New Hybrid GR6J-Wavelet Genetic Algorithm-Artificial Neural Network (GR6J-WGANN) Conceptual-Data Model Approaches for Daily Rainfall-Runoff Modelling

Abstract

Rainfall-runoff modeling is significant for efficient water resources management and planning. The hydrological conceptual models can have challenges, such as dealing with nonlinearity and needing more data, whereas data-driven models are generally lacking in reflecting the physical process in the basin. Accordingly, two-hybrid model structures, namely Genie Rural a 6 parametres Journalier (GR6J)-wavelet-based-genetic algorithm-artificial neural network(1) (GR6J-WGANN(1)) and GR6J-wavelet-based genetic algorithm-artificial neural network(2) (GR6J-WGANN(2)) models, were proposed in this study to develop rainfall-runoff modeling performance. The novel GR6J-WGANN(1) model used the routing store outflow (QR), exponential store outflow (QRexp), and direct flow (QD) obtained from the GR6J, and the GR6J-WGANN(2) model used the soil moisture index (SMI) obtained from the GR6J as input data. The wavelet transformation and Boruta algorithm were implemented to decompose the input data into components and select important wavelet components, respectively. The performance of the GR6J, standalone WGANN models, and hybrid models were tested in three sub-basins of Konya Closed Basin, Turkey, which generally has arid and changing climate conditions. The hybrid models performed better than the conceptual and data-driven models, particularly regarding the extreme flow predictions. Using soil moisture index, routing store outflow, exponential store outflow, and direct flow as the output of the GR6J in GR6J-WGANN(1) and GR6J-WGANN(2) improved the rainfall-runoff modeling performance remarkably. The findings of this study indicated that hybrid models, which integrate strong sides of conceptual and data-driven models, can be more useful for producing more accurate forecasting results.