Abstract: In order to quickly and accurately grasp the train operation status and future operation trend, it is necessary to deeply study the prediction method of train operation delay. Based on the analysis of the train operation data of Beijing-Shanghai High-speed Railway in 2020, including the impact of dwell time on delay and the number of propagation stations under different initial delay time, this paper put forward a full section prediction method based on Recurrent Neural Network(RNN), and used the RNN model of synchronous many to many mode as the basic model structure, established a prediction model of train operation delay. In the selection of eigenvalues, the paper used the integrated gradient scoring method to select the 12 most significant variables from multiple eigenvalues as the independent variables of the prediction model, and verified the delay data of Beijing-Shanghai High-speed Railway in 2020 by the prediction model. The verification results show that the model can achieve 89% accuracy within the error range of 5 minutes on the verification set. The prediction method can meet the needs of actual production, help the dispatching department make scientific decisions and improve the quality of railway passenger service.