A decomposition-based forecasting method with transfer learning for railway short-term passenger flow in holidays

Abstract

The forecasting of the railway short-term volume plays an essential role in the railway company, which is the basement of the ticket assignment, the line planning, and the passenger station management, and so on. However, the short-term flow forecasting tends to yield poor performance in holidays in China and leads to a low service-of-level for the passengers. In order to improve the accuracy of the short-term passenger flow volume forecasting of high-speed railway, and overcome the challenges in holidays, transfer learning with a time series decomposition is managed. The proposed decomposition-based forecasting method with transfer learning is based on a boosting method which is an instance-based transfer. Firstly, the time series is decomposed into linear time series and nonlinear time series. Using the SARIMA model to predict linear time series, the nonlinear time series is acquired and transformed as feature-label samples with a feature selection to the transfer learning. To limit the negative transfer, two sample filtering methods are proposed by the computing of similarity of time series. Finally, the samples in the source domain are learned by a TrAdaboost model which is embedded with Random Forest adjusting the weights of training samples to reduce the negative transfer. The 24 stations of the Jinghu High-speed railway are tested in our experiments. Our transfer learning model outperforms the baselines. The results also show the effectiveness of feature selection and sample filtering methods. It is proved that this method can be applied to the short-term passenger flow prediction of high-speed railway efficiently, and it is beneficial to improve the efficiency of resource allocation and the service level of high-speed railway transportation.