Solar Irradiance Prediction Based on Satellite Image Data in Different Regions: A Deep Learning-Based Approach

Abstract

Solar irradiance is considered a fundamental parameter for photovoltaic systems. Stations established worldwide to measure solar irradiance are limited in terms of coverage. Therefore, many solar irradiance prediction models and techniques are being developed to obtain solar irradiance data. The influence of climatic factors on solar irradiance, its non-stationary nature, random variability, and complexity make prediction challenging. This study presents a short-term (minute) approach to predict global solar irradiance on a horizontal plane from satellite images, addressing all complexities and without relying on ground stations, providing broad coverage. In the proposed approach, convolutional neural network-based deep neural network architectures are employed to extract convolutional features from satellite images for the purpose of global solar irradiance on a horizontal plane prediction. Data obtained from satellite images were trained in an artificial neural network and tested and validated in regions of Turkey with different climatic conditions. Subsequently, the performance of the model is assessed and compared using common performance metrics in the literature, such as root mean square error and correlation coefficient (R), without considering the data of the predicted regions. The obtained results show that the proposed prediction model demonstrates high performance in some regions with an R value of over 97%. The prediction model may perform better if the larger training dataset is more similar to the test dataset. Indeed, successful prediction results have been obtained in regions located close to the training dataset in some areas. These results highlight that deep learning architectures can make significant contributions to research on renewable energy sources, such as solar energy.