Short-term Metro OD Passenger Flow Prediction Based On Deep Learning

In the public transportation system,the metro is widely popular among people because of its punctuality,environmental protection and other advantages.However,with the growth of the urban population,the passenger flow of the metro continues to rise,and the resulting problems such as queuing at the station and crowded carriages have brought inconvenience to the commuters.To improve the service level of metro,more intelligent operations and regulations should be carried out.The metro OD flow prediction is of great practical significance in grasping the dynamic changes of metro travel demand in advance.On the one hand,metro operators can sense the evolution trend of travel demand in advance and adopt some active control strategies to avoid overcrowding.On the other hand,metro passengers can learn future passenger flow patterns and flexibly arrange their departure time and routes.In addition,it is of great theoretical significance to study how to deal with the high dimensionality,sparsity and complex spatiotemporal correlation of OD passenger flow data in the forecasting task.Based on the theoretical knowledge of traffic engineering and deep learning methods,this thesis proposes metro OD passenger flow prediction models.The main research contents and results include:(1)A hybrid model with traffic mechanisms and deep learning tools is proposed.Integrating the deep learning model with the traffic mechanism model,a hybrid OD prediction model called Hybrid model combining Temporal Convolutional Auto-Encoder and Modified Gravity Model(Hybrid-TCAE-MGM)is proposed.The model uses a convolutional auto-encoder structure to solve the sparsity problem of the OD matrix and completes the extraction of spatial correlation features.Meanwhile,the introduction of traffic domain knowledge assists model training.The test results of the real metro passenger flow dataset show that introducing the mechanism model is beneficial to improve the prediction performance of the model and Hybrid-TCAE-MGM is superior to the traditional time series models,machine learning models and deep learning models.(2)A metro OD passenger flow prediction model considering full spatial correlation extraction of the OD matrix is designed.The heterogeneous data OD spatio-temporal tensor and geographic information tensor are constructed,and an innovative model called Heterogeneous Data Feature Extraction Machine(HDFEM)is proposed.The model fuses the heterogeneous data OD spatio-temporal tensor and geographic information tensor through the attention mechanism,and extracts the long-distance features in the OD matrix that the Hybrid-TCAE-MGM model cannot deal with.Experimental results show that the prediction accuracy of HDFEM is higher than that of various comparison models including the Hybrid-TCAE-MGM model,which verifies the importance of extracting long-distance spatial features.(3)A metro OD passenger flow prediction model based on the multi-hypergraph structure is developed.A multi-hypergraph structure is constructed based on the correlation between OD pairs and two strategies(CLOSEST strategy and SAMPLE & CLUSTER strategy),and a novel model called Spatio-Temporal Dynamic Attentive Multi-Hyper Graph Network(ST-DAMHGN)is proposed.This model not only overcomes some shortcomings of the previous models(e.g.,there are long-term low value and random OD passenger flows,and the order of the stations is artificially specified when constructing the OD matrix),but also refines the research granularity to the level of OD pairs.Moreover,the model realizes the dynamic adaptive adjustment of the node perception field in different correlation feature dimensions by the multi-graph attention mechanism.Compared with the above two models proposed in this thesis,ST-DAMHGN further improves the prediction accuracy of short-term OD passenger flow and the experiment results verify the rationality of the multi-hypergraph structure design and the powerful learning ability of hypergraph convolution algorithm.