<p>Rockburst, as a typical dynamic disaster in deep underground engineering, poses significant challenges to construction safety control. To achieve real-time and accurate rockburst prediction, this study proposes a rockburst grade prediction method based on microseismic parameter optimization and improved XGBoost. First, feature mining was conducted on 136 sets of multi-source engineering case data, and four microseismic parameters—cumulative event count, energy, apparent volume, and microseismic event rate—were selected as prediction indicators through mutual information analysis and information gain calculation. To address the parameter sensitivity issue of the XGBoost model, the whale optimization algorithm (WOA) was introduced for adaptive hyperparameter tuning, establishing a WOA-XGBoost prediction model that integrates feature selection and parameter optimization. The accuracy and reliability of the model were validated from three aspects: model evaluation, model comparison, and engineering verification. The results demonstrate that the improved model achieves a prediction accuracy of 89.28%, representing a 7.15% improvement over the unoptimized XGBoost model, while also outperforming other benchmark algorithms in key metrics such as precision and recall. Engineering validation using four case studies, including the Gaofeng Mine and Qinling Tunnel, showed highly consistent predictions. This method provides a scientific reference for short-term rockburst prediction in deep underground engineering projects.</p>