Enhancing representation of data-scarce reservoir-regulated river basins using a hybrid DL-process based approach

Abstract

River discharge estimation in regulated river basins requires the inclusion of specific reservoir operation modules. However, human impacts remain challenging to depict in regions where upstream operational information (such as inflow and storage observations) is not available. Here, we develop a hybrid deep learning (DL)-process based approach that couples a conceptual hydrological model with simplified generic reservoir operation schemes and uses embedded neural networks (ENNs) to parameterize the conceptual model and optionally replace its reservoir operation module. We assess the ability of ENNs to compensate for the structural inability of simplified operation schemes to capture complex human impacts, while maintaining the advantage of minimal upstream operational record requirements. The hybrid models were tested across 43 regulated basins in the Continental USA in terms of their ability to simulate reservoir outflow, and evaluated in terms of their physical interpretability. Results show that the hybrid models outperformed both the conceptual and the LSTM models for outflow simulations, achieving a median NSE (KGE) of 0.648 (0.700) compared to 0.426 (0.415) for the conceptual models and 0.581 (0.636) for the LSTM model, with percentage improvements of 52.47 % (68.67 %) and 11.53 % (10.06 %) respectively. The dynamic parameterization by the ENNs compensates for the structural limitations of the simplified operation schemes to reproduce different operational patterns. Beyond their superior accuracy, the hybrid models also preserve physical interpretability, producing meaningful untrained internal variables such as inflow and evapotranspiration, and coherent parameters. The physical operation modules, while not improving final output accuracy, play a key role in supporting the physical interpretation of internal inflow processes. This highlights the importance of evaluating hybrid models comprehensively, rather than relying solely on final output performance. Our study offers deeper insights into hybrid modeling and provides a promising solution for system representation of data-scarce regulated river basins.