Using GRACE TWS to predict reservoir height in the Upper Parana River Basin

All over the world, water levels are constantly changing. From lakes to rivers to oceans, the patterns of the water levels change due to different factors. With hydrological extremes increasing in intensity and duration around the world, it is important to understand what changes these levels in order to better predict and mitigate the negative impacts of changing water levels. We use estimates of terrestrial water storage (TWS) variability from the Gravity Recovery and Climate Experiment (GRACE) satellite missions to predict reservoir operation in Brazil. To do this, reservoir water elevations are derived from multi-satellite radar altimetry (RA) data and used as a proxy of their operation. 16 reservoirs in Southern Brazil are analyzed.

For each reservoir, the Pettitt test was used to identify the point break within the TWS data, and the Mann-Kendall test was used to identify trends before and after these breaks. A machine learning approach was used to reconstruct RA-based water elevations using GRACE data. The approach considered numerous geomorphologic and meteorologic characteristics of reservoir including precipitation (from GPM IMERG) and temperature (GLDAS). For some of the reservoirs, various ML models were run with a 5-day forecast horizon and the outputs were compared to determine which model predicted most accurately for each reservoir. Some of the models incorporated in this study include decision tree regressor, kernel ridge model, linear regression, random forest regression, and support vector regression. The findings of this study will give insight into what variables affect the relationship be-tween TWS and RA height in the Upper Parana Basin in Southern Brazil to improve prediction measures for reservoir height.