Aplicação de redes neurais artificiais e do algoritmo dos k-vizinhos mais próximos para modelar o processo chuva-vazão na Bacia do Rio Piauitinga (SE)

• Autor(a) principal: Dantas, Luiz Antônio Muñoz
• Data de Publicação: 2022
• Tipo de documento: Trabalho de conclusão de curso
• Idioma: por
• Título da fonte: Repositório Institucional da UFS
• Texto Completo: http://ri.ufs.br/jspui/handle/riufs/17147
• topic: Engenharia Civil; Ensino de engenharia civil; Modelagem chuva-Vazão; Redes Neurais Artificiais (RNA); Algoritmo dos k-vizinhos; Weka; Rio Piauitinga (Bacia); Rainfall-runoff Modeling; Artificial Neural Networks; K-Nearest Neighbors; Weka; ENGENHARIAS::ENGENHARIA CIVIL::ENGENHARIA HIDRAULICA::HIDROLOGIA

This work presents the application of artificial neural networks and the k-nearest neighbors algorithm to model the rainfall-runoff process in the Piauitinga river basin, Sergipe. The input data used in the modeling were those from the Estância streamgauge station and from the Salgado raingauge station, which were grouped into calibration and validation sets, selected randomly. The simulation runs took place by means of the machine learning software Weka, and four different formulations were tested for the daily and monthly situations. The efficiency of the simulations performed was verified by the Nash-Sutcliffe coefficient, according to which considerable results were obtained for the situations presented, mainly for the models based on artificial neural networks. The formulation containing as features rainfall and streamflow of up to three previous periods was the one with best results, providing efficiency classified as good for the daily approach and very good for the monthly approach. Alternative modeling was carried out to compare the model using only one raingauge station with another using the average rainfall in the basin, and it was noted that the model that uses only the Salgado raingauge station achieved better or equal results than the others. Finally, new simulations were carried out using the initial periods for calibration and the final ones for validation, and it was noticed that the simulations that randomly classified the data gave much better results.