Developing a seamless medium- to-long-range flow forecast to improve the prediction of hydropower generation in Brazil

Abstract

The management of water resources is of great importance for many human activities, especially in the operational context of the Brazilian electricity sector, where we have a predominance of hydroelectric generation. In this context, streamflow forecasts are the main source of information for the optimization of the electric system. This PhD thesis describes a series of steps developed to deal with uncertainties in the forecasts, in order to obtain reliable flow predictions. This research work aims to develop and evaluate a chain of data and models for short (several days) to long (several months) term hydrometeorological forecasting, which can be used for forecasting the hydroelectric production in Brazil. The study is based on 41 river basins, which represent 30 hydroelectric plants in South America, with high interest for energy production. For the uncertainty in the observed precipitation data, we investigated the differences between the TRMM_MERGE and CPC datasets, considering different resolutions (daily, monthly and annual precipitation). Substantial differences were found between the two data sources, which seem to be amplified in the period 2008-2017. A spatial trend was found, with TRMM-MERGE precipitation values higher than those of the CPC dataset when moving towards north and west in the study area. Based on these observed uncertainties, and seeking a better estimate of precipitation, the combination of these two sources was evaluated. Inspired by the modified water balance equation, uncertainties involving the precipitation datasets were identified and quantified. Hydrological modeling was also used to choose and validate the precipitation datasets. The results indicated that the combination of real-time precipitation (TRMM-MERGE and CPC), weighted by the uncertainty of the original sources, outperforms the isolated use of only one of the data sources. Another source of uncertainty analyzed was the precipitation forecast. In this step, the performance of two different bias correction methods, QM - Quantile Mapping and LS - Linear Scaling, was compared. In terms of seasonal precipitation forecast (up to 7 months of forecast horizon), the results indicated that the errors observed in the raw forecasts are more dependent on the month of the year than on the forecast horizon, with systematic overestimation during the rainy season and underestimation observed during the dry season for most of the basins studied. The bias correction methods were effective, especially during the rainy season, with the QM method showing better performance. Based on the good performance obtained with this method, the bias correction was applied to the medium (up to 45 days) and short (up to 15 days) term forecasts of the European center ECMWF, but with a different application. At each initialization (time when a forecast is issued), the medium-term model also generates the reforecast of the last 20 years, for the same 46 calendar days. Thus, the parameters for the QM correction were recalculated at each initialization, generating an on-the-fly correction, which depends only on the forecast horizon. These same parameters were used to correct the short-term model, and the results obtained were equal to or better than the results obtained with the parameters calculated only with the historic time series of short-term reforecasts. This method is advantageous as it does not need a long time series of reforecasts to calibrate the bias correction parameters, which allows to better follow the evolution of meteorological models (in the case of this thesis, the ECMWF models). To build the precipitation forecast in a continuous (seamless) way, based on the coupling of the three forecast horizons of the ECMWF models (15 days, EPS model; 45 days, extended model and 7 months, seasonal model), several coupling methods were tested. The member-by-member method was chosen as it presented equal performance when compared to the more sophisticated methods, but with the advantage of not requiring great mathematical efforts or data manipulation. The main result obtained indicated that the greater number of initializations of the short- and medium-term models (initialized every day and every 15 days, respectively) improves the performance of the precipitation forecasts, especially for the month following the start date of the forecast. The seamless precipitation forecast was applied to a hydrological modelling framework. Two post-processing techniques were tested to deal with the uncertainty of the hydrological forecasts, these being the assimilation of streamflow data in real time, and the application of an autoregressive correction (AR output-error correction) to adjust the model output (final streamflow predictions). The application of these techniques improved the performance of the forecasts, producing more reliable results and with lower average error, especially in the first two months of the forecast horizon. The streamflow forecasts obtained from the different stages of construction of the short to long-term hydrometeorological forecasting system were applied to predict the production of hydroelectric energy in the Brazilian electric system, based on the 30 plants associated with the 41 river basins of this thesis. The results showed a good performance of the forecasting system, which was able to predict when the production would be above or below the average production for the most distant forecast horizons (seasonal forecast). The work developed in this thesis proposes a tool that has great potential to be applied in the planning of the hydroelectric operation in Brazil, which can contribute to the optimization of the operation of the electrical system and the management of the use of the water stored in the reservoirs.