Ionospheric TEC forecast model based on support vector machine with GPU acceleration in the China region

Abstract

The work presents the application of support vector machine (SVM) with Graphics Processing Unit (GPU) acceleration in developing a regional forecast model for the ionospheric total electron content (TEC) over China region. In this study, the SVM model has the past TEC values as inputs as well as 26 input parameters, which include the ionospheric diurnal variation, seasonal variation, spatial variation, solar activity, geomagnetic activity, and thermospheric wind. The output is the TEC values up to 1 h ahead. Datasets for 2016–2017 are used to train the SVM, and datasets for 2018 are selected as the test dataset to verify the SVM model performance. Predictions from the SVM model, back propagation-based NN (BP-NN) model, and International Reference Ionosphere 2016 (IRI2016) model are then compared with the Global TEC grid data released by IGS in China. According to the predicting results, the root-mean-square-error (RMSE) of SVM model ranges from 1.31 to 1.64 TECU, the relative error (RE) is 10.73–15.86%, and the correlation coefficient falls within the range of 0.92–0.99. The BP-NN model’s RMSE varies between 1.61 and 2.26 TECU, RE is between 13.77 and 20.19%, and the correlation coefficient lies in the range of 0.90–0.98. For the IRI2016 model, the RMSE, RE and correlation coefficient ranges are 2.21–5.73 TECU, 27.36–40.31%, and 0.83–0.95, respectively. Combined with the comparison of diurnal variations of TEC, it suggests that the SVM model greatly outperforms the BP-NN and IRI2016 models. Furthermore, the variation of seasonal and local characteristics is also validated by the SVM model. The results indicate that the SVM model accelerated by GPU is very promising for applications in ionospheric studies.