Analysis of Deep Learning Architectures for Satellite Imaging-based Solar Irradiance Dense Time-Series Forecasting

Accurate forecasting of solar irradiance is essential for facilitating the integration of renewable energy into power systems and mitigating carbon emissions. The inherently dynamic nature of solar irradiance, shaped by temporal and spatial variability, necessitates the adoption of advanced time-series deep learning models. In this study, state-of-the-art deep learning architectures were evaluated for satellite-based solar irradiance time-series forecasting, with a specific emphasis on their application in smart grid management. Recurrent models, including ConvLSTM, PredRNN++, and PredRNNv2, were benchmarked against attention-based models such as SimVPv2, MMVP, and TAU, utilizing the MCD18A1 dataset while accounting for seasonal variations in Taoyuan City, Taiwan. The experimental results revealed that attention-based models, particularly TAU, outperformed recurrent models in both summer and winter scenarios. For the summer season, TAU achieved a root mean square error (RMSE) of 19.64 W/m<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> and a peak signal-to-noise ratio (PSNR) of 34.5 dB. In the winter season, TAU achieved an RMSE of 29.27 W/m<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> and a PSNR of 48.83 dB. Compared to ConvLSTM models, TAU demonstrated a 10% improvement in test RMSE and a 13% improvement in PSNR. These findings underscore the critical role of advanced deep learning frameworks in enhancing the accuracy of solar irradiance forecasting, thereby contributing to more efficient energy management and improved grid resilience.