Novel Python-based “ all-regressor model ” application for photovoltaic plant-specific yield estimation and systematic analysis

Abstract

ABSTRACT The deployment of photovoltaic technology in global electricity generation has reduced the dependence on fossil fuels and the adverse effect of global warming climate changes. However, the variability of PV output due to its reliance on meteorological parameters hinders the total maximization of the technology. In this paper, the interdependence of meteorological parameters on accurate estimation of the specific yield of the PV system is examined using machine learning algorithms. The analysis uses 2 years of meteorological data in a 10-min interval. The meteorological parameters used as input are global solar radiation (GR), wind speed (WS), wind direction (WD), the standard deviation of wind direction (WSD), air temperature (AT), and relative humidity (RH), while the output variable is the PV specific yield (PVSY). This study uses a Python-based “all regressor” machine learning algorithm, to analyze their accuracy for the estimation analysis at different input feature space. The accuracy of the models and algorithms is analyzed using mean absolute error (MAE), mean square error (MSE), root mean square error (RMSE), and coefficient of correlation (R2). The best performance in the two feature space was retrieved for AT, and GR with MSE and MAE values of 0.113 and 0.252, respectively. The best performance in the three feature space was retrieved for WSD, AT, and GR with MSE and MAE values of 0.109 and 0.255, respectively. The results show that the most accurate PVSY estimation was for the four input feature space with input variables of WS, AT, GR, and RH. The K-nearest neighbor (KNN) proved to be the most appropriate algorithm for this feature space with an RMSE and MAE score of 0.101 and 0.266, respectively. The study also shows that the most significant meteorological inputs in PV yield estimation are air temperature and global solar radiation. The study also noted that a large feature space adds uncertainties to the training process. Finally, the paper concludes that the sunshine-based parameters are key inputs for high-quality PV yield estimation regardless of the number of input combinations. It is also concluded that, in this study, the most accurate regression algorithms for estimation are the Least Absolute Shrinkage and Selection Operator (LASSO), KNN, Elastic Net (EN), Decision Tree regressor (DT), and Extra Tree regression (ET).