Modelos empíricos e machine learning na estimativa da radiação global diária na Amazônia

Abstract

Global radiation (Hg) directly influences various chemical, physical and biological processes, including evapotranspiration, photosynthesis, photovoltaic energy generation and others. However, its measurement is restricted to certain regions and historical series are necessary for scientific studies, agricultural, environmental and industrial projects. Some meteorological variables such as air temperature or insolation correlate well with Hg. This study evaluated 87 empirical models for estimating global radiation using insolation, air temperature, relative humidity, astronomical variables and hybrid combinations with more than one variable, and also evaluated MLP and SVM machine learning algorithms with 40 different input combinations, for 20 cities distributed in the Brazilian Amazon Biome. With the empirical models, using insolation, the best performance was obtained with the potential model, with S, So and Ho as input variables, while for air temperature, the best performance was obtained with the model with ΔT, Tmed and Ho as input variables, The hybrid model with ΔT, S, So and Ho as input variables had the best performance, while the models with astronomical variables were non-significant (NS) for all the weather stations evaluated and the models based on relative humidity showed poor performance. In the evaluation using the MLP and SVM algorithms, the greater the number of input variables, the better the performance, especially with the inclusion of insolation and air temperature, which reduces scattering on days with high and low atmospheric transmissivity, but when all the variables are available it is recommended to use the combination RHmax, RHmed, RHmin, Tmax, Tmed, Tmin, S, So, Ho. Overall, there was no significant difference in the performance of the empirical models and the MLP and SVM algorithms with local meteorological data for estimating global radiation in the Brazilian Amazon biome.