Forecasting of Global Vertical Total Electron Content Based on Trigonometric B-Spline with Long Short Term Memory

Abstract

Short term forecasting of Ionosphere is not only an important topic for both near real-time applications such as single frequency point positioning and navigation, but also monitoring the ionosphere by data assimilation methods. In this study, short term forecasting of global ionosphere on the basis of Trigonometric B-splines is stud- ied with both Deep Learning methods such as LSTM and also conventional methods such as SARIMA. In addition, dimension reduction with Principal Component Anal- ysis is also investigated. The Trigonometric B-spline coefficient time series of Global Ionosphere is obtained by generating coefficients using approximately 20 years of IGS global ionosphere maps in IONEX format. After examining the data, two different methods are proposed on the basis of trends. One is assuming the trend as constant, and the other is a combination of linear and annual trend by Facebook Prophet li- brary. Performance of LSTM and SARIMA models are investigated in the forecasting of individual B-spline coefficient, and also in terms of forecasting Spatial Mean and Principal Components. In addition, a block based LSTM model is also proposed. Best model for each method is established by means of hyper parameter search. Then these best models are compared on days of both quiet and storm ionospheric conditions. According to the results, the combination of dimension reduction with SARIMA model performs better in both quiet and storm days, with 56.17% and 32.59% improvement with respect to persistent ionosphere model, respectively. The proposed block based LSTM model and PCA LSTM provide close results to the SARIMA model with 15% and 21% improvement especially around 00:00 UT. In addition, up to 56% improvement is achieved in the PCA SARIMA model in 2010 on selected days. Although LSTM provides a blackbox model building, feature engineering based on SARIMA model parameters in LSTM models may provide better results.