Abstract: Deformation as an intuitive monitoring indicator reflects the operation state of a dam. Existing intelligent methods for dam deformation prediction are prone to local optimum and are inapplicable in the case of large-scale data. In this paper, we combine a fast and efficient gradient boosting framework Light Gradient Boosting Machine (LightGBM) based on the decision tree with the global optimization algorithm, Bayesian optimization, to predict dam deformation. Taking two concrete dams as case study, we compared the modelling result with those of multiple linear regression, support vector regression, and multi-layer perceptron to verify the applicability of the present model. The RMSE and MAE of the proposed model are both superior to those of other methods, which manifests the feasibility and superiority of the model. In addition, LightGBM can evaluate the importance of input parameters and select the features that affect dam deformation, so as to determine the factors that have more significant influence on dam deformation and offer reference for following safety assessment.

Key words: dam deformation prediction, Bayesian optimization, LightGBM, multiple linear regression, support vector regression, multi-layer perceptron