River water temperature prediction using REMD, RLMD, and SVMD coupled with deep learning models

Abstract

River water temperature (T w ) is a critical parameter for the freshwater ecosystems and the aquatic life. Daily water temperature time series can be used for many applications and included as input variable for various modelling tools. In this study, four deep learning (DL) models (i.e., LSTM, BiLSTM, GRU, and BiGRU) were used for modelling daily water T w measured at four rivers stations in Poland. Furthermore, a new method is adopted in the present study based on signal decomposition (SD) for improving the performances of the single DL models. Three different SD algorithms were used in the present study namely: robust empirical mode decomposition (REMD), robust local mean decomposition (RLMD), and successive variational mode decomposition (SVMD). For each DL we have adopted four scenarios: ( i ) using only air temperature (T a ) as input variables, and ( ii ) by using the product functions (PFs) obtained using the RLMD and the intrinsic mode functions (IMFs) obtained using the REMD and SVMD. All models were first calibrated using 70% of the dataset and validated using 30% of the dataset, and their performances were evaluated using the root mean squared error (RMSE), the mean absolute error (MAE), the coefficient of correlation (R), and the Nash-Sutcliffe efficiency (NSE). In general, the smallest RMSE and MAE are obtained for the models using the signal decomposition, whereas the largest errors metrics are achieved by the single DL models. Among all stations, the best performances were achieved using the REMD-BiGRU with R, NSE, RMSE, and MAE of 0.986, 0.973, 0.966, and 0.752, respectively, while the poorest performances were obtained by the GRU with R, NSE, RMSE, and MAE of 0.815, 0.659, 3.456, and 2.613, respectively.