The swift formulation of dispatch adjustments amidst disturbance scenarios poses a critical challenge in the daily operations of high-speed railways. Accurate prediction of train delays, along with their potential impact, is essential for effective dispatching and command operations. Therefore, a hybrid method named Density clustering - Bayesian optimization - Memory network (DBM) hybrid method is proposed, merging data mining and machine learning to increase prediction efficiency and accuracy. Leveraging techniques like Hierarchical Density-Based Spatial Clustering of Applications with Noise, the DBM hybrid method uncovers train delay evolution patterns. Within each pattern, Bayesian Long short-term memory is used to predict train delays. This DBM hybrid method integrates density clustering and Bayesian optimization parameter of the memory network, offering a forward-looking approach to dispatch adjustments. Utilizing real operation data from the Beijing-Guangzhou high-speed railway, the results demonstrate a prediction accuracy of 93.475% with a permissible error of 1 min. This improvement in accuracy and efficiency has significant practical implications.