Redes neurais convolucionais recorrentes aplicadas a previsão meteorológica de curto prazo através de imagens de radar

• Autor(a) principal: Rossatto, Felipe Copceski
• Data de Publicação: 2023
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Repositório Institucional da UFPel - Guaiaca
• Texto Completo: http://guaiaca.ufpel.edu.br/xmlui/handle/prefix/14182
• topic: Redes neurais; Meteorologia; Nowcasting; Radar; Neural networks; Meteorology; CIENCIAS EXATAS E DA TERRA; MATEMATICA; METEOROLOGIA

In this study, a model based on Recurrent Convolutional Neural Networks (RCNN) is proposed for short-term meteorological forecasting (nowcasting). This approach presents an alternative to traditional statistical extrapolation techniques. The chosen methodology involves the implementation of a supervised predictive learning RCNN model known as PredRNN++. The model is trained and evaluated using data in the form of images captured by four radars situated in the southern region of Brazil. These radar images are made freely accessible via the website of the National Institute for Space Research (INPE). The images serve as both input and output for the neural network. For the output, the model aims to predict images ranging from 6 to 120 minutes ahead in time, relative to the provided input images. The forecasted images are compared against the actual observations to assess the accuracy of the PredRNN++ model. This evaluation process involves empirical analysis of the predicted images, along with the application of statistical metrics such as Root Mean Square Error (RMSE), Structural Similarity Index (SSIM), and Mean Absolute Error (MAE). The assessment focuses on an extreme weather event that occurred in the southern region of Brazil on June 12, 2018. The outcomes of the study indicate that the PredRNN++ model demonstrates promise as a viable alternative for nowcasting predictions. This is attributed to its ability to effectively replicate the intensity and spatial distribution of the emulated meteorological systems.