Research On Short-term Passenger Flow Forecasting Method Based On Deep Learning

Predicting the short-term passenger flow of subway is an important part of rail transit operation management and it plays an important role in optimizing dispatching plans,formulating operation plans,and preparing train operation diagrams.Accurate and timely subway short-term passenger flow forecasting results can facilitate subway operating companies to actively plan and take appropriate passenger flow control and diversion measures,thereby alleviating the pressure on passenger flow at stations and realizing safe operation of the subway network.In the context of the widespread application of the automatic fare collection system(AFC)in the subway,a large amount of inbound and outbound passenger flow data can be obtained,which provides a basis for the subway short-term passenger flow prediction.Research shows that the combination of deep learning methods and big data has a better performance than traditional prediction methods in short-term passenger flow prediction.Therefore,by combining AFC data,the use of deep learning methods has important practical significance for subway short-term passenger flow prediction.This paper takes the short-term inbound passenger flow of subway stations as the main research object,uses K-means clustering algorithm and correlation analysis to process passenger flow data,and studies the effects of different hyperparameters on BP neural network prediction model,convolutional neural network(CNN)prediction model,long-short-term memory network(LSTM)prediction model.Combining variational modal decomposition(VMD),echo state network(ESN)and long-short-term memory network,a VMD-LSTM-ESN combined prediction model is established.On the premise of the same data,the prediction performance of different models is compared,and the effectiveness of the combined prediction model is verified.The major research contents and conclusions of this paper are as follows.(1)On the basis of Hangzhou AFC data,the original data is preprocessed,and the inbound passenger flow data of each subway station in Hangzhou is further extracted with a granularity of 15 minutes,which provides a good data set for the short-term passenger flow prediction model.Establish BP neural network prediction model,CNN prediction model and LSTM prediction model respectively and optimize network parameters.In order to compare the performance of these three models,let them train and predict on the basis of the Qianjiang Road inbound passenger flow sequence with a time granularity of15 minutes.The results show that the prediction accuracy of the LSTM network model is higher than that of the BP neural network model and the CNN network model.(2)This paper considers the combination of variational modal decomposition,long and short-term memory network and echo state network to establish a VMD-LSTM-ESN combined prediction model.First,use VMD method to decompose the passenger flow sequence,and obtain new sub-sequences through sample entropy calculation and recombination.Second,each sub-sequence is sent to different LSTM predictors for training and the training error is fed back.Third,use the training error as the secondary training data and then send it to the ESN error predictor for error modeling.Finally,the trained combined prediction model is obtained.Under the same data set,autoregressive moving average model(ARIMA),support vector regression model(SVR),LSTM model and VMD-LSTM combined prediction model are used for comparison.The results prove that the VMD-LSTM-ESN combined prediction model has good prediction performance in peak,flat peak and all-day phases,which verifies the effectiveness and robustness of the model.