Forecasting Local Ionospheric Parameters Using Transformers

TL;DR

The paper presents LIFT, a transformer-based framework for probabilistic, local forecasts of foF2, hmF2, and TEC with uncertainty quantified via nonparametric quantiles. By combining a linear baseline with a transformer that ingests a rich set of exogenous drivers and climatology, the method delivers 24-hour forecasts and outperforms the IRI climatology, while generalizing to unseen locations. The approach leverages an encoder-decoder transformer with convolutional embeddings and a hybrid loss that jointly optimizes linear trends and nonlinear residual quantiles, yielding calibrated prediction intervals. The results demonstrate practical value for space weather operations, including improved radio-propagation planning via uncertainty-aware forecasts, and point to future work on global grids and enhanced tail calibration.

Abstract

We present a novel method for forecasting key ionospheric parameters using transformer-based neural networks. The model provides accurate forecasts and uncertainty quantification of the F2-layer peak plasma frequency (foF2), the F2-layer peak density height (hmF2), and total electron content (TEC) for a given geographic location. It supports a number of exogenous variables, including F10.7cm solar flux and disturbance storm time (Dst). We demonstrate how transformers can be trained in a data assimilation-like fashion that use these exogenous variables along with naïve predictions from climatology to generate 24-hour forecasts with non-parametric uncertainty bounds. We call this method the Local Ionospheric Forecast Transformer (LIFT). We demonstrate that the trained model can generalize to new geographic locations and time periods not seen during training, and we compare its performance to that of the International Reference Ionosphere (IRI).

Figures (42)

• Figure 1: Geographic map of GIRO stations used. The size of each marker indicates roughly the relative amount of data obtained from each location. Larger marker sizes represent more valid data obtained between the years 2000 to 2023. Some stations only had a few segments of usable data and thus are not visible in this figure.
• Figure 2: All data for the parameters foF2, hmF2, and TEC obtained from a Digisonde sounder in Boulder, Colorado, USA from 1 January 2000 to 1 January 2023.
• Figure 3: Number of observations obtained for training (blue), validation (green), and testing (red) grouped by year.
• Figure 4: Indices used as exogenous inputs to the model for the time period 2004-2023.
• Figure 5: An illustration of how the LIFT model generates point forecasts and uncertainty quantification. Each row provides results for a different target variable (foF2, hmF2, TEC). The linear (first column) and nonlinear (second column) components add together to generate an estimate of the quantile distribution for each target (third column). The final median prediction for each target variable is represented by the bold weighted lines in the third column.