Short-term Prediction Of Urban Rail Passenger Flow Based On Deep Learning

Under the background of resource depletion and energy control,urban rail transit has become an inevitable need in the process of urbanization,and the short-term passenger flow prediction of urban rail transit will help the efficient operation of urban rail transit,improve passenger travel experience and reduce urban pollution.The passenger flow distribution of urban rail transit is the result of many factors,shallow machine learning model can not effectively learn it,therefore,this paper will study the short-term passenger flow forecast of urban rail transit based on the deep learning model—Convolutional Long Short-Term Memory(Conv LSTM).In this paper,the k-convlsm single station prediction model is designed by combining the Conv LSTM and k-means clustering algorithm.Specifically,to ensure the full integration of characteristics,this paper creats "distribution images" of passenger flow based on the spatial association between adjacent stations,and uses Conv LSTM to extract the temporal and spatial characteristics at the same time,also,the similarity of passenger flow in certain time slots will be extracted by adaptive k-means clustering algorithm.By comparing with parallel CNN-LSTM model,Conv LSTM,LSTM,Bi-LSTM,SVR and BP neural network,the effectiveness of the model was verified.On the basis of single station short-term passenger flow research,forecast of multi station passenger flow will bring more effective information for managers.To reduce the randomness of multi-station forecast as much as possible,this paper selected 13 indicators(such as convenience,importance and surrounding development of the station)to evaluate the external characteristics of passenger flow,and built a multi-station passenger flow forecast model based on Stacked Auto Encoder and Conv LSTM.To help Conv LSTM fully extract inherent laws of the above indicators,the Stacked Auto Encoder is used.Considering the time cost of autoencoder and the effectiveness of encoding results,the Latent Motion Monte Carlo Tree Search is designed to optimize the hyperparameters of Stacked Auto Encoder,and the output of encoder will be combined with historical spatio-temporal data as the basic data for passenger flow prediction.Finally,the validity and stability of the model were verified by the application in Shenzhen Metro data.