A multi-step strategy for enhancing the rainfall-runoff modeling: combination of lumped and artificial intelligence-based hydrological models

Abstract

Accurate rainfall-runoff (RR) modeling holds significant importance in environmental management, playing a central role in understanding the dynamics of water cycle. In this respect, the precision in the determination of RR is crucial for mitigating the adverse effects of both water scarcity and excessive runoff, ensuring the sustainable management of ecosystems and water resources. As a primary hydrological variable, runoff engages in direct interactions with other hydrological variables. Due to the complexity of the RR process, two primary approaches are commonly used in modeling, namely conceptual (lumped) models and artificial intelligence (AI) models. Conceptual approaches are based on hydrological processes and use a larger number of hydrological variables, yet they often exhibit lower performance compared to AI models. In contrast, AI models rely on fewer parameters and lack physical interpretability, but demonstrate high performance. This study merges the advantages of both lumped and AI techniques to develop an advanced RR model. Hence, the applicability of several lumped and AI-based models in estimating the streamflow rates with the help of basic meteorological variables is investigated. The lumped hydrological models, namely the Modello Idrologico SemiDistribuito in continuo (MISD), Identification of Unit Hydrographs and Component Flows from Rainfall, Evaporation, and Streamflow (IHACRES), and G & eacute;nie Rural & agrave; 4 param & egrave;tres Journalier (GR4J), are employed in conjunction with AI algorithms as Radial Basis Function (RBF) neural networks, Adaptive Neuro-Fuzzy Inference System (ANFIS) and Multilayer Perceptron (MLP). An ensemble of conceptual models (MISD, IHACRES, and GR4J) and three AI models (MLP, RBF, and ANFIS) with various lag times are considered as effective variables, where Support Vector Machine (SVM) was utilized as a feature selection method with five different kernels in determining the best inputs. Afterward, the SVM-ANFIS model, as the best model, is hybridized with Ant Colony Optimization (ACO) to develop the SVM-ANFIS-ACO model. It is found that the coupling of lumped and AI methodologies considerably enhanced the accuracy of the RR models; and SVM-ANFIS-ACO outperformed other models in streamflow computation.