Coupling Differentiable Modules of Reservoir Operation and Rainfall‐Runoff Processes for Streamflow Simulation

Abstract Reservoir operation modules are essential for hydrological modeling in human‐regulated catchments. This paper is concentrated on testing the coupling of reservoir operation and rainfall‐runoff processes under the framework of differentiable parameter learning (dPL). Specifically, the Community Water Model (CWatM)'s reservoir operation module is coupled with the Hydrologiska Byråns Vattenbalansavdelning (HBV) model; the differentiable fully coupled model (FCM) uses one long short‐term memory (LSTM) network to calibrate all parameters using outflow; and the differentiable loosely coupled model (LCM) sets up two LSTM networks respectively for inflow and outflow. For comparison, the differentiable HBV is calibrated by outflow. The results of 77 reservoirs highlight that the dPL is effective in improving the efficiency of the conventional models. The median Kling‐Gupta efficiency is improved from 0.53 for HBV to 0.59 for differentiable HBV, from 0.52 for FCM to 0.61 for differentiable FCM and from 0.54 for LCM to 0.60 for differentiable LCM. Zooming into the hydrological processes, it is found that the differentiable HBV fits reservoir outflow by underestimating recession coefficients and overestimating the baseflow index. The differentiable FCM fits the outflow but not the inflow since it tends to overestimate the maximum storage of the upper soil layer. The differentiable LCM fits both inflow and outflow with one LSTM estimating the parameters of HBV and the other LSTM estimating those of CWatM's reservoir operation module. For ungauged catchments, the differentiable LCM outperforms differentiable FCM in reproducing inflow and outflow. Overall, the dPL is effective in simulating the hydrological processes for human‐regulated catchments.