A Comparative Study on Modeling Methods for Deformation Prediction of Concrete Dams

A series of machine learning models have been proposed in the past decades, but it remains undetermined which is optimal for specific applications. Establishing mathematical prediction models for dam deformation and structural health monitoring based on environmental factors is crucial to dam safety assessment. This paper takes Zhexi Dam, a concrete gravity-type dam in China, as an example to conduct a comparative study on the performance of deformation prediction models. The physical factors that cause dam deformation include the air temperature, reservoir water temperature, reservoir water level, and dam aging. The correlations between environmental factors and dam deformation are evaluated by maximum information coefficient (MIC) and Pearson, Kendall, and Spearman correlation coefficients. The monitoring data reveal that the deformation has a high correlation with environmental factors. A number of the most representative monitoring points from hundreds of monitoring points are selected for modeling. For comparison, seven modeling methods, i.e., multiple linear regression (MLR), gradient boosting decision tree (GBDT), random forest (RF), support vector machine (SVM), and long short-term memory network (LSTM), weighted average model (WAM) of the above five algorithms, and Transformer-based neural network, are introduced to establish dam deformation prediction models. The experimental results indicate that both the weighted average model and the Transformer-based neural network achieve consistently high accuracy, showing strong agreement with the monitoring data generally. However, in scenarios involving small sample sizes, the SVM model demonstrates relatively superior predictive performance compared to the other models.