A lightweight time series method for prediction of solar radiation

Solar radiation (Rs) is vital and profoundly influences the environment. Accurate forecasting of Rs is crucial in renewable energy applications, despite its nonlinearity and dependency on loads. To overcome limitations in measurement tools, various methodologies are employed to estimate Rs using alternative environmental parameters. In our article, we present an innovative framework that explores the impact of feature selection (FS) on time series for accurate global Rs forecasting. This framework provides a holistic approach to recursive feature elimination (RFE) and its integration with various models such as random forest (RF), Decision Tree (DT), Logistic Regression (LR), Classification and Regression Tree (CART), Person (Per) and Gradient Boosting Models (GBM). The obtained results reveal that the CART, LR, and GBM models exhibit strong predictive accuracies of 0.894, 0.884, and 0.882, respectively. Notably, these three methods demonstrate a consistent standard deviation (std) of 0.033, indicating stability in their performance. Evaluating the normalized mean absolute error (nMAE) standard deviation (std), the models achieve values of 0.892 (0.029), 0.885 (0.034), and 0.885 (0.035) respectively. Additionally, the RFE algorithm showcases the significant impact of input lags as features and delivers good performance. Beyond accuracy, our findings hold practical implications for renewable energy planning, daily operation of solar power plants, and investment decision-making, contributing to the optimization and sustainability of solar energy systems.