Study On Deep Sequential Model-Based Metro Passenger Mobility Prediction And Its Mechanism Analysis

Individual mobility prediction(as the modeling of future travel characteristics of individuals)is of great significance to the operation and management of personalized and differentiated transportation applications(such as online car hailing,sharing services,etc.)in the context of "smart transportation".The accurate prediction of individual travel behavior can provide reliable guidance,thus improving management and service level.However,individual travel behavior is highly complex,and it is a challenging problem to accurately predict individual travel behavior.Based on the historical travel records of subway passengers extracted from the data of Automatic Fare Collection(AFC)system,this study makes an in-depth study on the problems existing in the prediction of individual travel behavior.The main contents are as follows:(1)To solve the problem of missing and wrong OD information in AFC data caused by missing or wrong information in the corresponding table between station and card reader,a method of feature extraction based on tensor decomposition of card reader associated passenger flow is proposed to describe the characteristics of passenger flow through each card reader.For the problem of missing information,a missing information filling method based on the fusion of neural network and decision is proposed to repair the missing information in the corresponding table.To solve the problem of information error,an anomaly detection method based on isolated forest and neural network was proposed,which automatically detected the information error of corresponding table and inferred the correct corresponding information.The results show that the proposed method has good performance in the problem of missing information filling and error repair,and the accuracy can reach 80-90% when the ratio of missing/error is small.(2)In view of the lack of in-depth study on the predictability of the data set used in the current research and the lack of reference for the prediction accuracy,this study uses the upper limit quantization model of predictability based on entropy rate to conduct an in-depth study on the predictability of individual travel behavior sequences extracted from AFC data.Through the predictability experiment of single attribute sequence and compound attribute sequence,it is found that the individual travel sequence extracted from AFC data has a strong predictability(up to 70%),but is slightly less predictable than other non-mobile associated data sets(such as mobile phone location record data).At the same time,the relationship between individual travel behavior characteristics and predictability is studied.The results show that individuals with fewer visits and longer travel sequences tend to have higher upper limits of predictability.(3)Aiming at the limited ability of feature extraction and sequence modeling of existing prediction models,a Sequence to Sequence(seq2seq)prediction model based on attention mechanism is proposed for the prediction of individual travel behavior in subway system.Firstly,four feature extraction methods,including attribute embedding,normalized representation,mapping and overlapping coding,were proposed to extract the feature of travel attributes from discrete and continuous data types.At the same time,a model training algorithm based on paired time pointer is proposed to make the model have the ability to grasp the predicted time information.Experimental results show that compared with the traditional machine learning model,the prediction accuracy of the proposed model is significantly improved(more than 10%).Compared with the deep learning model without the predictive time information,the accuracy of the proposed model is improved by 3%.(4)Based on the "black box" characteristics of deep learning prediction model is difficult to get policymakers trust issue,interpretability definition of individual travel forecasting model is put forward,and put forward a kind of based on weights erasing the interpretability of attention assessment framework,research from two aspects of decision-making flip and output probability distribution interpretability of individual travel behavior forecasting model and its significance.The results show that the attention mechanism can be interpreted to some extent in the deep learning-based individual travel behavior prediction model.The significance of the single weight of attention is weak.The combined attention based on the descending order of weights can significantly affect the model decision.