Ionospheric Total Electron Content Forecasting at a Low-Latitude Indian Location Using a Bi-Long Short-Term Memory Deep Learning Approach

Forecasting ionospheric total electron content (TEC) has been of great interest among ionospheric researchers and radio propagation scientists as it critically affects the operation of dynamic technological systems relying on space-based radio propagation. In this context, this article attempts to forecast ionospheric TEC using a bi-long short-term memory (Bi-LSTM) deep learning neural network model at a low-latitude location (KLEF Campus; geographic 16(degrees)26'N, 80(degrees)37'E; dip 22(degrees)11') in Guntur, India. We employed the TEC observables from a specialized global navigation satellite system (GNSS) ionospheric scintillation and TEC monitoring (GISTM) receiver as input data for the deep learning model. To include external ionospheric characteristics, geomagnetic and solar indices (Kp, ap, F10.7, sunspot number (SSN), and Sym-H) were also considered whose individual effects are realized before incorporating them in the model. The performance of the Bi-LSTM model was compared with the conventional LSTM, time-series autoregressive integrated moving average (ARIMA), the latest edition of global empirical international reference ionosphere (IRI-2020) models, and the TEC extracted from hourly global ionospheric maps (GIMs) using statical metrics like root-mean-square error (RMSE), maximum absolute error (MAE), and R error measurements. The Bi-LSTM model showed a low RMSE value of 7.73 TEC unit (TECU), while the LSTM and ARIMA models had values of 8.16 and 8.39 TECU, respectively. The Bi-LSTM model also had the highest correlation (R = 0.962), followed by the LSTM (R = 0.957) and ARIMA (R = 0.956). The day-to-day performance of the TEC forecast was evaluated for quiet and disturbed days, confirming a substantial performance irrespective of geomagnetic activity. Further assessment of the model performance at other latitudes in a wider range of the Indian longitude sector may provide more insight into its predictability for regional deployment.