The Use of Long Short-Term Memory Artificial Neural Networks for the Global Prediction of Atmospheric Density

Abstract

Atmospheric drag is one of the most significant disturbing forces in the orbit determination process for satellites in low-Earth orbits. Subsequently, the prediction of atmospheric density is one of the largest sources of error within the calculation of atmospheric drag. Thus, the accuracy of the orbit determination process for satellites in low-Earth orbits, which affects the mission operations of such satellites, is limited by the accuracy of atmospheric density predictions. Atmospheric density models, such as the Mass Spectrometer Incoherent Scatter (MSIS) models and the Jacchia models, provide global predictions of atmospheric density, but such models can have large errors due to the complex relationship between atmospheric density and space weather phenomena. Now, artificial neural networks have been proven to be a valid method for atmospheric density predictions with accuracies meeting or, in most cases, exceeding the accuracies of atmospheric density models. However, previous research on the use of artificial neural networks for atmospheric density predictions has focused on localized, not global, predictions of atmospheric density. In other words, rather than being used to develop a global model, artificial neural networks have been used to forecast atmospheric density along the orbit of a particular satellite. Thus, this research focused on the development of a global model for atmospheric density using artificial neural networks, particularly artificial neural networks with long short-term memory (LSTM) units, trained and tested on data from the Challenging Minisatellite Payload (CHAMP) and the Gravity Recovery and Climate Experiment (GRACE). Overall, for higher solar and geomagnetic activity, the LSTM artificial neural networks were more accurate than NRLMSISE-00 and JB2008 more consistently when tested on shorter timespans. For lower solar and geomagnetic activity, the LSTM artificial neural networks were consistently more accurate than NRLMSISE-00 and JB2008, regardless of the length of the timespan for the testing data.