MR and stacked GRUs neural network combined model and its application for deformation prediction of concrete dam

Abstract

Deformation prediction is crucial to dam safety management and economical operation. However, most existing prediction techniques, such as statistical methods, support vector machine based models and other traditional machine learning methods, fail to consider the dynamic evolution characteristics of dam deformation and just treat this process as a problem of static fitting, which leads to the fact that these models not only have limited prediction accuracy and generalization ability, but also can hardly carry out long-term prediction. This study proposed a statistical-deep learning combined strategy to improve the accuracy and stability of concrete dam deformation prediction and realize long-term prediction. The multiple regression (MR) model is utilized to decompose the observed displacement into four components: hydrostatic, temperature, aging and reminder. Then the stacked gated recurrent units (GRUs) neural network is employed to explore the berried chaotic and dynamic change mechanism in the remainder component and make prediction based on historical data. Meanwhile, the other three components are predicted by the fitted MR model. Finally, the predicted displacement is obtained by aggregating the results of the MR model and the stacked GRUs neural network. A large hydropower dam on the Yellow River is introduced to demonstrate the application of the proposed strategy. Results show that the model proposed in this paper achieves excellent performance in both prediction accuracy and stability in long-term prediction tasks. In a one-year arc dam deformation prediction, the relative root mean square error of the proposed model is as small as 1.34%, and the R2 is as big as 0.9981. The proposed model has certain engineering application value in the long-term prediction of concrete dam deformation.