Machine learning for time series prediction of valley deformation induced by impoundment for high arch dams

Abstract

This study develops a data-driven hybrid model using machine learning methods to predict the valley deformation induced by impoundment for high arch dams. The elastic net and random forest are employed to identify and rank the key influencing factors of valley deformation as the best input features in the model. The variational mode decomposition (VMD) is introduced to decompose the original nonstationary valley deformation time series. The least square support vector machine (LSSVM) model is constructed to generate the predictions of valley deformation. The sparrow search algorithm (SSA) is utilized to find the optimal model parameters of LSSVM. A practical example involving the time series prediction of valley deformation for the Baihetan high arch dam in China is presented to validate the developed model. The developed model can generate long-term predictions of valley deformation efficiently and accurately based on the present monitoring valley deformation time series. Both the daily reservoir water level and 8-week accumulated rainfall show a significant influence on the valley deformation for the Baihetan high arch dam. The utilization of the VMD and SSA improves the accuracy of the developed model substantially. Thus, the developed VMD-SSA-LSSVM model produces better predictions of the fluctuation trend and turning points of the monitoring time series of valley deformation than the LSSVM model without the VMD and SSA. In addition, the role of VMD is more significant than that of SSA in improving the accuracy of the LSSVM model.