Abstract: To accurately predict train the delay time caused by accidents such as urban rail transit equipment failure, and improve the emergency response disposal efficiency and the passenger guidance service level, the association fusion of internet data and field data of metro accidents is carried out. According to the unbalanced metro accident random undersampling data, a cascade classification prediction method based on GBDT (gradient boosting decision tree) is proposed to predict the delay time of metro accidents. The results indicate that when the delay time allowable deviation of GBDT cascade classification method is 0~5 min, the predicted delay time accuracy of the method is 20%~25% higher than that released on site, and 5% higher than that of GBDT multi-classification prediction method.