2026-05-122026-05-122026-01-19SANTOS, Wanderley Pereira dos. Modelagem da dinâmica hidrológica para previsão do nível de jusante da UHE Tucuruí por redes neurais artificiais. Orientador: Raphael Barros Teixeira; Coorientador: Cleison Daniel Silva. 2026. [12], 57 f. Dissertação (Mestrado em Computação Aplicada) – Núcleo de Desenvolvimento Amazônico em Engenharia, Universidade Federal do Pará, Tucuruí, 2026. Disponível em: https://repositorio.ufpa.br/handle/2011/18222. Acesso em:.https://repositorio.ufpa.br/handle/2011/18222This work investigates the application of LSTM, GRU, and TCN artificial neural networks in forecasting the downstream level of the Tucuruí Hydroelectric Plant, whose dynamic, nonlinear, and multivariable nature requires modeling formalisms based on deep learning. The need for the study is linked to the fact that the city of Tucuruí is located The need for the study is linked to the fact that the city of Tucuruí is located downstream from the plant and has inhabited areas near the banks of the Tocantins River, and this scenario requires control and monitoring of water levels in the city. For this to happen, there is a need for accurate downstream level forecasts in a timely manner so that intervention measures, when necessary, can be carried out by the competent authorities at the right time. To this end, in this work, several models were developed to be trained by artificial neural networks. The trained networks seek to model the complex relationships between inflow and outflow and the target variable, the downstream level. Hourly sampling series of real data from the period 2010 to 2025 are used, totaling 140,279 samples. The models are trained to predict up to 10 hours ahead, a window considered appropriate for the desired real scenario. Data preprocessing played a key role in the quality of the forecasts, with the moving average being applied, followed by normalization by maximum value, a strategy that contributed to the reduction of high-frequency noise and the stabilization of the training process. Subsequently, denormalization allowed the analysis of the results in real physical units, enabling (or providing) consistent hydrological interpretations. The quantitative evaluation was conducted using the MSE, FitNRMSE, and FitR² metrics, which enabled a complementary analysis of precision, absolute error, and explanatory power of the real signal variance. The results show high FitR² values (above 99%) for all networks, showing that the models were able to explain most of the variability in the downstream level. In addition, FitNRMSE values also remained high, demonstrating that forecast errors are small when compared to the natural occurrence of the system. After comparative analysis between the models, it was observed that GRU presented the best overall performance, less degradation of forecasts, and slight superiority in metrics, especially in the H3 model. LSTM performed similarly to GRU, but with greater sensitivity in regions of greater signal variability, while TCN demonstrated good average performance, but with a tendency to smooth predictions in more abrupt transitions. Thus, it is concluded that the H1, H2, and H3 models, together with the proposed LSTM, GRU, and TCN architectures, are capable of adequately representing the dynamics of the downstream level, with emphasis on H3 and GRU, which together showed superior performance to the other combinations.Figures 42 and 43 highlight the results obtained in the forecast analysis, called “real alerts,” whose simulation shows two events in sequence recorded at 11:00 a.m. and 12:00 pm on March 10, 2025, with sliding windows for the 10 future samples, including the actual values proven to validate the performance of the forecasts. These results constitute a solid basis for future applications in operational forecasting, decision support, and iintegration with predictive control strategies, such as MPC.Acesso AbertoAttribution-NonCommercial-NoDerivatives 4.0 Internationalhttp://creativecommons.org/licenses/by-nc-nd/4.0/Modelagem hidrológicaVazão afluenteVazão defluenteNível de jusanteIdentificação de sistemasMétricas estatisticasRedes neurais LSTM, GRU e TCNDefluent flowDownstream levelSystem identificationStatistical metricsLSTM, GRU, and TCN neural networksGRUTCNDissertaçãoCNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO::METODOLOGIA E TECNICAS DA COMPUTACAODESENVOLVIMENTO DE SISTEMASCOMPUTAÇÃO APLICADA