Integrating Reservoir Dynamics Into Differentiable Process‐Based Hydrological Model for Enhanced Streamflow Estimation

Reliable hydrological predictions are crucial for water allocation, reservoir operations, and flood control. While data‐driven models like Long Short‐Term Memory (LSTM) offer high accuracy, they lack physical interpretability and cannot estimate internal hydrological states or reservoir‐specific variables. To overcome these limitations, we propose a differentiable framework, dPLHBVRes, which integrates a simplified process‐based model (HBV) enhanced with a reservoir module and neural network‐based parameterization. This approach improves streamflow prediction while enabling the estimation of untrained hydrological variables, including evapotranspiration, soil moisture, reservoir storage, and outflows. For 38 regulated catchments in Peninsular India, dPLHBVRes achieved streamflow accuracy comparable to LSTM (median NSE 0.66 vs. 0.67), but with the added advantage of enabling investigation of internal states. Furthermore, in the absence of observed reservoir‐specific inputs, incorporating remote sensing‐based observations, such as water spread area, into dPLHBVRes further enhanced model performance. For instance, in the Rengali catchment, this additional input improved streamflow prediction from NSE 0.65 to 0.74 while also enhancing simulations of untrained variables: evapotranspiration (bias reduced from 0.46 to 0.39 mm), soil moisture (bias reduced from −21.42 to −7.34 mm), reservoir outflows (NSE from 0.42 to 0.49), and reservoir storage (correlation from 0.37 to 0.74). Overall, dPLHBVRes offers a practical and interpretable alternative to black‐box models for regulated catchments, providing accurate streamflow predictions while maintaining physical interpretability through access to internal untrained hydrological variables.