Leveraging Satellite Imagery and Deep Learning for Solar Radiation Forecasting in Data-Scarce Regions

In this paper, an Artificial Intelligence based methodology for solar radiation forecast in areas where data is limited is presented, based on satellite images and deep learning approaches. Reliable ground-based solar measurements are still limited, especially in developing countries, which is still a significant challenge in the deployment of solar energy. The proposed solution is to use widely available satellite data that can be derived from data on cloud cover images and land surface indicators as proxies for solar resource assessment. To enhance forecasting accuracy over different geographical regions, the use of various deep-learning architectures such as convolutional neural networks, transformers, and spatiotemporal models is explored. Transfer learning and domain-adaptation techniques are also incorporated into the framework allowing transfer of models to areas with lower historical data. In addition, pre-processing methods like normalization, interpolation, augmentation and feature engineering are used to improve the robustness of the model. Methods to aid reliable decision-making and investment planning, such as probabilistic forecasting, uncertainty quantification and calibration methods are included. The proposed approach has promising prospects for enhancing solar-resource assessment, planning for renewable-energy expansion, and for sustainable energy development in areas with a significant data deficiency.