Modelling Ionospheric State and Dynamics Using Machine Learning Techniques

Kurzfassung

Our society becomes increasingly dependent on reliable space-based observations. For instance, millions of people across the world rely on Global Navigation Satellite Systems (GNSS) for daily navigation. The radio signals transmitted by GNSS can be delayed or advanced by the ionosphere. To correct ionospheric propagation effects, it is essential to provide accurate ionospheric predictions. Using the dispersive (frequency dependent) characteristics of the ionosphere, the ionospheric delay can be cancelled out using multiple frequencies. However, this technique does not apply to single frequency users who require ionospheric models to correct for this propagation effect. The ever-changing state and dynamics of the ionosphere can be observed by using the total electron content (TEC) which is the line-in-sight number of free electrons, where one TEC unit (TECU) equals 10^16 free electrons/m2. The International GNSS Service (IGS) has been providing global ionosphere maps (GIMs) since 1998, covering more than two solar cycles of data, making this large dataset useful for training neural networks (NNs). Space weather events (e.g., geomagnetic storms) can have huge effects in the ionosphere, where increases (or decreases) in TEC of more than 200% have been observed during extreme events. It is, therefore, important that ionospheric models also perform well during periods of geomagnetic disturbance. In this work several NN-based solutions are deployed for providing predictions of the ionospheric state (TEC), not only during quiet time but also during geomagnetically disturbed conditions. In total, three models have been developed and presented in this dissertation using the feedforward neural network (FNN) or the long short-term memory (LSTM) architecture. The first model is the global quiet-time model using an FNN architecture. This model is trained using GIMs from the Center for Orbit Determination in Europe (CODE). To reduce the dimensionality of the data, Carrington rotation averages (approximately 27 days) were used. The model uses the solar flux index F10.7, solar zenith angle, geographic longitude, geomagnetic latitude, universal time (UT) and day of year (DOY) as input parameters. The quiet-time model can predict not only large-scale features, such as seasonal behaviour or diurnal variations, but also small-scale features such as the separation of the ionization crests at both sides of the geomagnetic equator and the Nighttime Winter Anomaly (NWA). The ability of the model to reproduce the evolution of the NWA shows how well this approach can learn very small features from the data. The performance of the quiet-time model in terms of root mean square error (RMSE) is in the order of 6 and 2.5 TECU for the high solar activity year 2015 and low solar activity year 2020, respectively. The quiet-time model outperforms the Neustrelitz TEC Model (NTCM) by approximately 1 TECU. The quiet-time model does not use any geomagnetic indices, and the performance during geomagnetic disturbed periods decreases. Therefore, the second model proposed in this thesis focuses on storm-time modelling. For this work an FNN-based European storm-time model was proposed that predicts the relative TEC with respect to the preceding 27-day median TEC, using the TEC from GIMs produced by the Universitat Politècnica de Catalunya (UPC). The network uses the median TEC, geographic latitude and longitude, UT, storm time, DOY, F10.7, global storm index SYM-H and geomagnetic activity index Hp30 as driver parameters. The model was trained, validated and tested using a database of about 400 storms. An event was marked as a storm when the disturbance storm-time Dst index was < - 50 nanotesla. This model is able to show seasonal storm-related perturbations, i.e., increased TEC values for winter storms, negative storm phases in case of summer storms and a mixture of both during equinox storms. The European storm-time model has an RMSE of 3.38 TECU for the test dataset consisting of 33 storm events during 2015 and 2020, which is a performance increase of approximately 1.3 TECU compared to the quiet-time model. Providing relative TEC values instead of the direct TEC has the advantage that other inputs can be used for the median TEC, e.g., our quiet-time model or other models such as the NTCM. The last ionospheric model investigated during this work was a global LSTM-based model. This model is able to provide forecasts up to 24 hours ahead and uses only the 3-day historic TEC, geographic latitude and longitude, UT and DOY. The performance of the model was investigated during high and low solar activity conditions as well as during geomagnetic disturbed and quiet periods. The average RMSE for high solar conditions (the year 2015) and low solar conditions (the year 2020) is 3.3 and 1.5 TECU, respectively. During storm conditions the RMSE increased to 4.2 TECU in 2015 and did not significantly change in case of the year 2020 (0.1 TECU increase). The performance was also compared to the performance of the quiet-time model and the NTCM. Near real-time applications were also investigated by providing the real-time GIMs from IGS as input. The LSTM-based model was outperforming the quiet-time and NTCM model for all cases. This dissertation demonstrates that NN-based models are capable of providing accurate TEC predictions including small-scale features not previously seen in predictions by other models. It shows that the machine learning-based models can successfully describe storm-related perturbations and have a good accuracy during geomagnetic storms. This work also demonstrates near real-time applications such as using the real-time products from IGS as input for the historical TEC. The findings of this dissertation contribute to a more accurate and stable prediction of the ionospheric TEC that can correct GNSS signals.