ML-derived reservoir operations for 24,000 dams implemented in a global hydrological model

Globally there are over 24,000 dams that greatly alter river connectivity and streamflow regimes of the world’s large rivers. To capture the impact of dams, global hydrological models implement simplified reservoir operations that use modelled inflow, static storage capacity values, and downstream demand to calculate reservoir releases and better understand large-scale streamflow dynamics. According to Steyaert et al., 2023, these approaches typically overestimate the amount of water stored in reservoirs, smooth out the seasonality in storage, and may miss long term trends. All of which can underestimate the impact on streamflow regimes as the operations are not necessarily derived from historic time series. To assess the importance of reservoir operations on global streamflow regimes, we update the number of reservoirs in the PCRGLOBWB 2.0 hydrologic model from 6,000 to 24,000 using the georeferenced global dams and reservoirs dataset (GeoDAR (Wang et al., 2022)) and derive dynamic storage thresholds using freely accessible remotely sensed storage data and a new reservoir algorithm developed by Turner et al., 2021. We obtained the reservoir specific parameters required for the Turner algorithm using a ML based approach to enable global simulations of all 24,000 dams. We also employ a sensitivity analysis across multiple command areas (250, 650 and 1100 km) to assess the impact reservoirs have on the global streamflow and downstream water demand. Preliminary results in the Rhine basin show that increasing the number of dams and using data derived methods provides more realistic streamflow regimes. We observe an improvement in the KGE of discharge simulations from 0.43 to 0.67, also reducing the bias from -2012.23 to -316.94 compared to the old reservoir implementation currently used in PCR-GLOBWB 2.0. This significant improvement in model performance highlights the importance of observation derived rule curves for reservoir management in global hydrological models.