UTILIZAÇÃO DOS MODELOS DE REDE NEURAL ARTIFICIAL E REDE NEURAL RECORRENTE PARA REALIZAR PREVISÕES DO NÍVEL DO RIO CAÍ - SÃO SEBASTIÃO DO CAÍ, RIO GRANDE DO SUL

This academic work presents a study on the use of neural networks in predicting the water level of the Caí River, in the Brazilian state of Rio Grande do Sul. In this study, algorithms based on Real-Time Recurrent Learning (RTRL), Multilayer Perceptron (MLP) and GRU (Gated Recurrent Units) neural networks were evaluated to compare the effectiveness of such models in hydrological applications, especially in identifying signals that precede floods. The data used in this study covers a period of 46 years and was provided by the Agência Nacional de Águas (ANA), being used in the training of neural networks. In addition, six hydrological alert events recorded by the station located in the municipality of São Sebastião do Caí were selected. The results obtained made it possible to identify the model with the best performance and demonstrate its importance in reducing the impacts of floods, especially when used in conjunction with other existing monitoring systems. The accuracy of the models was tested in real scenarios, comparing the predictions of the algorithms with historical data and information collected in real time during flood periods, and this process allowed us to evaluate the ability of the models to predict hydrological time series efficiently. The results obtained in this study indicated that the GRU neural network showed greater effectiveness in predicting river levels, with 91% accuracy, but both algorithms were able to identify critical moments before floods, visually providing both an estimate of the time until the maximum river level is reached and the period necessary for the river to return to its usual level.