Solar Energy Datasets of Deep Learning Models Incorporating with GK-2A and ASOS Ground Measurements

This study presents the construction and evaluation of a dataset for estimating solar energy using the GK<sup>-2</sup>A satellite and deep learning. The GK<sup>-2</sup>A is currently utilized in real-time for weather observations over the Korean Peninsula. The GK<sup>-2</sup>A satellite features 16 channels, producing radiative channel images at spatial resolutions ranging from 500 m to 2 km, with temporal intervals as short as 2 minutes depending on the area. These satellite data are used in various fields, including meteorology, oceanography, vegetation monitoring, and renewable energy. In this study, we used spectral channel data from the GK<sup>-2</sup>A expended local area satellite from January 2021 to December 2022. For training and evaluating the accuracy of the deep learning model, we utilized data from 98 automated synoptic observing system ground observation sites operated by the Korea Meteorological Administration. A back-propagation neural network model, which showed meaningful results in estimating solar energy, was applied. Various hyperparameters were optimized, and data preprocessing and separation were conducted to optimize the model. The study also compared the performance of the deep learning model with physical models. The BPNN deep learning model achieved a statistical accuracy of root mean squared error (RMSE) 77.32 Wm<sup>-2</sup>, mean bias error (MBE) -0.48 Wm<sup>-2</sup>, and R<sup>2</sup> 0.91, indicating high accuracy. In contrast, the physical model showed an RMSE of 132.01 Wm<sup>-2</sup>, MBE -76.51 Wm<sup>-2</sup>, and an R<sup>2</sup> of 0.74, displaying relatively lower accuracy compared to the deep learning model. Additionally, the spatio-temporal map of solar energy generated by the deep learning model successfully captured the attenuation of radiation due to clouds and the variation in solar energy based on the position of the sun. The solar energy data produced in this study are expected to be useful as input data for various fields such as meteorology, agriculture, environmental monitoring, and marine sciences.