Análise e previsão da energia armazenada em reservatórios hidrelétricos utilizando técnicas de aprendizado de máquina

This work proposes an approach for forecasting energy stored in hydroelectric reservoirs, through the application of machine learning techniques, with emphasis on the use of the LSTM recurrent neural network. The data source used was the Daily Affluent Natural Energy (ENA) database by Subsystem, which consists of values of energy produced by plants in the Southeast/Central-West, made available by the National Electric System Operator (2024), performing a pre-processing process to adapt the data and then applying machine learning techniques. This approach allowed the exploration of historical patterns of energy generation and storage, aiming at the development of accurate forecasting models. Subsequently, an analysis of the results obtained was conducted, highlighting the techniques that demonstrated the best performance, through the evaluation of their pertinent metrics. The experiment that stood out the most was Experiment 3, which used moving averages to smooth the signal, obtaining the best results for the MASE (1.1009) and POCID (93.5094%) metrics with a 15-day smoothing window.