Prediction of wall deflection and ground settlement during deep excavation using hierarchical Bayesian model-based stacking ensemble learning framework

This study presents a hierarchical Bayesian model-based stacking ensemble learning framework (HBM-based SELF) for predicting the wall deflection and ground settlement during deep excavation in soft clays. The framework integrates twelve base learners of varying complexity, combining three model structures (a generic model, the original HBM, and a HBM with a generic covariance) with database pre-processing methods [tailored clustering enabled regionalisation (TCER) and dimension reduction (DR)]. The best learner is selected as the learner that produces the least predictive entropy. Validation using the Formosa case in Taipei shows that the HBM-based SELF tends to yield more accurate predictions for the wall deflection and ground settlement at each excavation stage as compared to single base learners. Extensive leave-one-site-out cross-validation further demonstrates its superiority over the original HBM. Comparisons with regression models in the literature also confirm that the HBM-based SELF provides more accurate predictions of wall deflection and ground settlement.