Improving the Prediction Capacity of a Hybrid Method of SWAT and LSTM by Reducing Parameter Uncertainty.

For effective water resources management, a hybrid modeling approach that couples process-based models (PBMs) with data-driven models (DDMs) has been proposed. The common approach is to use the outputs of PBMs as input for DDMs. Because the performance of the hybrid method depends on the prediction accuracy of PBMs, the uncertainty in PBMs should be carefully considered. Remotely sensed evapotranspiration (RS-ET) has been often invited as additional constraint of PBMs to reduce their parameter uncertainty. This study tested whether the parameter uncertainty of PBMs affects the hybrid model accuracy. In this study, Soil and Water Assessment Tool (SWAT) and Long-short term memory (LSTM) were selected as a representative PBM and DDM, respectively. At first, SWAT was calibrated against streamflow and suspended solids and these SWAT outputs were assumed to include parameter uncertainty. SWAT was calibrated against streamflow, suspended solids, and further RS-ET, which were viewed as SWAT outputs with less parameter uncertainty. Then, we applied two SWAT outputs into LSTM to explore predication capacity of two hybrid models on suspended solids according to presence and absence of uncertainty. The hybrid model results with consideration of SWAT uncertainty showed 0.75 and 0.62 KGE values, which were averagely 0.1 and 0.3 greater than those with disregarding uncertainty for the training and validation periods, respectively. The outcomes of this study demonstrated that the additional constraint of PBMs can reduce the uncertainty of parameters and contribute to improving the accuracy of the hybrid model.