Previsão de vazão de curto prazo em bacias hidrográficas do Sistema Interligado Nacional utilizando novos modelos fuzzy evolutivos gaussianos multivariáveis aprimorados

Abstract

Efficient water resource management is a fundamental pillar of the hydropower generation sector worldwide, with significant social, environmental, and economic implications. In this context, accurate streamflow forecasting methods constitute essential tools to support reliable and sustainable energy planning. This study investigates the applicability and effectiveness of evolving fuzzy systems aimed at improving short-term streamflow forecasting in the Brazilian hydrological system. As an original contribution, three novel approaches are proposed by incorporating adaptive hyperparameter update mechanisms into the evolving Multivariable Gaussian (eMG) model structure, introducing a methodological modification based on the dynamic adaptation of the significance level (λ) according to the prediction error. This is the first known application of such adaptive mechanisms in evolving fuzzy systems for streamflow forecasting. The performance of the proposed models was evaluated against the benchmark model currently used by the National System Operator (ONS), the SMAP/ONS, as well as other classical approaches in the literature for this application, such as deep learning (LSTM) networks. The analysis used streamflow and satellite-based precipitation data, covering an eleven-year period (2011–2021). Six strategic river basins of the National Interconnected System (SIN) were considered: Furnas, Itaipu, Itumbiara, São Simão, Três Marias, and Tucuruí, encompassing small and large-scale basins with a wide diversity of hydrological regimes. A robust evaluation protocol was developed, and results demonstrate that the proposed models consistently outperformed the benchmark models, showing a significant reduction in forecast errors and near-zero bias. To corroborate these findings, a statistical validation protocol confirmed the significance of these improvements in most tested scenarios. Due to their adaptive nature and computational efficiency, the proposed models effectively handle the variability inherent in streamflow observations, offering robust and promising solutions for real-time applications in national hydropower management, and consequently contributing to a safe and reliable electricity generation system capable of meeting the country’s demand.