Predicting global ionospheric TEC maps using Gaussian process regression

Abstract

The integration of machine learning technology into ionospheric prediction has emerged as a burgeoning field of research. However, several challenges still exist, such as the disregard of delays in real-world data acquisition, variations in accuracy due to different input data and methods. In this study, we present an ionospheric Gaussian process regression (GPR) model with multiple input parameters to forecast global ionospheric total electron content (TEC). We assess two solutions based on GPR model that leverage different ionospheric data, namely the vertical TEC (VTEC) and spherical harmonic coefficients (SHC) prediction solutions. Our findings demonstrate that both GPR models exhibit strong accuracy and stability. Specifically, for 1-day-ahead-predicted global ionospheric maps (GIMs) in 2015 year and 2019 year, the root mean square error (RMSE) of the SHC prediction solution is 4.343 TECU and 1.702 TECU, respectively, while the VTEC prediction solution has RMSE values of 4.321 TECU and 1.673 TECU. Moreover, in high and low solar activity, over 29% and 68% of the absolute residuals are within 1.0 TECU, respectively. Furthermore, we compare the GPR model with conventional methods (such as Adaptive Autoregressive Model (AAR)) and observe that the RMSE of the GPR model is lower than that of the AAR model for GIMs predicted under different solar and geomagnetic activities, with a difference ranging from 0.2 to 0.4 TECU. In addition, the GPR model can provide prediction intervals for the predicted values. These intervals are typically represented using the mean and standard deviation, where the mean represents the predicted value and the standard deviation represents the uncertainty associated with the prediction. Prediction intervals can assist users in understanding the uncertainty of the model's predictions.