Hybrid prediction method of solar irradiance applied to short-term photovoltaic energy generation

One of the most promising renewable energy sources used as a solution to supply the increase in electricity consumption is photovoltaic solar energy. This source has intrinsic and uncontrollable, peculiarities that cause intermittencies in its generation, due to climatic factors. Therefore, it becomes relevant to the existence of solutions for the prediction of solar photovoltaic energy generation, enabling increased security in the generation and distribution of electricity. In this context, this research proposed a new hybrid prediction method applicable to short-term solar irradiance. The proposed method employs a set of image processing metrics, to extract all-sky image features, used as input in machine learning-based prediction models. The set of all-sky image metrics, extracted with image processing, represent complementary characteristics of sky that provided an overall average accuracy ≅ 30 % higher compared to using traditional meteorological information. The experimental results of the hybrid prediction method with Artificial Neural Network and Light Gradient Boosting Machine, considering six short-term prediction horizons, showed an overall average prediction accuracy of 17.5 % better than the Persistence model. The proposed approach is more interpretable than several literature studies, demonstrated competitive results with more robust deep learning models and, represents a new path for future studies in the prediction of solar photovoltaic energy generation. [Display omitted] •New hybrid prediction method that combines image processing and machine learning.•Evaluation of a new set of explicit metrics extracted from all-aky images.•Promising results and competitive with literature studies using deep learning.•The proposed prediction method has a “gray-box” approach that can be optimized.