Advancements in streamflow modeling and forecasting in Brazil

Abstract

This doctoral thesis introduces two significant advancements in streamflow modeling and forecasting in Brazil. The first contribution lies in the methodological domain, wherein a novel statistical technique for multisite streamflow simulation is proposed, the GLM-Copula which couples Generalized Linear Models (GLM) to represent the temporal dependencies and the Copula to represent the spatial dependencies. This technique enables the modeling of dependence structures using various probability functions, thus eliminating the need for normalization. The results showed that the GLM-Copula approach ability to preserve summary statistics from the historical data was similar to the classical multivariate ARMA and the state- of-art COPAR models. For the dependency structures, the GLM-Copula reproduced what was narrowly the best in reproducing the short-term temporal dependence (lag-1 autocorrelation), narrowly the worst in reproducing the spatial dependence (lag-0 cross-correlation) and reasonable the best in reproducing the total association (copula entropy).The proposed model allows the modelling of both spatial and temporal dependencies without normalization, is computationally efficient and can be used as a dimensional reduction, which we suggest justifies its addition to the time series generation toolbox.The second contribution of this thesis lies in the dynamic streamflow forecasting literature domain, where a multimodel probabilistic approach for seasonal streamflow forecasting is applied for all Brazilian hydropower catchments. This contribution seeks to fill the scientific literature gap in dynamic seasonal streamflow forecasting in Brazil by proposing a mutilmodel ensemble approach to generate monthly probabilistic streamflow forecasts from a hydrological model forced by both NMME and SUBX predictions for all the 87 hydropower catchments that are monitored by the electrical system’s operator in Brazil. The findings of this study demonstrate that the combination of multiple forecast models into a Multimodel probability forecasting approach yields improved performance and greater robustness compared to individual models, including the best individual model for each catchment. These results hold true for streamflow forecasting across all Brazilian regions.