Quantification of Uncertainty in the Hydrological Forecasting System of Salto Grande Reservoir

Hydrological forecast is a key product for the operation of reservoir systems. The anticipation of a given discharge enables a more efficient use of the water resource by taking early actions and adapting the daily operation of the system. However, hydrological forecasting has multiple sources of uncertainty that cause the deviation of the given forecast from the actual observation. The identification of the sources of uncertainty as well as its quantification therefore is a necessary step towards a better estimation of the forecast and the production of more reliable information for reservoir operation.

The sources of hydrological uncertainty are commonly divided in the initialization of the model, the meteorological uncertainty, and the model uncertainty (both structural and parametric uncertainty). This research aims at identifying which of these sources has a bigger role in the short-term hydrological forecasting system of Salto Grande hydropower plant. The forecasting system estimates the inflows of the reservoir using a complex cascade of processes including the Sacramento Soil Moisture Accounting model, unit hydrographs, Muskingum hydrological routing, and auto-regressive model corrections, among others. It also incorporates three sources of meteorological forcing variables, namely data coming from the Numerical Weather Prediction (NWP) systems GFS, GEFS, ECMWF. While the GFS contains a deterministic estimation, GEFS and ECMWF provide probabilistic estimates using an ensemble of forecasts (20 and 50 respectively).

We analyze deterministic and probabilistic daily forecasts produced at Salto Grande using a hindcasting experiment and obtain performance metrics based on the verification of the forecast. The system has been extended to include the use of observed precipitation as a forcing variable, therefore adding an important benchmark when comparing results using the NWP forecasts. Particularly, we used metrics such as Receiver Operating Characteristics (ROC) scores, Continuous Ranked Probability Skill Score (CRPSS), Brier Skill Score (BSS), Rank Histograms (RH), as well as Mean Absolute Errors (MAE), to identify the impact of meteorological numerical weather predictions in the production of hydrological forecasts in Salto Grande. This information provides important characteristics of uncertainty that will be used towards the improvement of the hydrological forecasting system.