Examining the applicability of different sampling techniques in the development of decomposition-based streamflow forecasting models

Abstract The applicability of the traditionally used overall decomposition-based (ODB) sampling technique in the development of forecasting models is controversial. This study first conducts a systematic investigation of the performance of models developed using the ODB sampling technique. A stepwise decomposition-based (SDB) sampling technique that is consistent with actual forecasting practice is then proposed. Moreover, a novel calibration algorithm that couples a two-stage calibration strategy with a shuffled complex evolutionary approach is formulated to help maintain the performance of models. The application of models produced using these different sampling techniques to four gauging stations in China and Canada indicates that (1) the ODB sampling technique that employ the discrete wavelet transform (DWT), empirical mode decomposition (EMD) and variational mode decomposition (VMD) as series decomposition techniques do not produce convincing forecasting models because additional information on the future streamflow that is to be predicted is introduced into the explanatory variables of the samples; (2) the SDB sampling technique strictly excludes information on future streamflow from the explanatory variables and is thus as an appropriate alternative for developing forecasting models; (3) the DWT and VMD techniques benefit models by enhancing their performance; on the other hand, the EMD is unsuitable for use in forecasting, due to the variable number of subseries that result from the implementation of the stepwise decomposition strategy. Finally, methods that can be used to enhance the performance of decomposition-based models and the prediction accuracy of nonstationary streamflow are suggested.