2021 Volume 2 Issue J2 Pages 872-882
This study tried to use an ensemble learning approach, stacking, to improve the accuracy of river flow estimation at a study watershed, the Tedori River watershed. Multiple deep learning methods (MLP, CNN, LSTM) with several combinations of hyperparameters were employed as weak learners of the ensemble learning. For the weak learners, the daily meteorological data and the daily average river flow discharge at the study watersheds were utilized as the input and target data, respectively. The outputs of the weak learners were utilized as input to the stong learner. As the strong learner, XGBoost was used. In this study, not only the entire datasets obtained from the weak learners, but also some parts of them were utilized as inputs to XGBoost. Then, the results were compared so as to investigate the effects of the selection of the weak learners on the modeling accuracy. The resutls indicate that the use of LSTM as weak learners has possibility to improve the modeling accuracy by ensemble learning.
