Forecasting Daily Seepage Discharge of an Earth Dam Using Wavelet–Mutual Information–Gaussian Process Regression Approaches

Abstract

Because of their sensitive structure, earth dams might face failure due to seepage phenomenon. In order to prevent such failure, some equipment like piezometers are installed in the body or foundation of earth dams. This study investigated the importance of piezometer installation level in dam body or foundation using mutual information–wavelet–Gaussian process regression. 27 Piezometers in three section along with reservoir level were employed to predict one-step-ahead seepage discharge of Zonouz earth dam. The daily data of 1 year of piezometer level and reservoir level were collected for this purpose. In order to find the best possible input combination, three groups of modeling scenarios were defined using piezometers and reservoir level time series. As some input combinations had more than two variables, decomposed time series were imposed into mutual information (MI) tool in order to decrement input variables and find the most correlated input–output features. Afterward, mentioned features were imposed into optimized Gaussian process regression (GPR) to be predicted. Different kernels were selected as core tool of GPR, but results demonstrated the capability of radial basis function (RBF) kernel. GPR–RBF structure were optimized using cross-validation technique. Results indicated that input combination including piezometer level and reservoir level of section II, especially piezometer 203 time series led to the best result among all scenarios.