A data derived workflow for reservoir operations for simulating reservoir operations in a global hydrologic model

Abstract. Globally there are over 24,000 storage structures (e.g. dams and reservoirs) that contribute over 7,000 km³ of storage. Until recently, most of the data regarding these reservoirs was not openly accessible. As a result, many studies rely on generalized operations based on generalized assumptions about reservoir storage dynamics and management. With the creation of global datasets such as the Global Reservoirs and Dams (GRanD), RealSat, GloLakes, and the International Coalition for Large Dams database (iCOLD) as well as localized datasets such as ResOpsUS for the contiguous United States, and the Mekong Data Monitor for the Mekong River basin, the inference of reservoir operations using data derived techniques has become much more ubiquitous regionally. Yet to our knowledge, there has been no global application of data-derived methods due to their model complexities and data limitations. Therefore, our analysis aims to fill this gap by providing a workflow for implementing data derived reservoir operations in the large scale hydrologic models with an application in the PCRGLOBWB 2 global hydrologic model. This methodology uses global satellite altimetry data from GloLakes, a parameterization methodology developed by Turner et al. (2021), and a random forest model. We also test the sensitivity of our reservoir scheme to downstream demand by selecting three different downstream areas presumed to be served by reservoirs release (command areas): 250 km, 650 km, and 1100 km. Our results demonstrate that our random forest algorithm is able to capture the storage dynamics and that the errors are mainly due to the errors in using remotely sensed storage data. Additionally, we observe in many cases that deriving operational bounds from historical reservoir time series has minimal impact on streamflow at the basin outlets nor is the scheme sensitive to the downstream command areas. We do observe that streamflow is affected directly downstream from the reservoirs and that the data-derived methodology does increase the accuracy of simulated global reservoir storage when compared to observations. In fact, the derived operations have much lower storage values that align better with both direct and remotely sensed reservoir storage observations. This demonstrates that generic operations overestimate the total amount of water stored in reservoirs and, as a result, are potentially overestimating water availability. Ultimately, our workflow allows global hydrologic models to capitalize on recent data acquisition by remote sensing to provide more accurate reservoir storage and global water security.