Ionospheric Electron Density Profiles Prediction using Temporal Deep Learning Models

Detailed ionospheric parameters during certain time periods, e.g. the electron density (Ne), can be observed by the incoherent scatter radar (ISR). However, the ISR observations are sporadic and the models for their prediction are in need. Previously, we developed static perceptron-based neural network (PNN) models for Ne prediction aided by neural architecture search (NAS). In this work, we treat Ne as a dynamic time series at different altitudes and further develop Ne prediction models using recurrent neural network (RNN) or transformer-based neural network (TNN). The Millstone ISR data are used to train and test these models. Our results show that the dynamic models, RNN and TNN, which take the historical data into the modeling mechanism, can predict Ne better than the static PNN model and the empirical model, International Reference Ionosphere (IRI).