Eventos Anais de eventos
COBEM 2021
26th International Congress of Mechanical Engineering
Solar irradiance forecasting using machine learning, comparison of XgBoost and Boosting models.
Submission Author:
David Mickely Jaramillo Loayza , MG
Co-Authors:
David Mickely Jaramillo Loayza, Paulo Alexandre Costa Rocha, Maria Eugenia V. da Silva
Presenter: David Mickely Jaramillo Loayza
doi://10.26678/ABCM.COBEM2021.COB2021-1090
Abstract
One of the most important tools for electricity generation based on solar energy is the solar irradiance forecasting. This is because the solar irradiation depends on mostly by weather conditions variability, and an accurate forecasting is needed to control and distribute efficiently the electricity demand of generation systems. Machine learning models are being used to develop many types of forecasting models, namely the Extreme Gradient Boosting (XGBoost) model, which has high capability and works faster than the others tree boosting models. Many forecasting models that are robust have a significant demand of resources to work, hence representing a high cost of implementation. On the other hand, the XGBoost is a model that has high efficiency and low resources demand. That capability allows to test models with different time scales, from minutes to days. In this sense, the objective of this research is to analyze the performance of XGBoost model on solar irradiance forecasting and compare with the Tree Boosting model applied on the same forecasting situations. The XGBoost and Tree Boosting models were analyzed using different time scales (2 min, 10 min, 30 min, 1h, 1 day) from the same dataset that was partly used to train the models. This performance comparison was done using error metrics like RMSE, nRMSE, MAE, nMAE and others. The results were obtained using 50% of the original dataset, above of this percentage the values of errors do not have a significantly variation and allows to decrease the time of computing. Boosting model presents better performance than XGBoost in this kind of work, but with a slight similarity; XGBoost was faster to compute than the Boosting model. The performance of the models was affected by the ONI predictors in a positive way, where La Niña presents better FS values.
Keywords
solar irradiance forecasting, Solar irradiance, Solar Energy, machine learning, Extreme Gradient Boosting, XGBoost, Renewable energies
