Enhancing solar radiation prediction accuracy: A hybrid machine learning approach integrating response surface method and support vector regression

Abstract

An accurate solar radiation (SR) prediction with a practical training approach is vital in estimating solar energy. A hybrid machine learning (ML) model is proposed for estimating the monthly SR. The proposed model includes two ML approaches: the response surface method (RSM) and support vector regression (SVR). The RSM is used to optimize the input variables and handle the data points for the prediction of SR. The first ML approach presents two input variables to estimate data handling. In the second ML process, the SVR model provides a nonlinear regression for handling data supplied by RSM. A new model was employed to predict the SR data taken from two stations in Turkey, as the temperature and extraterrestrial radiation were used as the model inputs. The RSM, artificial neural networks (ANNs), SVR, multivariate adaptive regression spline (MARS), M5 model tree (M5Tree) and convolutional neural networks (CNN) methods as existing ML approaches were employed to compare the predictions proposed hybrid ML approaches using several criteria. Data were split into training and testing sets, and two scenarios were established to compare models’ efficiencies according to different sets. The outcomes showed that the proposed model provides better accuracy for estimating SR using limited input data than other alternatives. The accuracy of the ANNs, SVR, MARS, M5Tree, RSM and CNN models was improved using a hybrid ML model. The proposed RSM-SVR method enhanced the efficiency of the ANN, SVR, MARS, M5Tree, and RSM methods by RMSE margins ranging from 0.1% to 5.6%, 2.8% to 7.3%, 1.0% to 8.3%, 0.1% to 28%, and 2.0% to 5.9%, respectively.