Forecasting 24-Hr Total Electron Content With Long Short-Term Memory Neural Network

Abstract

An accurate prediction of the ionospheric state is important for correcting ionospheric propagation effects on Global Navigation Satellite Systems (GNSS) signals used in precise navigation and positioning applications. The main objective of the present work is to find a total electron content (TEC) model which gives a good estimate of ionospheric state not only during quiet but also during perturbed ionospheric conditions. For this, we implemented several long short-term memory (LSTM)-based models capable of predicting TEC up to 24 hr ahead. For the first time, we used the solar wind forcing parameters Wprot (a measure of the ionospheric disturbance during storm time) and Econv (measure of the solar wind parameters) as driver parameters. We found that using external drivers does not improve the accuracy of TEC predictions significantly. The final model is trained with data from the last two solar cycles using TEC from the rapid UQRG global ionosphere maps (GIMs). Data from the years 2015 and 2020 were excluded from the training data set and used for testing. The performance of the LSTM-based TEC model is tested for near real-time (RT) cases as well by using RT products (IRTG GIMs) as historical TEC inputs. We compared the performance of the LSTM-based model to a quiet-time feed forward neural network (FNN)-based model and the Neustrelitz TEC model (NTCM). The results indicate that the LSTM-based model proposed here is outperforming the FNN-based model and NTCM in both cases, that is, using the UQRG or the IRTG GIMs as input for the historical TEC.