Comparison of physical and machine learning models for estimating solar irradiance and photovoltaic power

Abstract

Conventional models to estimate solar irradiance and photovoltaic power rely on physics and use empirical correlations to handle regional climate and complex physics. Recently, machine learning emerges as an advanced statistical tool to construct more accurate correlations between inputs and outputs. Although machine learning has been applied for modeling solar irradiance and power, no study has reported the accuracy improvement by machine learning compared to conventional physical models. Hence, this study aims to compare the accuracies of physical and machine learning models at each step of solar power modeling, i.e., modeling of global horizontal irradiance, direct normal irradiance, global tilted irradiance, and photovoltaic power. Comparison results demonstrated that machine learning models generally outperform physical models when input parameters are appropriately selected. Machine learning models more significantly reduced the mean bias difference (MBD) than the root mean square difference (RMSD). For global horizontal irradiance and photovoltaic power, machine learning models led to substantially unbiased estimations with 0.96% and 0.03% of MBD, respectively. Among machine learning algorithms, long short-term memory and gated recurrent unit were more recommendable. However, the physical model for solar power estimation was more efficient to reduce RMSD because of their ability to consider constant parameters as input.