A Transfer Learning Forecasting Method With Knowledge Discovery On Railway Frequent Flyer Program For Railway Short-Term Passenger Flow

According to the Medium and Long-term Railway Network Plan(2016-2030),by 2025,the scale of the railway network of China will reach about 175,000 kilometers,of which the high-speed railway will be about 40,000 kilometers.With the large-scale operation of highspeed railways in China,in order to satisfy the changes in the passenger transport market and passenger flow fluctuations in a timely manner,the railway department increasingly needs a scientific and accurate short-term passenger volume forecasting method.Moreover,the railway department could provide the decision-making input data for the passenger transportation plans and operation management plans more quickly,accurately,high-quality,and conveniently.At present,the short-term railway passenger flow forecasting for highspeed railways faces the following problems.First,it deals with the diversified characteristic ineffectively for passenger flow forecasting.The second is the rapid development and continuous evolution of China’s high-speed railway network.Third,it is challenging to predict the passenger flow during holidays,which requires high accuracy of the prediction results.Fourth,most of the prediction methods are unable to apply to specific groups.To deal with the above issues,this paper seizes the characteristics of historical and frequent flyer data,and leverages the machine learning methods,time series decomposition,similarly calculation,and transfer learning,to realize the specific forecasting for key time periods and key populations,and adapt to the background and requirements of the increasingly diverse application scenarios of short-period passenger flow forecasting.The main works are as follows:(1)Forecasting methods via Machine learning for railway short-term passenger flow.On the basis of literature review and the development of passenger flow forecasting and the progress of frequent flyers,this paper analyzes the impact of passenger travel behaviors,weather,holiday passenger demand characteristics,time series features,and historical data on the short-term passenger flow prediction for high-speed railways.Feature engineering is applied with the feature generation,and feature selection based on the importance score.Then,the random forest,Xgboost,LSTM,and GCN is developed with feature engineering to predict the railway short-term passenger.The experimental results show the effectiveness of feature engineering.(2)A decomposition-based forecasting method with transfer learning for railway shortterm passenger flow.The railway short-term passenger flow time series is decomposed into linear and nonlinear time series.For the prediction of nonlinear time series,a similarity calculation method considering the holiday time series is proposed for the sample filtering to effectively increase the train samples.Transfer learning combined with sample filtering is applied to reduce the "negative migration" to improve the performance on holidays.The experimental part reports the time series similarity calculation comparison and forecast performance analysis.(3)A transfer learning forecasting method with knowledge discovery on railway frequent flyer program for railway short-term passenger flow.Based on frequent flyer knowledge discovery,the method of time series decomposition combined with transfer learning is further proposed.The prediction problem is transformed into a multivariate time series forecasting problem,and an improved similarity calculation method for the multivariate time series is leveraged to effectively filter samples,and the prediction method of multidimensional time series based on deep learning is constructed.Then,the effectiveness of the improved similarity calculation method and prediction proposed method are analyzed.(4)Case study.An actual frequent flyer program data and passenger ticket data of China’s Beijing-Shanghai high-speed railway are managed for the case study.After the data cleaning,the necessity of transfer learning and the necessity of similarity calculation are shown in the quantitative analysis.Furthermore,the machine learning method for railway short-term passenger flow,the decomposition-based forecasting method with transfer learning for railway short-term passenger flow,and the transfer learning forecasting method with knowledge discovery on railway frequent flyer program for railway short-term passenger flow are tested.The experiment results outperform the baselines.