Accurate prediction of rockburst is fundamental for mitigating rockburst hazards and holds significant importance for the safety of deep underground engineering. A data set containing 241 rockbursts was constructed. Uniaxial compressive strength (UCS), uniaxial tensile strength (UTS), brittle-coefficient (BI), maximum tangential stress (MTS), elastic energy index (W_et), and the ratio of maximum tangential stress to uniaxial compressive strength (SCF) were selected as the prediction indexes of rockburst. Statistical analysis of the original dataset initially identified outliers through coefficient of variation. The t-distributed stochastic neighbor embedding(t-SNE) method was employed to reduce data dimensionality to three dimensions for visualization, which intuitively revealed data imbalance and outliers. Subsequently, Borderline-SMOTE oversampling was applied to balance the dataset, followed by Yeo-Johnson transformation for data normalization. Through different data preprocessing strategies, three random forest models were developed: the RF model, SMOTE-RF model, and Borderline-SMOTE-RF model. Grid search combined with 10-fold cross-validation was implemented for parameter optimization, and model performance was evaluated using confusion matrices and multiple metrics. Results demonstrate that the Borderline-SMOTE-RF model achieves the highest prediction accuracy of 82.7%. The proposed model was ultimately applied to rockburst prediction in the headrace tunnel of Jiangbian hydropower station, showing high consistency with actual situation. The research findings can provide valuable insights for rockburst disaster prediction.