Solar radiation times-series forecasting in southern Brazil: A comprehensive analysis

To enable the study of solar behavior without installing expensive sensor equipment, machine learning time-series models can be highly useful. In this study, we forecast future values of solar radiation incident on a horizontal surface by comparing five different models: Holt-Winters, LSTM, SARIMAX, SVM, and XGBoost, using a comprehensive dataset of satellite meteorological observations from NASA spanning over 30 years. 90 points in the Brazilian southeast (in the state of Minas Gerais and its surroundings) were analyzed using two different cross-validation methods (Fixed Start and Rolling Window) and compared using the Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Mean Absolute Percentage Error (MAPE) metrics. Our analysis revealed that the Holt-Winters model yielded the lowest error, with an MAE of 0.302 kWh/m 2 /day, followed by the LSTM (0.314), SARIMAX (0.338), SVM (0.39), and XGBoost (0.336) models. The statistical analysis of the cross-validation methods revealed that although the fixed start method yields lower error metrics, it requires substantially longer training times (due to the increased input data) and is only slightly superior to the rolling window method. The most significant divergence between the models and the actual solar radiation values was observed along the eastern border of the state. An exploratory analysis of solar behavior showed that greater data variability (standard deviation and variance) is associated with worse forecasting performance. Given the worldwide availability of the data, the methodology presented in our work can be replicated to make solar radiation predictions anywhere, facilitating new developments in sustainable renewable energy production.