Abstract: During the operation of a dam, its original monitoring data exhibit complex, diverse, and time-varying characteristics, leading to gradual reduction in the effectiveness and accuracy of long-term monitoring warnings and thereby increasing disaster risks. Therefore, developing efficient and accurate deformation monitoring models is crucial to dam safety assessment. Traditional deterministic point predictions of a dam system, due to its inherent uncertainty, are faced with unavoidable challenges in error, bringing in low accuracy and a difficulty in determining the main factors of dam deformation. This paper presents a novel method that combines eXtreme Gradient Boosting with Bootstrap to construct prediction intervals. We use Elastic Net to extract the features of displacement influencing factors, and Bayesian Optimization to search for its optimal parameters. It can effectively estimate its own bias by combining multiple XGBoost models through Bootstrap; through residual training of the ensemble model, it further estimates the variance of random noise, quantifying the uncertainty of dam deformation. We validate this method in engineering case studies against the monitoring data from the Baihetan extra high arch dam under operation. Comparison of its predictions with the measurements and those predicted using a single model verifies its high accuracy and robustness, showing its root mean square error of only 0.0112. The accuracy of the model reaches 96%, and the efficiency is raised by up to 71% compared with the single model.

Key words: hydraulic engineering, deformation prediction, XGBoost, interval analysis, Bayesian optimization