In this paper, we proposed an innovative encoder‐decoder structure with a convolution long short‐term memory (ED‐ConvLSTM) network to forecast global total electron content (TEC) based on the International GNSS Service (IGS) TEC maps from 2005 to 2018 with 1‐hr time cadence. The ED‐ConvLSTM model is used to forecast TEC maps 1–7 days in advance through iterations. To investigate the model's performance, we compared the model with International Reference Ionosphere (IRI2016) model in 2014 and 2018, and compared the model with 1‐day Beijing University of Aeronautics and Astronautics (BUAA) model in 2018. The results show that our 7‐day ED‐ConvLSTM model (ED‐ConvLSTM model that forecasts 7 days in advance) outperforms IRI2016 in 2014 and 2018, and our 5‐day ED‐ConvLSTM model (ED‐ConvLSTM model that forecasts 5 days in advance) outperforms 1‐day BUAA model. Furthermore, the root mean square error (RMSE) from the 1‐day ED‐ConvLSTM model with respect to the IGS TEC maps decreases by 51.5% and 43%, respectively, in 2014 and 2018 compared with that from IRI2016 model. The RMSE from the 1‐day ED‐ConvLSTM model is 20.3% lower than that from the 1‐day BUAA model in 2018. In addition, our model has the highest RMSE in the Equatorial Ionospheric Anomaly (EIA) region, but can roughly predict the features and locations of EIA. However, the model fails to forecast localized TEC enhancement and the sudden ionospheric response to the geomagnetic storms. Overall, the model shows competitive performance in medium‐term global TEC maps prediction during geomagnetic quiet periods.

In recent years, transformer has been widely used in natural language processing (NLP) and computer vision (CV). Comparatively, forecasting image time sequences using transformer has received less attention. In this paper, we propose the conv-attentional image time sequence transformer (CAiTST), a transformer-based image time sequences prediction model equipped with convolutional networks and an attentional mechanism. Specifically, we employ CAiTST to forecast the International GNSS Service (IGS) global total electron content (TEC) maps. The IGS TEC maps from 2005 to 2017 (except 2014) are divided into the training dataset (90% of total) and validation dataset (10% of total), and TEC maps in 2014 (high solar activity year) and 2018 (low solar activity year) are used to test the performance of CAiTST. The input of CAiTST is presented as one day’s 12 TEC maps (time resolution is 2 h), and the output is the next day’s 12 TEC maps. We compare the results of CAiTST with those of the 1-day Center for Orbit Determination in Europe (CODE) prediction model. The root mean square errors (RMSEs) from CAiTST with respect to the IGS TEC maps are 4.29 and 1.41 TECU in 2014 and 2018, respectively, while the RMSEs of the 1-day CODE prediction model are 4.71 and 1.57 TECU. The results illustrate CAiTST performs better than the 1-day CODE prediction model both in high and low solar activity years. The CAiTST model has less accuracy in the equatorial ionization anomaly (EIA) region but can roughly predict the features and locations of EIA. Additionally, due to the input only including past TEC maps, CAiTST performs poorly during magnetic storms. Our study shows that the transformer model and its unique attention mechanism are very suitable for images of a time sequence forecast, such as the prediction of ionospheric TEC map sequences.

No abstract

In this paper, a deep learning long-short-term memory (LSTM) method is applied to the forecasting of the critical frequency of the ionosphere F2 layer (foF2). Hourly values of foF2 from 10 ionospheric stations in China and Australia (based on availability) from 2006 to 2019 are used for training and verifying. While 2015 and 2019 are exclusive for verifying the forecasting accuracy. The inputs of the LSTM model are sequential data for the previous values, which include local time (LT), day number, solar zenith angle, the sunspot number (SSN), the daily F10.7 solar flux, geomagnetic the Ap and Kp indices, geographic coordinates, neutral winds, and the observed value of foF2 at the previous moment. To evaluate the forecasting ability of the deep learning LSTM model, two different neural network forecasting models: a back-propagation neural network (BPNN) and a genetic algorithm optimized backpropagation neural network (GABP) were established for comparative analysis. The foF2 parameters were forecasted under geomagnetic quiet and geomagnetic disturbed conditions during solar activity maximum (2015) and minimum (2019), respectively. The forecasting results of these models are compared with those of the international reference ionosphere model (IRI2016) and the measurements. The diurnal and seasonal variations of foF2 for the 4 models were compared and analyzed from 8 selected verification stations. The forecasting results reveal that the deep learning LSTM model presents the optimal performance of all models in forecasting the time series of foF2, while the IRI2016 model has the poorest forecasting performance, and the BPNN model and GABP model are between two of them.

In this paper, we present a novel approach to improve the accuracy of TEC prediction through data augmentation. Prior works that adopt various deep-learning-based approaches suffer from two major problems. First, from a deep model perspective: LSTM models exhibit low performance on long-term data dependency, while self-attention-based methods ignore the temporal nature of time series, which results in an information utilization bottleneck. Second, the existing TEC actual data is limited and existing generative models fail to generate sufficient high-quality datasets. Our work leverages a two-stage deep learning framework for TEC prediction, stage 1: a time series generative model synthesis of sufficient data close to real data distribution, and stage 2: an Anto-correlation-based transformer to model temporal dependencies by presenting series-wise connections. Experiment on the 2018 TEC testing benchmark demonstrates that our method improves the accuracy by a large margin. The models trained on synthetic data had a notably lower RMSE of 1.17 TECU, while the RMSE for the IRI2016 model was 2.88 TECU. Our results show that the model significantly reduces monthly RMSE, displaying higher reliability in mid, high, low latitudes. Our model shows higher reliability and significantly reduces monthly RMSE and latitude RMSE. However, although our model performs better than IRI2016, low latitudes RMSE needs improvement, as values are generally above 2.5 TECU. This finding has important implications for the development of advanced TEC prediction models and highlights the potential of transformer models trained on synthetic data for a range of applications in ionospheric research and satellite communication systems.