Analysis of Ensembles Applied to Time Series Forecasting of Hydroelectric Dams

Context: Dams are essential for electricity generation in Brazil. Given the countrys vast territorial extent, it is crucial to optimize the use of its energy sources, particularly the hydrological potential of rivers. Problem: Flow forecasting in river basins is critical for the safety and efficiency of operations. Underestimation errors can increase the risk of dam overflow and failure, while overestimations compromise water storage, affecting supply and power generation during dry periods. Solution: This study proposes a dynamic selection method of machine learning models to achieve more accurate and stable flow forecasts. IS Theory: The research is based on Dynamic Capabilities Theory, emphasizing the ability for internal reconfiguration to address complex issues such as time series forecasting. Method: Data were initially normalized and temporally adjusted to ensure consistency. A diverse set of base models was selected based on their intrinsic characteristics, then integrated into a dynamic ensemble framework. Performance was evaluated using MAPE and RMSE metrics, allowing comparative analysis between the proposed dynamic ensemble and traditional deep learning models. Results: The dynamic selection approach showed significant improvements, achieving errors of 23.88 (MAPE) and 852.97 (RMSE), outperforming traditional and deep learning models. Contributions and Impact in the IS Field: This work demonstrates that dynamic model selection is a superior and promising approach compared to isolated model use, offering a significant advancement for flow forecasting in dam systems and contributing to more efficient water resource management in the Brazilian context.