Three Days-Ahead Solar Irradiance Forecasting in Burkina Faso: A Comparison of WRF-Solar Model and Machine Learning Approaches.

Abstract

The increasing production on solar photovoltaic (PV) energy in Burkina Faso necessitates accurate forecasting of solar irradiance to optimize grid production and power planning. This study evaluates and compares the performance of the physics-based WRF-Solar model and three machine learning (ML) techniques such as Random Forest (RF), Support Vector Regression (SVR), and Long Short-Term Memory (LSTM) for three-day-ahead forecasting of Global Horizontal Irradiance (GHI). Using ground-based measurements from the Zagtouli solar farm and reanalysis data from ERA5, CAMS, and ECMWF-HRES for the year 2020, the research analyses model performance under diverse atmospheric conditions, including clear-sky, cloudy, high aerosol, and mixed (cloudy + aerosol) scenarios. The methodology includes quality control of observations, sky condition classification using the clearness index, and model evaluation using statistical metrics such as RMSE, MAE, R², and IOA. Results demonstrate that WRF-Solar outperforms 3 ML models used in this thesis under clear sky condition. Under cloudy condition, ML models particularly LSTM has given better prediction compared to WRF-Solar models. No clear winner between models under aerosol condition, while in mixed condition, the WRF-solar and LSTM have shown a continue competency for the three days ahead solar irradiance forecasting. The study concludes that ML methods offer a promising alternative or complement to traditional physical models in solar forecasting in West Africa’s dynamic climate.