Exploring possibilities for solar irradiance prediction from solar photosphere images using recurrent neural networks

Studies of the Sun and the Earth's atmosphere and climate consider solar variability and its constant monitoring an important driver in climate models. Solar irradiance is one of the main parameters that allow monitoring this variation, which can be studied in spectrum ranges or in its version that integrates all those ranges. Some physical and semi-empirical models were developed and made available and are very relevant for the reconstruction of irradiance in periods of data failure or absence in the collection. However, the solar irradiance prediction could benefit ionospheric and climate models through prior knowledge of irradiance values hours or days ahead, without the need to know or have available other parameters that would be necessary for their reconstruction. This paper presents a neural network based approach, which uses images of the solar photosphere to extract sunspot and active region information and thus generate inputs for recurrent neural networks to perform the irradiance prediction. Experiments were performed with two recurrent neural network architectures for short- and long-term predictions of total and spectral solar irradiance along three wavelengths. The results show good quality of prediction results for TSI and motivate individual effort in improving the prediction of each type of irradiance predicted in this work. The results obtained for SSI point out that photosphere images do not have the same influence on the prediction of all wavelengths tested, but encourage the bet on new spectral lines prediction. The quality closeness in neural networks and physical models results raise the possibility that prediction is an option to be considered in studies for which only reconstructed data are considered.