@MastersThesis{Almeida:2021:EsCaAm,
author = "Almeida, Adriano Pereira",
title = "Uso de redes neurais para previs{\~a}o de descargas
el{\'e}tricas nuvem-solo a partir de dados de radar: um estudo de
caso na Amaz{\^o}nia",
school = "Instituto Nacional de Pesquisas Espaciais (INPE)",
year = "2021",
address = "S{\~a}o Jos{\'e} dos Campos",
month = "2021-05-28",
keywords = "tempestades severas, raios, RADAR, an{\'a}lise explorat{\'o}ria
de dados, intelig{\^e}ncia artificial, severe storms, lightnings,
exploratory data analysis, artificial intelligence.",
abstract = "A regi{\~a}o central da Bacia Amaz{\^o}nica {\'e} caracterizada
por ter uma vegeta{\c{c}}{\~a}o densa e condi{\c{c}}{\~o}es
atmosf{\'e}ricas favor{\'a}veis para o desenvolvimento de
tempestades severas e com alta incid{\^e}ncia de raios. Diante
disso, foi desenvolvido neste trabalho um estudo de algumas
parametriza{\c{c}}{\~o}es atrav{\'e}s de t{\'e}cnicas de
aprendizado de m{\'a}quina para prover subs{\'{\i}}dios para o
desenvolvimento de sistemas de previs{\~a}o de descargas
el{\'e}tricas a curto prazo, levando em considera{\c{c}}{\~a}o
as condi{\c{c}}{\~o}es locais dessa regi{\~a}o. Essas
parametriza{\c{c}}{\~o}es partiram de conclus{\~o}es obtidas na
an{\'a}lise explorat{\'o}ria dos dados de radar
meteorol{\'o}gico e da sele{\c{c}}{\~a}o de vari{\'a}veis por
meio de classifica{\c{c}}{\~o}es com a {\'a}rvore de
decis{\~a}o. Al{\'e}m disso, foram criados modelos baseados na
rede neural perceptron multicamadas (MLP) com diversas
configura{\c{c}}{\~o}es de entrada e varia{\c{c}}{\~o}es dos
seus principais hiper-par{\^a}metros, que tiveram como intuito
avaliar a sensibilidade das vari{\'a}veis em diferentes
intervalos de tempos anteriores {\`a} previs{\~a}o. As
previs{\~o}es foram feitas para 12 minutos {\`a} frente,
utilizando os valores das vari{\'a}veis em instantes anteriores
{\`a} previs{\~a}o: t-1 (12 minutos); t-2 (24 minutos); e t-3
(36 minutos). Embora a maioria das vari{\'a}veis
meteorol{\'o}gicas utilizadas possuam depend{\^e}ncias temporais
mais restritas, os modelos que utilizaram os tr{\^e}s instantes
de tempo, tiveram melhor desempenho. Junto com a
degrada{\c{c}}{\~a}o dos dados de entrada e
informa{\c{c}}{\~o}es sobre as caracter{\'{\i}}sticas dos
perfis de refletividade radar foi poss{\'{\i}}vel definir
crit{\'e}rios que podem ajudar na melhorias das previs{\~o}es de
raios via redes neurais. ABTSRACT: Lightning associated with
severe storms can cause several socio-economic problems. Due to
its high convective activity, predominantly equatorial location,
and large territorial extension, Brazil is one of the countries
where the highest incidence of lightning strikes occurs globally.
Studies show that the most significant cause of human fatalities
related to storms in Brazil is associated with lightning and that
a large part of them appear in rural areas. The material damages
caused by these events are also very recurrent. Therefore, the
interest in creating forecasting systems for this natural event is
of paramount importance. However, several challenges are
associated with the creation of lightning forecasting systems.
Sometimes these challenges are related to the characteristics of
the event itself, such as its high variability and its
relationship with complex physical processes that are difficult to
model computationally. The local factors that can influence the
development of electrically active storms should also be
considered. The central region of the Amazon Basin is
characterized by dense vegetation and favorable atmospheric
conditions for the development of severe storms and a high
incidence of lightning. Therefore, a study of some
parameterizations through machine learning techniques was
developed in this work to provide subsidies for the development of
lightning forecasting lightning in the short-term systems,
considering the local conditions of the central region of the
Amazon Basin. These parameterizations started from conclusions
obtained in the exploratory analysis of meteorological radar data
and the selection of variables through classifications with the
decision tree. In addition, models based on the neural network
multilayer layer perceptron (MLP) with different input
configurations were created to evaluate the sensitivity of the
variables at different time intervals before the forecast. The
forecasts were made for 12 minutes ahead, using the previous times
values of the variables: t-1 (12 minutes); t-2 (24 minutes); and
t-3 (36 minutes). Although most meteorological variables used have
more restricted temporal dependencies, the models that used three
previous time instants performed better. Along with the
degradation of the input data and information on the
characteristics of the radar reflectivity profiles, it was
possible to define criteria that can help to improve the lightning
forecasts via neural networks.",
committee = "Santos, Rafael Duarte Coelho (presidente) and Calheiros, Alan
James Peixoto (orientador) and Shiguemori, Elcio Hideiti and
Rodriguez, Carlos Augusto Morales",
englishtitle = "Use of neural networks to predict cloud-to-ground lightning from
radar data: a case study in the Amazon",
language = "pt",
pages = "208",
ibi = "8JMKD3MGP3W34R/44JMNDL",
url = "http://urlib.net/ibi/8JMKD3MGP3W34R/44JMNDL",
targetfile = "publicacao.pdf",
urlaccessdate = "26 abr. 2024"
}