High latitude ionospheric scintillation forecasting using Deep Learning
Résumé
Irregularities of ionospheric layer can lead to rapid fluctuations of the amplitude, phase and direction of arrival of a received GNSS signal. Recent advances in deep-learning bring many approaches to study ionospheric scintillation. So that this paper presents models belonging to the field of artificial intelligence, based on solar wind and geomagnetic indices, to predict Rate of TEC index (ROTI) which is well related to high latitude ionospheric disturbances [Fabbro et al. 2021]. To forecast ROTI index taking into consideration the complex and nonlinear dynamics of the Earth’s magnetosphere and ionosphere, nonlinear models such as Recurrent Neural Network (RNN) have been investigated. First, we have worked on the development and the optimization of Long Short Term Memory (LSTM) neural networks. These models have proven their efficiency in space weather domain to forecast geomagnetic indices specific to Earth magnetic current systems, like am index in [Gruet, 2018]. ROTI databases used come from the Greenland GNSS Network (GNET) and the Norwegian Mapping Authority (NMA). The advantage of such databases is their latitude coverage, offering a large geographic survey of the ionosphere conditions in Arctic region during the 2010-2020 period. ROTI measurements have been aggregated into time series using similar, but not identical, super-observations as introduced in [McGranaghan et al., 2018]. Superobservations should be considered as a statistical summary of the ground receiver-to-individual
GNSS satellite links available at a given time and in a given Magnetic Latitude (MLAT) interval. Feature parameters that permit to predict ROTI time series are extracted from the fiveminute resolution HRO dataset, which is provided, by the National Aeronautics and Space Administration (NASA)’s Space Physics Data Facility. The Kendall correlation coefficients have been used to study the existence of a potential relationship between each HRO feature and ROTI super-observations for a 1-hour lead-time forecast. A special attention has been paid to provide operational models based on data recorded by the ACE satellite located at the Lagrangian point L1 and ground geomagnetic indexes computed with magnetometers data on the northern hemisphere, available in near real-time. Finally after studying the ability of LSTM neural networks to predict the ROTI index, a model to predict ROTI probability of occurrence at high latitude, based on LSTM neural networks and using lognormal ROTI distribution has been developed and evaluated.