Modelagem do potencial de geração de energia hidrelétrica: uma contribuição para o planejamento energético na Amazônia

Abstract

The feasibility studies for hydroelectric power plants (HPP) use observed flow data to estimate the potential for energy generation. According to the Brazilian rule, the records need to cover a minimum period of seventy years of observation. However, in some regions the flow records comprise short periods, so that the time series does not capture the natural variability of the climate and this can lead to errors in the interpretation of the information, such that the power generation of the plant can be lower than expected. To address this problem, this thesis proposes an innovative methodology to estimate the potential of power generation. The principle is to use precipitation data, which has long time series, to estimate the potential of power generation through a modeling that employs artificial intelligence techniques. Two distinct methods of machine learning were tested. The first is a deep learning technique called Group Method Data Handling (GMDH). The second uses artificial neural networks with distinct options for optimization algorithms, Levenberg-Marquardt and Bayesian regularization. The methodology was applied to the Jatobá hydroelectric project in the Tapajós river basin in Pará. The evaluation indicators showed that the models have the skill to simulate the power generation, with a better performance for the GMDH, which reached 95% correlation with the actual data and only 12% average error percentage. The simulations obtained better performance during the dry season, which is fundamental, since this critical period for the generation of hydropower defines the firm energy of the enterprise. Statistical analysis on the observed data detected a significant tendency of precipitation reduction in some sub-basins of the Tapajós River. Simulations incorporating a climate change scenario proposed by the Intergovernmental Panel on Climate Change, as well as a scenario of long-term statistical trend, both indicated a reduction in power production capacity for the next twenty years, suggesting that the HPP Jatobá may not meet the project's firm energy demand for long-term under such conditions. The success of this approach can contribute to reducing uncertainties and subsidizing preliminary studies for the implementation of hydropower plants, as well as simulating scenarios to support planning, reduce costs and generate synthetic data for time series of power generation covering periods without observational field data.