Eventos Anais de eventos
COBEM 2021
26th International Congress of Mechanical Engineering
EVALUATION OF MACHINE LEARNING MODELS FOR VERY SHORT-TERM SOLAR IRRADIANCE FORECASTING: A CASE STUDY FOR PETROLINA / PE, NORTHEASTERN BRAZIL
Submission Author:
Nadja Gomes de Oliveira , CE
Co-Authors:
Nadja Gomes de Oliveira, Paulo Alexandre Costa Rocha, Mostafa Elsharqawy
Presenter: Nadja Gomes de Oliveira
doi://10.26678/ABCM.COBEM2021.COB2021-1902
Abstract
This work uses the SONDA network irradiance data to forecast global horizontal and direct normal irradiances (GHI and DNI) in very short-term intra-hour for 5 minutes resolution during the period of four years for one solarimetric station in the northeast of Brazil, Petrolina/PE. Four different machine learning models were tested, namely 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 with the RMSE and Forecast Skill relative to the persistence model. Results show that the machine learning models achieve significant forecast improvements over the reference model. In addition, the Ensemble with Ridge Regression and XGBoost models have rarely been used for short-term solar forecasting. This framework can be used to select appropriate machine learning approaches for short-term solar power forecasting and the simulation results can be used as a baseline for comparison. The four methods presented similar behavior for raw and normalized variables, with the RMSE values ranging between 72.85 W/m² and 76.12 W/m² for GHI and values between 104.22 W/m² and 104.66 W/m² for DNI. It is worth to mention that the Ensemble with Ridge model outperformed all the other methods for GHI, obtaining RMSE values between 72.85 W/m² and 74.02 W/m² and for DNI, the XGBoost model outperformed obtaining RMSE values between 104.22 W/m² and 104.66 W/m². Regarding the Forecast Skill (FS), the Ensemble with Ridge model performs 0.60% better than the XGBoost for GHI, although the XGBoost performs 1.58% better than the Ensemble with Ridge model for DNI. Simulation results for FS also showed that the use of the clear-sky index for GHI increased most of the model’s performance between 0.70% to 1.15% for GHI, excepted for the LASSO, and between 0.27% and 0.36% for DNI with the LASSO and XGBoost, although a decreased between 0.67% and 0.83% was observed for the kNN and the Ensemble with Ridge models.
Keywords
machine learning, global solar irradiance, Direct Normal Irradiance, intra-hour forecasting
