Global TEC forecasting based on deep learning techniques: a comparative
study and perspectives for a Space Weather operational service
Abstract
The aim of this work is to present a global ionospheric prediction model
based on deep learning (DL) to forecast Total Electron Content 24 hours
in advance under different space weather conditions. Three different DL
techniques have been compared to select the most suitable for the
purpose of an operational service: Long Short Term Memory (LSTM), Gated
Recurrent Units (GRU) and Convolutional Neural Networks (CNN). The
modeling approach inherits and extends what has been proposed by
Cesaroni and co-authors (2020). We use TEC on 18 selected grid points of
Global Ionospheric Maps (GIMs) as the target parameter and Kp index as
the external input. We use a dataset from 2005-2016 for training and
testing, we also analyze case studies from 2017 under different
geomagnetic conditions. Results show that CNN models have better
predictive capabilities than the other two DL models, even under
geomagnetically disturbed conditions. Considering the first 24 hours of
forecasting, CNN exhibits errors between 0.5 and 2 TECu, while LSTM and
GRU errors can reach 3 TECu. We also show how all the proposed models
outperform the two naive models: the so-called “frozen ionosphere” and
a 27 days averaged model.
Moreover, we implemented the models using incremental training to update
them as new data arrives and thus the trained model is able to adapt to
rapid changes within the previous 24 hs to the forecasting. Thus, the
proposed model can be implemented in an operative manner for Space
Weather applications and services.