Bayesian optimization of stacking ensemble learning model for HPC compressive strength prediction
Qingfu Li, Xiang Wang
Abstract
This paper presents a new strategy that combines a Bayesian algorithm with stacked ensemble learning models to predict the compressive strength (CS) of high-performance concrete (HPC). Four machine learning methods-XGBoost (Extreme Gradient Boosting), Random Forest (RF), Gradient Boosting Decision Trees (GBDT), and CatBoost (Symmetric Decision Trees)-are employed to generate initial predictions. Bayes was used as a black-box algorithm to optimise the hyperparameters of a single model. The optimal data split ratio of 9:1 was tested, and the sensitivity ranking of input variables was determined using the Pearson correlation matrix. Stacking is an ensemble method that can improve prediction accuracy by integrating multiple base learners (ensemble models) in the first layer, and using a linear regression model (LR) as the second-layer meta -learner. The optimal base learner is identified through various configuration schemes. The proposed stacked model successfully combines the outputs from the base learners, resulting in significantly better prediction accuracy compared to previous studies. When applied to a new dataset, the method demonstrated strong performance. Furthermore, interactions between input variables and their effects on HPC were analyzed using Shapley Additive Explanations (SHAP) and Partial Dependence Plots (PDP). The results indicate that age, Water Reducing Agent (WRA), and Granulated Blast Furnace Slag (GGBFS) are the three most influential parameters affecting the compressive strength of HPC. This provides valuable insights for the research and application of high-performance concrete.