Prediction model for the compressive strength of rock based on stacking ensemble learning and shapley additive explanations

Accurately predicting the compressive strength of rock (RCS) is crucial for the construction and maintenance of rock engineering. However, RCS prediction based on single machine learning (ML) algorithms often face issues such as parameter sensitivity and inadequate generalization. To address these challenges, a new (RCS) prediction model based on a stacking ensemble learning method was proposed. This method combines multiple ML algorithms to achieve more accurate and stable prediction results. Firstly, 442 sets of rock mechanics experimental data were collected to form the prediction dataset, and data preprocessing techniques, including missing value imputation and normalization, were applied for data cleaning and standardization. Secondly, nine classic ML algorithms were used to establish RCS prediction models, and the optimal configurations were determined using k-fold cross-validation and Bayesian optimization. The selected base learners were LightGBM, Random Forest, and XGBoost, and the meta-learners were Ridge, Lasso, and Linear Regression. Finally, the models were verified using the testset, and the comparison showed that the proposed stacking models were better than all single models. Notably, the Stacking-LR model exhibited the best predictive accuracy(R2=0.946, MAE=5.59, MAPE=9.94%). Furthermore, the Shapley Additive exPlanations (SHAP) method was introduced to analyze the impact and dependencies of input features on the prediction results. It was found that both Young's modulus and confining pressure are the most critical parameters influencing RCS and exert a positive impact on the prediction results. This finding is consistent with domain expert knowledge, enhances the model's interpretability, and provides robust support for the predicted results.