Data-driven hierarchical Bayesian model for predicting wall deflections in deep excavations in clay

Abstract

The current paper presents a data-driven hierarchical Bayesian model (HBM) for predicting maximum lateral wall deflections in deep excavations in clay. The presented approach can address a number of challenges, including handling missing input parameters, incorporating the observational method, and accounting for site uniqueness. A new database, EXCA-CLAY/11/901, comprising 302 excavation sites worldwide, is compiled to train the HBM. The trained HBM is applied to a real case study in downtown Shanghai. The HBM can be applied to predict the maximum lateral wall deflection at each excavation stage independent of measured deflections from past stages (non-observational approach) or the more typical observational approach that includes past deflection measurements. Extensive cross-validation analysis validates the performance of the HBM and compares it with three existing regression models. The results show that the regression models exhibit slightly better performance within their applicable range when compared to the non-observational HBM prediction, but are outperformed by the observational HBM. The HBM is shown to offer several advantages over the conventional regression models, including the ability to produce 95% confidence intervals, handle missing input parameters, accommodate a diverse database, and continuously refine predictions as new information becomes available.