How can observational data be used to improve the modeling of human-managed reservoirs in large-scale hydrological models?

Abstract. Human-managed reservoirs alter water flows and storage, impacting the hydrological cycle. However, modeling reservoir outflow and storage is challenging because it depends on human decisions, and there is often limited access to data on inflows, outflows, storage, or operational rules. Consequently, large-scale hydrological models either exclude reservoir operations or use calibration-free algorithms for modeling reservoir dynamics. Nowadays, remotely-sensed information on reservoir storage anomalies is a potential source for calibrating reservoir operation algorithms. However, it is not yet clear what impact calibration against storage anomalies has on simulated reservoir outflow and absolute storage. In this study, we introduce two reservoir operation algorithms that require calibration: the Scaling Algorithm (SA) and the Weighting Algorithm (WA). These algorithms were implemented in the global hydrological model WaterGAP and compared with the widely used Hanasaki algorithm with both default (DH) and calibrated (CH) parameter values. We calibrated all three algorithms against outflow, storage, storage anomalies, and estimated storage (based on storage changes and reservoir capacity) observed for 100 reservoirs in the USA to understand the information content of the observation variables. As expected, calibration against all three types of storage-related variables improved the storage simulation. Storage simulation using DH resulted in only 16 (15) skillful simulations (where the Kling–Gupta Efficiency with a trend component > -0.73) out of 100 reservoirs. In contrast, calibration against storage anomalies resulted in 64 (39), 68 (45), and 66 (45) skillful storage simulations for CH, SA, and WA, respectively, during the calibration (validation) period. However, calibration against storage-related variables barely improved the performance of the outflow simulation, which strongly depends on the accuracy of the simulated inflow. In fact, using observed inflow instead of simulated inflow has a more significant effect on improving outflow simulation than calibration, whereas the opposite is true for storage simulation. We found that the default parameters of the Hanasaki algorithm rarely matched the calibrated parameters, highlighting the benefit of calibration. Moreover, taking into downstream water demand in the reservoir operation algorithm does not necessarily improve modeling performance due to high uncertainty in demand estimation. Overall, the SA algorithm outperforms the other algorithms. Therefore, to improve the modeling of reservoir storage and outflow, we recommend calibrating the SA reservoir operation algorithm against remote sensing-based storage anomalies and improving reservoir inflow simulation.