Prediction of daily direct solar energy based on XGBoost in Cameroon and key parameter impacts analysis

Abstract

This study explores the ability of Extreme Gradient Boosting (XGBoost) to predict the direct normal irradiation (DNI) under clear sky conditions in Cameroon. The satellite data used are DNI clear sky, air Temperature, Relative Humidity, Wind Speed, Wind direction, irradiation at Top of Atmosphere (TOA) and Aerosol Optical Depth at 550 nm (AOD550) for each aerosol type (Black Carbon : BCAOD550; Organic Matter : OMAOD550; Sea Salt : SSAOD550; Sulphate : SUAOD550 and Dust: DUAOD550). To achieve this aims and build a worst case prediction scenario, K-means clustering algorithm with Elbow and Silhouette analysis are used to select training and validation data sets. The coefficient of determination R2, root mean square error RMSE and the interpretation of the model outputs in the light of the state of the art confirm the robustness of the used model. The interpretation of the XGBoost outputs using the Shapley’s value shows that the amount of energy in the study area is most impacted by DUAOD550, OMAOD550, temperature, SUAOD550, TOA, SSAOD550 and relative humidity respectively. Results suggest also that, if dust and organic matter aerosols are present in the same proportion, the attenuation produced by them can be 4 to 10 times higher than those induced by black carbon and sea salt aerosols.