Modeling with Artificial Neural Networks to estimate daily precipitation in the Legal Amazon

Abstract

Abstract Hydrological analyzes carried out from precipitation in the Legal Amazon (AMZ-L) are essential due to their importance in climate regulation, regional and global atmospheric circulation. However, in this region, there are limitations related to data series with short periods and many gaps, especially in the daily scale. Thus, to improve precipitation analyses, a non-parametric stochastic model based on Artificial Neural Networks (ANNs) was used to estimate daily precipitation in AMZ-L. For this, 22 rainfall stations were adopted over a period of 18 years (1998-2016) and with <1% missing data, which were organized considering the complete series and the seasonal periods (rainy and dry). The results obtained demonstrate the good capacity of the model to preserve the precipitation characteristics of the evaluated rainfall stations, mainly those with a more humid climate and with more frequent precipitations during the year, as is the case of those located in the Amazon Biome. However, in regions that suffer prolonged periods of drought, such as the Amazon-Cerrado Ecotone areas, the results were less satisfactory due to the greater recurrence of zeros in the historical series. The seasonal division into dry and rainy periods did not provide better estimates to the model, except for some rainfall gauge stations located at latitudes close to the equator. However, this study may support future research on the estimation of daily precipitation in the region.