Advanced Ensemble Machine Learning for Photovoltaic Production Forecasting in Tropical Microgrids: Application on the Katsepy Site, Madagascar

Abstract

This article presents a Machine Learning ensemble model for photovoltaic (PV) production forecasting, developed to address energy management challenges in rural tropical microgrid environments. The site studied is located in Katsepy, Madagascar, where the reliability of energy forecasts remains a significant challenge due to high climate variability. The main objective is to design a more accurate and adaptable forecasting tool to enable adequate energy planning and system operation. The methodology is based on historical photovoltaic production data and meteorological data collected between 2005 and 2023 from the PVGIS online platform. Several regression algorithms were evaluated, including Random Forests, Bagging, and Gradient Boosting, to identify the models best suited to the local context. Among these, Gradient Boosting showed the best performance according to RMSE, MAE, MAPE and R² measurements, followed closely by Random Forest and Bagging. The experimental process consists of two stages: first, validation using actual 2023 data, and then forward-looking forecasts for 2024 incorporating real temperature data. To improve accuracy and robustness, a Stacking ensemble model was constructed, combining the three best performing algorithms as base estimators and the Extra Trees Regressor as the meta-model. This ensemble approach consistently outperformed the individual models and provided realistic production estimates for 2024, indicating a moderate decline in photovoltaic production compared to 2023, driven by observed climate variations. The proposed forecasting framework provides a solid foundation for future work on optimal energy management and fault diagnosis in the Katsepy microgrid system, with great potential for adaptation to other tropical coastal regions.