Abstract: Ship operation time is a crucial factor in developing berth plans. While most existing research focuses on container terminals, few scholars conduct in-depth studies on bulk terminals due to the unique nature of their cargo and complex operational processes. Considering the similarities and differences among different berths at bulk cargo terminals and the limitations of a single machine learning model, this paper first proposes a berth clustering method based on K-means to group berths with similar operational traits. Then, an adaptive stacking ensemble learning combined with simulated annealing algorithm (SA-Stacking) is proposed for predicting ship operation time in various types of berths at bulk cargo terminals. This paper is based on the real historical operation data of Qingdao dry bulk terminal. Experimental results demonstrate that the SA-Stacking model can adapt to select the most appropriate combination of base models according to the data distribution characteristics of different berth categories, and has better prediction effect and generalization ability than a single machine learning model. Furthermore, the prediction approach based on berth characteristic classification can better capture the operational characteristics of different berths. Compared to prediction without berth clustering, the mean absolute error is reduced by nearly 2 hours, and both the root mean square error and the mean absolute percentage error are significantly reduced. At the same time, the R-square is increased, enhancing the model's interpretability and prediction accuracy.