Demand Prediction Of Urban Rail Transit Passenger Flow Based On Machine Learning

Urban rail transit has become an important part of urban public transportation by virtue of its safety,reliability,and large travel volume.In recent years,the rail transit has been constructed from the single line to the networked operation with the expansion of the city.Passenger flow is the foundation of urban rail transit construction and operation.Demand prediction of urban rail transit passenger flow can help managers know the passenger flow distribution and its dynamic evolution in advance and adjust passenger transportation service plan in time,so that rail transit can operate more safely and efficiently.Therefore,demand prediction of passenger flow is of great significance to scientific and efficient operation management for urban rail transit.Based on the AFC data,this article lays the foundation for the establishment of the origin-destination(OD)prediction model by in-depth analysis of the macroscopic passenger flow characteristics and individual passenger travel characteristics of urban rail transit.Then,this article predicts the macro inbound passenger volume,OD volume and micro individual OD(passenger travel destination).These predictions can obtain the distribution of passenger flow,the precise passenger OD travel information in the rail transit network,and the real-time changes of passenger flow.The main research work of this article is divided into the following three parts:(1)A multi-feature Wavelet-LSTM model of inbound passenger volume prediction is established.Firstly,gray relational analysis and Granger causality test are used to extract relevant stations.Then in order to deal with volatility of the time series,the wavelet transform is used to decompose the time series data to obtain stable subsequences.Finally,each subsequence is input into long short-term memory network(LSTM)which can learn the temporal characteristics to predict inbound passenger volume.This method is applied to predict inbound passenger volume of Tiantongyuan station,where the MAPE is 7.441%.The prediction performance is better than the ARIMA,the model which only input single station's passenger flow,and the model which does not do wavelet transform.(2)Based on the prediction result of inbound passenger volume,a Kalman-Wavelet-LSTM OD volume prediction model is proposed.As the AFC system generally uploads the volume data of each station with a granularity of 15 min,the destination of passenger flow cannot be obtained in advance.It is difficult to obtain the OD volume data of the current period.This article proposes a Kalman filter prediction method to solve the passenger flow state space,and the Wavelet-LSTM correction process is added to Kalman filter to deal with the volatility and temporal characteristics of passenger flow.The method is applied to the OD volume prediction in two directions of 15 min and 30 min between Tiantongyuan and Xierqi.The prediction accuracy of Kalman-Wavelet-LSTM is better than the traditional Kalman filtering.(3)Based on the massive passenger travel details information,this article proposes a model(RF-Light GBM)that uses Stacking strategy to combine Random Forest(RF)and Light GBM to predict passenger travel destination.This model achieves the precise individual OD prediction at the micro level.RF can effectively reduce variance,and the Light GBM which is the improved algorithm of the Gradient Boosting Tree has the advantage of reducing deviation.Therefore,the RF-Light GBM model can predict the passenger travel destinations according to a large number of passengers' feature information.The model is efficient and suitable for big data analysis.Through continuous learning of parameters,the prediction accuracy of the model can reach 86.682%.