Título: Training a Neural Network to Detect Patterns Associated with Severe Weather
Autores: Lima, Glauston R. T. de; Stephany, Stephan
Resumo: Early detection of severe convective events is essential to take preventive measures in order to reduce the related negative impacts. However, the increasing amount of data that meteorologists need to analyze nowadays precludes the chances of forecasting such events. Hence, the development of advanced data mining tools to support weather forecasting became a current research topic in Meteorology. This work presents the results obtained with an artificial neural network that was designed to recognize atmospheric patterns associated to severe convective activity. The network was trained with patterns given by a set of selected meteorological variables values of the numerical weather forecasting model Eta. The classes of atmospheric convective activity are defined by values of density of occurrence of atmospheric electrical discharges. It is assumed that such density can be correlated to the level of convective activity. The network architecture and training algorithm are based on heuristics derived from previous studies of the authors. Once validated, the proposed neural classifier would be able to screen the Eta model outputs and assist meteorologists in the early detection of severe convective events. In addition, it would also assess the Eta model skill to predict such events. In a first approach, only two data classes are considered, corresponding to severe convective activity and to moderate/weak/absent convective activity. The convenience of defining more classes will be evaluated in further work. The classification results are promising.
Palavras-chave: Supervised neural network; data mining; classification; convective activity; weather forecasting
Páginas: 30
Código DOI: 10.21528/lmln-vol11-no2-art5
Artigo em PDF: vol11-no2-art5.pdf
Arquivo BibTex: vol11-no2-art5.bib