A comparative assessment of a hybrid approach against conventional and machine-learning daily streamflow prediction in ungauged basins

This study focused on 671 ungauged basins across the contiguous United States, representing diverse climatic and physiographic setting, and leveraged the Catchment Attributes and Meteorology for Large-sample Studies (CAMELS) dataset. We assessed the applicability and potential of a hybrid model that couples a differentiable Parameter Learning (dPL) with a conceptual HBV model. The framework is evaluated in two ways: first, by benchmarking its predictive performance against a traditional regionalized HBV and a standalone LSTM model; and second, by exploring its utility as a diagnostic tool to understand the root causes of model failure modes based on hydrological processes understanding. The regionalized LSTM model yielded the highest predictive accuracy (mean KGE of 0.57), outperforming both the hybrid (0.41) and HBV (0.46) models. We revealed distinct failure patterns for each model, which were most pronounced in arid regions. The hybrid model’s primary failure mode—low-flow truncation—was attributed to a combination of two factors: systematic biases in the dPL’s parameter estimation and the inherent structural deficiencies of the HBV model’s groundwater module. Although the hybrid model did not surpass the standalone approaches in predictive performance, our findings suggest that the multifaceted advantages of the hybrid framework may offer a promising pathway to bridge the gap between learnable, black-box approaches and conceptual methods rooted in hydrological understanding. • Hybrid dPL–HBV tested on 671 CAMELS basins with 10-fold PUB splits. • Regional LSTM was most accurate for PUB with mean KGE of 0.57. • dPL regionalization matched proximity-based HBV performance. • LSTM missed peaks, HBV stretched recessions, hybrid cut low flows. • Hybrid revealed parK2 bias and HBV groundwater limits in low flows.