Abstract: For a dam curtain-grouting project, accurate prediction of its unit grout injection volume is crucial to engineering cost control and effective mitigation of its seepage-related risks. To address the complicated nonlinear relationship between its evaluation indicators and unit cement consumption, this paper presents a novel prediction model based on stacking ensemble learning, and compares it with the models of Random Forest (RF), XGBoost, and Support Vector Machine (SVM). We optimize the hyper-parameters of these models using a Newton-Raphson-based optimizer (NRBO), and test their predictive performance based on both training sets and validation sets. Results show the stacking model markedly outperforms the single-method models, with R2 = 0.998 and RMSE = 0.350 achieved on the training sets, and R2 = 0.971 and RMSE = 1.224 on the validation sets. By contrast, validation-set R2 values of the three single-method models are 0.905, 0.930 and 0.728 respectively, and RMSEs are 2.219, 1.896 and 3.748 respectively. Ensemble-learning models (Stacking, RF, and XGBoost) feature a stronger fitting capacity and robustness in the case of high-dimensional nonlinear data, while the stacking model, by leveraging the strengths of multiple base learners, further enhances predictive accuracy and robustness to outliers. Thus, our NRBO-Stacking model offers high accuracy, effective solutions, and better generalization performance for dam curtain-grouting projects.

Key words: dam engineering, curtain grouting, cement consumption prediction, random forest, NRBO algorithm, ensemble learning model