Aplicação de redes neurais para a interpolação da precipitação no Estado de Alagoas

Abstract

Tropical regions are characterized by presenting a great variability in the distribution of its rainfall regimes, however lack of information about the spatial and temporal distribution of rainfall is a serious obstacle to understanding and modeling its variability. Thus, the purpose of this study was to use Artificial Neutral Networks to study the variability of precipitation in order to to infilling missing, using the temporal and spatial interpolation of precipitation for the state of Alagoas. For the spatial interpolation were used monthly averages of 63 stations located in the state of Alagoas and 182 stations in the state of Pernambuco between the years 1965 to 1980, for the months of May and November. For the temporal interpolation was used daily data of 5 pluviometric and 1 fluviometric stations between the years 1993 to 2002. All stations used are from the database of the Agência Nacional of Água (ANA). For the interpolation were adopted some entry scenarios with different settings for the training process of neural networks, taking into account latitude, longitude, altitude, rainfall stations within the region of spatial dependence and rainfall data from previous days. The region of spatial dependence and spatial distribution of precipitation for the months in the study were obtained through the use of geostatistics. Neutral networks have proven to be a promising tool for the simulation of precipitation data, showing satisfactory results for the spatial interpolation with mean absolute error between 0,04 to 51,40mm and Mean Absolute Percentage Error from 0,03 to 30,48% for the month of May, and mean absolute error between 0,00 to 5,79mm and Mean Absolute Percentage Error between 0,00 to 30,65% for the month of November. Temporal interpolation obtained the best result for the scenario-1 with the use of five past data, with mean absolute error of 0,067mm and Mean Absolute Percentage Error of 5,176%.