Machine Learning Forecast of Ionosphere Total Electron Content

Abstract:

In the current era, the ionospheric total electron content (TEC) derived from multi-frequency Global Navigation Satellite System (GNSS) receiver is arguably the most utilized dataset in the ionospheric research area, and also has essential practical importance, as it is the largest naturally occurring error source for GNSS positioning, navigation, and timing (PNT) accuracy. The potential of using the GNSS data as a backbone of the space weather observational system has been demonstrated in the last decade with the GPS system, and as we moving into the multi-GNSS era, we are at the forefront of a new chapter by combining the traditional space science and the modern machine learning (ML) algorithms to make a leap forward in the specification and forecasting of ionosphere state and variability.

We use state-of-the-art ML algorithms to specify and forecast local, regional and global ionosphere TEC and its variability. Input of the forecast models includes actual and predicted TEC, solar wind and FISM solar radiation measurements. The single-station model, based on the Long Short-Term Memory (LSTM) networks, leverages on TEC values derived from GPS data collected at Continuously Operating Reference Stations (CORS), and can make predictions with the root mean square error (RMSE) less than 3 TEC units. In the global case, several data imputation and reconstruction techniques are firstly used to fill gaps in the global TEC maps. The completed maps are then used to develop Gaussian process based and neural network based forecast models.