Abstract: Using historical data from blasting operations of open-pit mines to predict the average rock fragment size is crucial for optimizing blasting parameters. However, existing methods face 3 major challenges: interference from high-dimensional input features, low computational efficiency, and difficulties in modeling with sparse datasets. This study therefore proposes a hybrid prediction model that integrates Random Forest (RF), Whale Optimization Algorithm (WOA), and Extreme Gradient Boosting (XGBoost). Specifically, the original rock fragment size dataset was subjected to multi-level preprocessing to enhance data quality; RF was employed to evaluate and select 19 input features to reduce dimensionality. WOA was integrated to intelligently optimize the hyperparameters of the prediction model; XGBoost was used to model the small-sample rock fragment size dataset. Comparative experiments showed that this model exhibited better prediction performance with an R² value of 0.93, outperforming other control group models. Additionally, the clear modeling process design further enhanced the operability and engineering application of the prediction model.