Random Forest, Support Vector Regression and Gradient Boosting Methods for Ionosphere Total Electron Content Nowcasting Problem at Mid-Latitudes

This paper illustrates data-driven machine learning approach for ionosphere total electron content (TEC) forecasting. The authors exploit different state-of-the-are machine learning algorithms like random forest, support vector regression, and gradient boosting to achive high accuracy (higher than conventional naive and linear models). The proposed approach allows to determine the most important parameters. The approach revealed that current TEC, first time derivative of TEC, cosine from local time LT, current F10.7 and SYM/H indexes, exponential moving averages of TEC (with 12, 24, 96 hour periods), 12h-lagged, 2-days and 15-days lagged F10.7 are the significant features for vertical TEC 4-hour nowcasting model. As the experimental data, the vertical absolute TEC was used. The time resolution of the data is 30 minutes. Initial phase and psueduorange slant TEC were recorded by the mid-latitude station IRKJ (52 N, 104 E) in 2014. All the models were evaluated and testing results comparison provided. Machine learning based models allow us to achive small RMSE and ap; 3 TECU, linear regression model based on significant features results in and ap; 4.5 TECU, while naive models results to huge RMSE.