Evaluation of machine learning models for solar irradiance prediction (GHI and DNI): a case study in Petrolina, PE, Brazil

Abstract

This work uses the SONDA network irradiance data to forecast global horizontal and direct normal irradiances (GHI and DNI) intra-hourly applying 5min and 60min forecast window resolution and five different time horizons (5min, 30min, 60min, 6 hours and 12 hours) during the period of four years for a solarimetric and anemometric station in the northeast of Brazil, Petrolina/PE. Five different machine learning models were tested, namely: Multivariate Adaptive Regression Splines (MARS), Least Absolute Shrinkage and Selection Operator (LASSO), k-nearest neighbors (kNN), Extreme Gradient Boosting (XGBoost) and an ensemble combination to form a final forecast (Ensemble with Ridge Regression). Their performance was compared using the RMSE and forecast skill (FS) relative to the smart persistence model. Results show that the machine learning models achieve significant forecast improvements over the reference model using only endogenous features. In addition, the Ensemble with Ridge Regression and XGBoost models have rarely been used for very short-term solar forecasting according to the literature. This framework can be used to select appropriate machine learning approaches for very short-term solar power forecasting and the simulation results can be used as a baseline for comparison. The XGBoost’s forecast skill model was not the winner in all time horizons and resolutions, but it is among the best results for GHI and DNI, with normalized variables. The XGBoost model prevails when the time resolution of 5 min is chosen, not considering other error metrics, such as MBE. It is worth to mention, for the time resolution of 5 min, that the XGBoost model has the best FS results in 66.66% of the time comparing to all the six results for GHI and DNI with raw and normalized variables. For the time resolution of 60 min, the MARS model has the best forecast skill’s results, dominating around 66.66% of all the outputs, including GHI and DNI for raw and normalized variables. Also, kNN is the Machine Learning model with the best outputs of MBE, proving that the model is more accurate and does not have huge estimations variations comparing to the other models.