Abstract
AbstractIt has been shown in ionospheric research that modelling total electron content (TEC) during storm conditions is a big challenge. In this study, TEC modelling was performed over Sutherland (32.38°S, 20.81°E, 41.09°S geomagnetic), South Africa, during storm conditions, using a combination of empirical orthogonal function (EOF) and regression analyses techniques. The neural network (NN) technique was also applied to the same TEC data set, and its output was compared with TEC modeled using the EOF model. TEC was derived from GPS observations, and a geomagnetic storm was defined for Dst≤−50 nT. The hour of the day and the day number of the year, F10.7p and A indices, were chosen as inputs for the modeling techniques to take into account diurnal and seasonal variation of TEC, solar, and geomagnetic activities, respectively. Both EOF and NN models were developed using GPS TEC data for storm days counted from 1999 to 2013 and tested on different storms. For interpolation, the EOF and NN models were validated on storms that occurred during high and low solar activity periods (storms of 2000 and 2006), while for extrapolation the validation was done for the storms of 2014 and 2015, identified based on the provisional Dst index data. A comparison of the modeled TEC with the observed TEC showed that both EOF and NN models perform well for storms with nonsignificant ionospheric TEC response and storms that occurred during period of low solar activity. For storms with significant TEC response, TEC magnitude is well captured during the nighttime and early morning, but short‐term features, TEC enhancement, and depression are not sufficiently captured by the models. Statistically, the NN model performs 12.79% better than the EOF model on average, over all storm periods considered. Furthermore, it has been shown that the EOF and NN models developed for a specific station can be used to estimate TEC over other locations within a latitudinal and longitudinal coverage of 8.7° and 10.6°, respectively. This is an important result as it reduces the data dimensionality problem for computational purposes.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have