Deep Learning Based Subway OD Passenger Flow Prediction

The subway has the characteristics of large volume,high timeliness and low pollution.Many large and medium-sized cities have already opened subway lines or put the subway construction plan on the agenda.The operation of the subway has alleviated the traffic pressure of densely populated cities to a certain extent,speeding up the efficiency of urban commuting,and occupying an important position in urban public transportation.In order to improve the service level of the subway operation and ensure the safe and stable operation of the subway,the prediction research of the subway passenger flow is an important issue.OD(Origin-Destination)passenger flow is one of the most intuitive manifestations of passenger travel dynamic distribution.An accurate OD passenger flow prediction model has important guiding significance for subway operation.Compared with the shallow model,the deep learning model has better feature abstraction ability and performance.Based on the spatio-temporal mining of subway passenger flow data,this paper constructs three different types of deep learning models for the characteristics of subway OD passenger flow,they are deep feedforward network(DFN)model,stacked autoencoder(SAE)model and long short-term memory(LSTM)network model.Using the DFN model as the basic model,the experimental comparison and analysis of the performance of different deep learning models.The main research work of this paper is as follows:(1)Based on the Spark platform,the raw data of the subway passenger flow is processed quickly,and the redundant fields in the data are eliminated.Through the analysis of the data samples,the error abnormal data is cleaned,and the holiday information fields are merged.The spatial and temporal characteristics of subway passenger flow data are analyzed and mined,and the evolution law of passenger flow is mastered to provide guidance for the establishment of deep learning prediction model.(2)Firstly,a deep feedforward network(DFN)prediction model is constructed.Considering the problem that DFN encounters difficult performance optimization and local optimization while improving the performance of layer number,a SAE-based prediction model is constructed to layer the deep network layer by layer.The pretraining performs feature extraction and fine-tunes the parameters for specific prediction tasks.According to the time series characteristics of subway OD passenger flow data,RNN has good learning performance for time series information.In order to solve the long-term dependence problem in basic recurrent neural network,the LSTM prediction model is constructed.(3)The prediction experiments of the short-time passenger flow of the key OD and the whole network OD are carried out respectively.The experimental results show that increasing the holiday information as the feature can effectively improve the prediction accuracy,and different time steps have a great influence on the prediction accuracy.Compared with different deep learning models,LSTM predicts the best performance,but the model training time is long;SAE performance is second,and its advantage is low model complexity and fast training speed.Through experimental comparison,the prediction accuracy of the whole network OD is lower than that of the key OD corresponding to a single OD pair,and it only needs to train one model,the prediction effect can also meet the subway operation demand.