Aplicação de redes neurais nos resíduos de deslocamentos horizontais de blocos da barragem principal da Usina Hidrelétrica de Itaipu
Resumo
Resumo: Este trabalho propõe a aplicação da metodologia de Redes Neurais Recorrentes nos resíduos de quatro Séries Temporais de dados dos Deslocamentos Horizontais dos Blocos da Barragem Principal da Usina Hidrelétrica de Itaipu, Esses resíduos resultaram da modelagem da série por meio da metodologia Box & Jenkins, Dessa modelagem obteve-se uma parte linear e depois aplicada a Rede Neural Recorrente chamada Long Short-Term Memory (LSTM) aos resíduos da série (parte não linear), A partir de então, foram gerados valores não lineares que combinados com a parte linear, constituem os valores de previsões, Foi feito um comparativo da previsão da série com a metodologia Box & Jenkins com o modelo misto. Os resultados mostram que o modelo misto apresenta melhores resultados em três das quatro Séries Temporais estudadas, com redução do erro de até 40% Abstract: This work proposes the application of the Recurrent Neural Networks Methodology in the residuals of a Temporal Series of data of the Horizontal Displacements of the Blocks of the Main Dam of the Raipu Hydroelectric Power Plant, These residues resulted from the modeling of the series by the Box & Jenkins methodology, From this model, a linear part was obtained and then a Recurrent Neural Network called Long Short-Term Memory (LSTM) was applied to the residuals of the series (non-linear part). Since then, nonlinear values have been generated that, combined with the linear part, are the prediction values, A comparison of the prediction of the series with the Box & Jenkins methodology with the mixed model was made. The results show th at the mixed model presents better results in three of the four Time Series studied, with a reduction of the error up to 40%,