Previsão da precipitação pluvial por meio de redes neurais artificiais treinadas utilizando diferentes variáveis climáticas

• Autor(a) principal: Silva, Mateus Alexandre
• Data de Publicação: 2021
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Repositório Institucional da UFLA
• Texto Completo: http://repositorio.ufla.br/jspui/handle/1/46327
• topic: Modelagem hidrológica; Multilayer perceptron; Redes neurais artificiais; Previsão da precipitação; Hydrological modeling; Artificial neural networks; Precipitation forecast; Recursos Hídricos

Methods for predicting rainfall help to avoid and mitigate damage caused by its deficit or excess, in addition to providing the necessary tools for decision-making in socio-economic sectors and for the development of water resources use adequate planning. Thus, the aim with this research was to predict monthly precipitation, one month in advance, in four municipalities of the metropolitan mesoregion of Belo Horizonte, MG, Brazil, using models of artificial neural networks trained with different climatic variables. Also, the aim was to indicate the suitability of such variables as these model inputs. The artificial neural networks were developed using MATLAB® software version R2011a, with NNTOOL toolbox. The artificial neural networks training was done using the multilayer perceptron architecture and the feedforward backpropagation algorithm. Initially, the sequential number corresponding to the month and total rainfall, on monthly scale, from the years 1970 to 1999 were used as input to training, to forecast the rainfall occurring in the years 2000 to 2009. Subsequently, the data of the sequential number corresponding to the month, total rainfall, compensated average temperature, average wind speed and average relative humidity, together with data of the ENSO phenomenon occurrence, also between the years 1970 to 1999, were used to training, to predict rainfall on a month scale, from 2000 to 2009, comparing the results. Finally, the correlation between the used variables and rainfall was calculated and, subsequently, the 3 most correlated variables were used for the training of the models. It was found that the variables most correlated to rainfall of the following month were the sequential number corresponding to the month, the total rainfall and the compensated average temperature, and that the training using them obtained a better performance than the others. It was concluded that artificial neural networks are suitable for forecasting rainfall but have a limitation for predicting months with high values of it.