Forecasting TEC During Geomagnetic Storms Using deep learning model : Performance and Limitations

This study investigates the prediction of ionospheric total electron content (TEC) during geomagnetic storm periods using a Convolutional Long Short-Term Memory (ConvLSTM) model. TEC maps were reconstructed using the DCGAN-PB (Deep Convolutional Generative Adversarial Network–Poisson Blending) method to ensure high-resolution, gap-free input data. Geomagnetic storm days were selected based on Dst index values below –50 nT, and for each event, a 24-hour dataset was constructed starting from the minimum Dst time. To address the limited number of storm events, overlapping time windows were applied to expand the training dataset. To further improve regional prediction accuracy, a region-weighted loss function was introduced, giving additional emphasis to the Korean Peninsula. Results show that ConvLSTM outperforms both the empirical IRI-2016 model and a baseline model that simply shifts TEC values from 24 hours earlier. The model maintained stable long-term forecasting performance and responded partially to abrupt local TEC variations, though its capacity to fully capture such events was limited. Adding quiet-day data to the training did not significantly improve storm-time performance, suggesting that ionospheric responses during geomagnetic storms are largely independent of prior-day conditions. These findings demonstrate that while ConvLSTM model is effective for general spatiotemporal TEC forecasting, further architectural improvements are needed to better capture event-scale variability.