Assessment of ANN and SVM models for estimating normal direct irradiation (Hb)

This study evaluates the estimation of hourly and daily normal direct irradiation (Hb) using machine learning techniques (ML): Artificial Neural Network (ANN) and Support Vector Machine (SVM). Time series of different meteorological variables measured over thirteen years in Botucatu were used for training and validating ANN and SVM. Seven different sets of input variables were tested and evaluated, which were chosen based on statistical models reported in the literature. Relative Mean Bias Error (rMBE), Relative Root Mean Square Error (rRMSE), determination coefficient (R2) and “d” Willmott index were used to evaluate ANN and SVM models. When compared to statistical models which use the same set of input variables (R2 between 0.22 and 0.78), ANN and SVM show higher values of R2 (hourly models between 0.52 and 0.88; daily models between 0.42 and 0.91). Considering the input variables, atmospheric transmissivity of global radiation (kt), integrated solar constant (Hsc) and insolation ratio (n/N, n is sunshine duration and N is photoperiod) were the most relevant in ANN and SVM models. The rMBE and rRMSE values in the two time partitions of SVM models are lower than those obtained with ANN. Hourly ANN and SVM models have higher rRMSE values than daily models. Optimal performance with hourly models was obtained with ANN4h (rMBE = 12.24%, rRMSE = 23.99% and “d” = 0.96) and SVM4h (rMBE = 1.75%, rRMSE = 20.10% and “d” = 0.96). Optimal performance with daily models was obtained with ANN2d (rMBE = −3.09%, rRMSE = 18.95% and “d” = 0.97) and SVM2d (rMBE = 0.60%, rRMSE = 19.39% and “d” = 0.97). ANN and SVM models improved Hb estimations as compared with other results from the literature. SVM has better performance than ANN to estimate Hb, and it should be the first option of choice.