Performance evaluation of coupled conceptual and machine-learning frameworks for streamflow prediction in ungauged basins

This study explores the potential of a hybrid streamflow model that addresses the interpretability limitations commonly associated with the ‘black-box’ nature of machine learning models. The hybrid model simulates the rainfall-runoff process through its conceptual structure, integrating a deep-learning neural network to estimate the associated parameters. Using the Catchment Attributes and Meteorology for Large-sample Studies (CAMELS) dataset, we evaluated the model’s performance across 671 catchments in the United States, and compared it with the widely used Hydrologiska Byråns Vattenbalansavdelning (HBV) and the Long Short-Term Memory (LSTM) neural network. In gauged basins, where the three models were directly trained using runoff observations, the LSTM showed superior performance, achieving a median Kling-Gupta Efficiency (KGE) of 0.79. In comparison, the HBV model and the hybrid model attained median KGE values of 0.66 and 0.68, respectively. However, when the same catchments were treated as ungauged and runoff was predicted using regionalization approaches, the performance of all three models declined: the LSTM experienced a 17% reduction in KGE (0.79 → 0.66), while the hybrid and HBV models showed reductions of 13% (0.68 → 0.59) and 11% (0.66 → 0.59), respectively. The largest performance degradation observed in the LSTM underscores the advantage of the physical constraints inherent in the HBV and hybrid models, which help mitigate potential information loss. However, the hybrid model exhibited a ‘lower boundary problem,’ where it failed to generate hydrographs below a certain threshold. Although the hybrid model did not surpass the regionalized LSTM in performance, this study emphasizes the interpretability benefits offered by its conceptual structure. Furthermore, it highlights the hybrid model’s potential as an effective regionalization approach, combining the learnability of machine learning with the physical consistency of conceptual models.

 

Acknowledgements: this work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (RS-2024-00416443).