Boosting total electron content forecasting based on deep learning toward an operational service

Abstract

We present a prediction model based on deep learning able to forecast ionospheric Total Electron Content at global level 24 h in advance. It has been conceived to operate under different space weather scenarios and in an operational framework. Three different deep learning (DL) techniques have been compared: Long Short Term Memory (LSTM), Gated Recurrent Units (GRU) and Convolutional Neural Networks (CNN). The modelling approach inherits by and extends what has been proposed by Cesaroni and co-authors (2020a). Specifically, the machine learning-based approach here reported is conceived to improve the first step of Cesaroni et al. (2020a), in which TEC is forecasted on 18 selected grid points of Global Ionospheric Maps (GIMs) using the geomagnetic global index Kp index as the external input.

CNN models provide better predictive capabilities than LSTM and GRU, and it has more robust behaviour under different space weather conditions. We also show how all the proposed models outperform the two naive models: the so-called “frozen ionosphere” or recurrence model and a 27 days averaged model.

The novelty of our approach is the operational capability based on an incremental learning method to prevent the aging of the trained models by updating the weights with little computational effort adding new information immediately after the 24-h forecasting. The improvement changed from RMSE of ∼6.5 TECu to ∼2.5 TECu.We also discuss limitations and the use of other space weather inputs (e.g. solar proxies, other geomagnetic indexes, etc) and the use of complementary data science techniques (e.g. data preparation, hyperparameter tuning, better data resolution, etc.) to enhance the forecasting in future works.