Abstract: Prediction of the development height of the water-conducting fractured zone is an important basic task for the safe mining of coal mines and the prevention and control of coal seam roof water disasters. The development mechanism of the water-conducting fractured zone in coal seam overlying strata is complex, and there are significant differences in the geological and engineering conditions of working faces, resulting in common problems such as low prediction accuracy and insufficient generalization ability in the prediction methods of water-conducting development height. To improve the accuracy and reliability of height prediction of the water-conducting fractured zone, 311 groups of measured data on the height of the water-conducting fractured zone were collected from 114 mines to construct a sample dataset covering 13 influencing factors including mining height, working face size, mining method, advancing speed, and overlying strata type. Aiming at the problems of insufficient prediction accuracy and poor generalization of single models, a prediction method for the height of the water-conducting fractured zone was proposed, which integrates eXtreme Gradient Boosting (XGBoost) and hyperparameter optimization based on Tree-structured Parzen Estimator (TPE). First, data preprocessing such as data cleaning and data standardization was performed on the raw data. Then, feature variables were comprehensively selected by three algorithms: Random Forest (RF), Gradient Boosting Decision Tree (GBDT), and XGBoost. The TPE algorithm and K-fold cross-validation were used to find the optimal hyperparameter combination, and the XGBoost-TPE prediction model for the height of the water-conducting fractured zone was established. The effects of four optimization algorithms (TPE, grid search, random search, and simulated annealing) on the prediction accuracy of the XGBoost model were compared. Comparative test results based on 8 evaluation indicators and 11 mainstream models show that, compared with single models, the XGBoost model reduces the mean absolute error (MAE), root mean square error (RMSE), median absolute error (MedAE), and training time (TT) by an average of 0.020, 0.019, 0.018, and 0.183, respectively, while the coefficient of determination (R²) and outlier-robust coefficient of determination (D²) are increased by 0.118 and 0.176, respectively. The results indicate that the XGBoost ensemble learning model has better generalization ability, prediction accuracy, and interpretability.