Sistema de previsão de eventos meteorológicos convectivos para a área terminal do Rio de Janeiro

Abstract

In this work, two weather forecasting approaches using machine learning algorithms for the flight terminal area of the International Airport of Rio de Janeiro are explored. The first method uses thermodynamic stability indices extracted from the airport soundings as input variable for the model and, as target variable, the prediction of (a) occurrence and (b) severity of meteorological convective events - classified from atmospheric discharges recorded by the RINDAT network between 2001 and 2016. The results found for the prediction of occurrence (severity) indicate more than 90% (84%) probability of detection and about 10% (20%) false alarm. A second method consists of training artificial neural networks for local data assimilation (surface and profile) in the regional atmospheric model Weather Research Forecasting applied to 6-hour forecast fields. This method proved to be efficient to emulate the traditional data assimilation approach (3d-var) without significant loss of information and with two great advantages: execution time around 70 times faster and can be used without high-performance hardwares.