Comparative study of Decision Tree, Random Forest and XGBoost performance in forecasting the power output of a photovoltaic system

In this paper, we make a comparative study of the performance of three methods for predicting the power output of a photovoltaic installation: Decision Tree, Random Forest and XGBoost. We performed these predictions in Python using as input meteorological data such as wind speed, sun position, temperature, direct irradiation, diffuse irradiation and reflected irradiation and as output data the power output of a 1000Wp panel. These data were downloaded from the PVGIS database for the city of Natitingou (Benin) and for a period of 12 years (from January 1st 2005 to December 31st 2016). We obtained as Mean Square Errors 2.195026, 3.058383 and 5.544319 respectively for the XGBoost, Random Forest and Decision Tree and for Regression Values 0.9999999194, 0.9999797366 and 0.9997013968 respectively for the XGBoost, Random Forest and Decision Tree. We conclude that all three models are effective for the forecasting task performed but that the XGBoost is the best performing model with Mean Square Error and Regression Value of 2.195026 and 0.9999999194 respectively.