Machine learning total electron content models based on F10.7

Abstract

The ionospheric total electron content (TEC) is a crucial parameter to calculate the ionospheric corrections for global navigation satellite systems and other systems that utilize the ionospheric high frequency radio band. To date, a number of empirical models have been developed to estimate both current TEC and forecasted TEC. However, with the exception of operational models, for which input parameters are broadcast, users do not typically have access to such parameters for other models. We present a methodology that enables the generation of the global TEC maps with a minimal set of input parameters (and their histories) for different forecast horizons, obviating the need to evaluate input parameters for the forecast date. The methodology employs the F10.7 index as the primary control parameter for the models, with additional features based on date and time. Various machine learning models were constructed, including gradient boosting, neural networks with recurrent LSTM cells, fully connected neural networks, and linear regression models for forecast horizons of 1, 3, 7, 30, 60, and 180 days. The main idea of the research was to change the generally used approach when you have to forecast drivers for a model before ionosphere forecasting. Our difference is to use previous indexes for TEC forecast. So to forecast TEC maps, there is no need to have the F10.7 index value for a forecast date. The created machine learning models demonstrate high quality. For example a 1-day forecast TEC model based on fully connected neural network / neural network with LSTM cells / gradient boosting / linear regression exhibited a root mean square error (RMSE) of 6.61 / 6.80 / 7.05 / 10.55 TECU, mean absolute error (MAE) of 4.51 / 4.70 / 5.03 / 7.87 TECU, and mean absolute percentage error (MAPE) of 20.87 % / 21.99 % / 25.57 % / 46.93 %, respectively. GEMTEC, IRI-2016 and Klobuchar models (used as a reference) demonstrated a RMSE of 10.52 TECU, 11.08 TECU and 15 TECU, respectively.