Aplicação de Deep Learning no preenchimento de falhas em dados micrometeorológicos

Abstract

This work aimed to to study the use of Deep Learning in gap filling in microclimatic data series. The databases used belong to automatic stations of the Institute of Meteorology (INMET), located in the cities of Campina Verde - MG, Sorriso -MT, Diamante do Norte -PR and Campo Bom -RS, collected from November 2014 to January 2015, composed by the following variables: solar radiation, wind speed, dew point, relative humidity, atmospheric pressure and air temperature. The experiment consisted in the construction of a model of machine learning and analysis of four basic factors, such as the number of hidden layers, number of neurons, number of independent variables used in the estimation of the target variable, and finally the impacts caused by the percentage of failures contained in the series. For the statistical analysis, Kruskal-Wallis and Mann-Whitney tests were used in the analysis of variance, Nemenyi test to paired comparison, and comparisons with other models widely used in the literature. The results showed that the number of 50 and 100 neurons does have a significant impact on the network response, that the use of only one independent variable has a significant difference in comparison to the use of more, and the percentage of failures influenced only the filling of the point time series of dew. It is concluded that the developed model presents good results for filling temperature, relative humidity and dew point faults.