Machine Learning in Estimating Daily Global Radiation in the Brazilian Amazon for Agricultural and Environmental Applications

Knowledge of global radiation (Hg) is essential for regional economic development and can help guide public policies related to agricultural and energy potential. However, its availability in several Brazilian regions is still limited. This work evaluates the predictive capacity of two machine learning (ML) techniques, such as multi-layer perceptrons (MLPs) and support vector machines (SVMs), in the estimation of Hg in 20 meteorological stations with 40 different input combinations involving insolation, air temperature, air relative humidity, photoperiod, and extraterrestrial radiation. It is also compared with three empirical models based on insolation, temperature, and a hybrid combination. In general, the greater the number of input variables, the better the performance of ML techniques, especially in combinations involving insolation that reduced the dispersion of estimated Hg on days with high atmospheric transmissivity and air temperature on days with low atmospheric transmissivity. The performance of SVM was better when compared to MLP in all statistical indicators. ML techniques presented better results than empirical models, and in general, the ordering of the best models in the three locations is achieved using SVM, MLP, and empirical models. Therefore, due to their easy implementation and generation of good results, the use of SVM models is recommended to estimate daily global radiation in the Brazilian Amazon.