Prediction of Global Ionospheric TEC based on Deep Learning and Singular Spectrum Analysis

To predict global ionospheric total electron content (TEC), this paper applies Singular Spectrum Analysis (SSA) for the extraction of trend, periodic, and irregular components form Spherical Harmonic (SH) coefficients. It further employs Convolutional Long Short-Term Memory (ConvLSTM)-based machine learning models to construct an encoder-decoder neural network for short-term prediction of TEC. The training set, ranging from March 2017 to October 2018, includes the trend, periodic, and irregular components form the SH coefficients, ap indices, DST indices, days of the year, and hours of the day. For testing set 1, which spans from November 2016 to February 2017, the root mean squared error (RMSE) is 0.72 TECU and 1.20 TECU for the first- and second-hour predictions, respectively ${(1 T E C U}=10^{16} \text{el} / \mathrm{m}^2)$ . For testing set 2, covering the period from November 2018 to February 2019, the RMSE of the prediction results is 0.55 TECU and 0.90 TECU.