Diversity-driven ANN-based ensemble framework for seasonal low-flow analysis at ungauged sites

Abstract

Low-flow estimation at ungagged sites is a challenging task. Ensemble-based machine learning regression has recently been utilized in modeling hydrologic phenomena and showed improved performance compared to classical regional regression approaches. Ensemble modeling mainly revolves around developing a proper training framework of the individual learners and combiners. An ensemble framework is proposed in this study to drive the generalization ability of the sub-ensemble models and the ensemble combiners. Information mixtures between the subsamples are introduced and, unlike common ensemble frameworks, are explicitly devoted to the ensemble members as well as ensemble combiners. The homogeneity paradigm is developed via a two-stage resampling approach, which creates sub-samples with controlled information mixture levels for the training of the individual learners. Artificial neural networks are used as sub-ensemble members in combination with a number of ensemble integration techniques. The proposed model is applied to estimate summer and winter low-flow quantiles for catchments in the province of Québec, Canada. The results show significant improvement when compared to the other models presented in the literature. The obtained homogeneity levels from the optimum ensemble models demonstrate the importance of utilizing the diversity concept in ensemble learning applications.