Research Information System
JOURNAL OF HYDROLOGY, vol.669, no.135002, pp.1-20, 2026 (SCI-Expanded, Scopus)
Hydrological modeling of karst systems is difficult due to their unique recharge, drainage and discharge behavior, which is often highly dynamic and nonlinear. It becomes even more challenging for elevated karst catchments, where the recharge process is additionally influenced by snow accumulation and melting. In this study, an innovative modelling approach was developed that hybridizing a process-based model and a data- driven model for the karst systems influenced by seasonal snow cover and its application was tested to a large, complex karst system in the Unica River catchment in Slovenia. For this purpose, the process-based model G´enie Rural `a 6 param` etres Journalier, including the CemaNeige snow routine (CemaNeige GR6J), was hybridized with the Stacked Autoencoder Deep Neural Networks (SAE-DNN). A 60-year period of catchment discharge observations, from 1962 to 2021, was used for model development, testing and evaluation. The performance of the stand-alone models, CemaNeige GR6J and SAE-DNN, as well as the hybrid model CemaNeige GR6J-SAE-DNN, was systematically compared. The results show that the hybrid model clearly outperforms both stand-alone models, especially during the extreme flow conditions. Additionally, the hybrid model performs better for more recent modelling periods than for longer ones. This is due to changes in climate conditions in historical datasets, which the hybrid model is limited to capture. Overall, the proposed hybrid modeling approach offers an innovative way to robustly predict the daily discharge behavior of karst systems influenced by seasonal snow cover, especially during extreme flow conditions, and could be applied to other karst systems with similar complexity and characteristics to support robust decision making in karst water resource management.