Artificial deep neural network modeling of solar- and atmospheric-driven ground magnetic perturbations at mid-latitude

Abstract

Ground magnetic observatories measure the Earth’s magnetic field and its coupling with the solar wind responsible for ionospheric and magnetospheric current systems. Predicting effects of solar- and atmospheric-driven disturbances is a crucial task. Using data from the magnetic observatory Chambon-la-Forêt at mid-latitude, we investigate the capability of our developed deep artificial neural networks in the modeling of the contributions above 24 hours and the daily variations. Two neural networks were built with the long short-term memory architecture with multiple layers. Using the data from 1995 onwards, the neural networks were trained with physical parameters indicative of solar variabilities and geographical daily and seasonal variations. By excluding the secular variation owing to the change of the Earth’s intrinsic magnetic field, we demonstrate that our approach can model the observed signals with overall good agreements for both a solar-quiet period in 2009 and a solar-active period in 2012. Particularly, using the walk forward training, we updated our models with new data leading up to the test year. The implication of this work is twofold. First, our approach can be adapted for near real-time prediction of intensity of solar and atmospheric disturbances. Second, the neural networks can be used to model the quiet variations when excluding the solar variabilities with important applications in the calculation of magnetic activity indices. This work is a proof-of-concept that deep neural networks can be used to model solar- and atmospheric-driven perturbations modulated by daily and seasonal variations as recorded at a ground magnetic observatory.