Data Augmentation-Based Estimation of Solar Radiation Components without Referring to Local Ground Truth in China

Abstract

The power generation of bifacial photovoltaic modules is greatly related to the diffuse solar radiation component received by the rear side, but radiation component data are scarce in China, where bifacial solar market is large. Radiation components can be estimated from satellite data, but sufficient ground truth data are needed for calibrating empirical methods or training machine learning methods. In this work, a data-augmented machine learning method was proposed to estimate radiation components. Instead of using observed ground truth, far more abundant radiation component data derived from sunshine duration measured at 2,453 routine weather stations in China were used to augment samples for training a machine-learning-based model. The inputs of the model include solar radiation (either from ground observation or satellite remote sensing) and surface meteorological data. Independent validation of the model at Chinese stations and globally distributed stations demonstrates its effectiveness and generality. Using a state-of-the-art satellite product of solar radiation as input, the model is applied to construct a satellite-based radiation component dataset over China. The new dataset not only outperforms mainstream radiation component datasets, but also has significantly higher accuracy than satellite-based datasets derived from other machine learning methods trained with limited observations, indicating the superiority of our data-augmented method. In principle, this model can be applied on the global scale without additional training with local data.