffic congestion is becoming more and more serious.As an important part of urban rail transit,the subway plays an irreplaceable role in alleviating traffic congestion.However,with the rapid development of rail transit today,the rapid increase of passengers has made the subway system in many cities in a long-term overload operation state,which poses a great threat to the safety of passengers.Subway passenger flow prediction is an important direction in the field of traffic flow prediction.Accurate subway passenger flow prediction can help managers to make reasonable vehicle scheduling,effectively reduce the risk of the subway load operation,and provide effective reference information for future rail transit design.The rapid development of deep learning provides a new means for traffic flow prediction research.Therefore,based on deep learning technology,this thesis focuses on how to improve the accuracy of subway passenger flow prediction,which mainly includes the following two aspects:(1)Aiming at the problem that the prediction peak value of Graph Convolutional Network(GCN)is smooth when forecasting passenger flow,a spatial-temporal Graph Convolutional network with enhanced Self-node weight(EST-GCN)is proposed.By increasing the proportion of node weights on the diagonal of the normalized matrix during information aggregation.The negative impact of the smoothing filter defined in the Fourier domain of GCN on the prediction results is solved,and the accurate prediction of subway passenger flow is realized.Firstly,the topological junction of the metro network was abstracted as an adjacency matrix,and the passenger flow data of each station at different times were extracted as the input of the proposed model.Then,the optimized graph convolutional neural network is used to extract the spatial features of traffic data.When nodes perform information aggregation,the ability of spatial feature extraction is enhanced by increasing the proportion of node weights on the normalized diagonal.Then,the time series data containing spatial features were used as the input of the Gated Recurrent Unit(GRU)model,and the dynamic changes were obtained through the gating mechanism and the information transmission between units,to capture the time features more efficiently.Finally,the passenger flow prediction results of subway stations with different time granularities were obtained.The experimental simulation analysis is carried out based on the data from Shanghai,Chongqing,and Hangzhou metro.The experimental results show that the method of maximizing the difference of aggregated information can accurately predict the peak passenger flow,and realize the passenger flow prediction of rail transit stations with different time granularities.(2)A regularized spatial-temporal graph convolutional network model(PMR-GCN)is proposed to aim at the impact of traffic noise data on prediction accuracy.A regularized dual structure is designed to reduce the amplification effect of the model on noise in the training process and enhance the robustness of the model.At the same time,the personalized enhanced graph convolutional network is used to improve the ability of the model to capture spatial features and improve the accuracy of passenger flow prediction.Firstly,a dual structure based on regularization is designed to solve the negative impact of inevitable noise in traffic data on model training,leading to low model robustness.At the same time,the Fourier transform module is used to capture the implicit information in the traffic data,improving the model’s robustness to offset the impact of noise data on prediction accuracy.In addition,a new personalized enhanced graph convolutional network(P-GCN)is proposed,and a trainable diagonal matrix is designed to realize the adaptive learning of the predicted peak passenger flow in the process of neighborhood aggregation,which improves the model’s ability to capture the spatial characteristics of traffic data.At the same time,the rich spatial-temporal features in the passenger flow data are obtained through the multi-head self-attention mechanism.Experiments based on the rail transit data of Shanghai,Chongqing,and Hangzhou show that the personalized enhanced graph convolutional network can effectively improve the accuracy of peak and passenger flow predictions,and the Fourier transform and regularized dual structure can effectively improve the robustness of the model.