A similarity-aware ensemble method for displacement prediction of concrete dams based on temporal division and fully Bayesian learning

Precisely predicting concrete dam displacements is crucial for assessing their structural behavior during operation. Many studies have testified that ensemble methods are more accurate and applicable in practice than individual predictive models. Nevertheless, the common way handling massive monitoring data is still conventional, that is, training and testing them as a whole, neglecting the internal law and pattern difference within data, which probably limits advancements in predictive effect. To this end, the patterns of monitoring data are identified and classified before model establishment, and a similarity-aware ensemble method (SAEM) using temporal division and fully Bayesian learning is presented for dam displacement prediction. Specifically, the unsupervised fuzzy C-means approach and sparrow search algorithm are fused for similar pattern clustering of environmental factors, thus achieving temporal division in displacement responses. Fully considering the adaptability of model structure and parameters to various data patterns, a non-parametric fully Bayesian Gaussian process regression (FBGPR) model is proposed by augmenting the standard GPR with Markov chain Monte Carlo simulation and Bayesian evidence evaluation theory. Different data clusters are then fed into FBGPR in chronological order, and the final results are derived through a grouping ensemble scheme. Multiple sets of monitoring data collected from a real-world dam project are employed for method verification. The results show that our proposed SAEM has better prediction accuracy compared to homogeneous clustering-based ensemble methods and commonly used individual models. The superior performance in two additional cases also verifies the adaptability and generalization ability of our method, which provides a new modeling tool for structural health monitoring of concrete dams.