Hydrological Modelling of Karst Catchment Using Lumped Conceptual and Data Mining Models

Abstract

Hydrological modelling is a challenging and significant issue, especially in nonhomogeneous catchments in terms of geology, and it is an essential part of water resources management. In this study, daily rainfall-runoff modelling was carried out using the lumped conceptual model, the artificial neural network (ANN), the deep-neural network (DNN), and regression tree (RT) data mining models for the nonhomogeneous karst Ljubljanica catchment and four of its sub-catchments in Slovenia with different geological characteristics. Model performance was evaluated using several performance criteria and additional investigation of low and high flows was carried out. The results of the study indicate that the Génie Rural à 4 paramètres Journalier (GR4J) lumped conceptual model yielded better modelling performance compared to the data-driven models, namely ANN, DNN and RT models. Moreover, the enhanced version of the GR4J model (i.e. GR6J) also yielded good performance in terms of the recession part. The RT model yielded the worst performance regarding runoff forecasting among the examined models in the case of all five investigated catchments. However, ANN and DNN data-driven models were slightly more successful in modelling the hydrograph recession in the case of karst sub-catchments compared to the GR4J lumped conceptual model structure. Inclusion of additional meteorological variables to ANN and DNN does not significantly improve modelling results. © 2019 Elsevier B.V.