Deformation is a crucial indicator to evaluate the safety of dam. With the increase in the number of deformation measuring points, however, prediction often lags because analyzing all the measuring points is time costly. Moreover, despite that traditional machine learning algorithms have ameliorated prediction accuracy, unreasonable selection of parameters has a great impact on prediction results and the process of establishing model is extremely complicated. In view of this, we introduced the fuzzy C-means clustering (FCM) and eXtreme Gradient Boosting algorithm (XGBoost) to partition the deformation measuring points according to the similarity of the change rules, and then established XGBoost prediction model for each partition. Taking the perpendicular line deformation monitoring data of the arch dam as an example, we verified the reliability of the clustering results and compared the XGBoost result with that of random forest prediction model. Result suggest superiority of the XGBoost prediction model in data pretreatment, modeling time, and prediction accuracy.