Estimation of Spectral Solar Irradiance Toward Deriving Spectral Mismatch Factor Across Diverse Photovoltaic Technologies by Using Aerosol Optical Properties

This study presents a machine-learning framework for predicting spectral solar irradiance from 300 to 1100 nm using the random forest and neural network models. Two approaches were compared: one using the real aerosol optical depth at discrete wavelengths and the other using the synthetic aerosol optical depth derived from the Ångström exponent. The models were trained on atmospheric variables and were validated against high-resolution spectroradiometer data from the Korea Institute of Energy Research. The models based on the synthetic aerosol optical depth outperformed real-data models in terms of accuracy, bias, and variance explained (γ2 = 0.979 vs. 0.967). These models also provide more reliable estimates of the spectral mismatch factor across the diverse photovoltaic technologies. The copper indium gallium diselenide and mono-crystalline silicon cells showed a high spectral mismatch factor stability, whereas Perovskite cells exhibited a greater sensitivity to spectral variations.