Forecasting Total Electron Content During Geomagnetic Storms Using Convolutional Long Short‐Term Memory (ConvLSTM): Performance and Limitations

Abstract This study investigates the effects of quiet time ionospheric conditions and the number of storm events used for training on the prediction of ionospheric total electron content (TEC) during geomagnetic storms using a deep learning method. A Convolutional Long Short‐Term Memory (ConvLSTM) model is employed for training and prediction of regional TEC maps around the Korean Peninsula. To ensure high‐resolution, gap‐free input data, TEC maps were reconstructed using a Deep Convolutional Generative Adversarial Network–Poisson Blending (DCGAN‐PB) method. Geomagnetic storm days were selected based on Dst index values below −50 nT, and for each event, a 24‐hr data set was constructed starting from the minimum Dst time. To address the limited number of storm events, partially overlapping 24‐hr segments were extracted from each storm using a sliding‐window approach to augment the training data set. To further improve regional prediction accuracy, a region‐weighted loss function was introduced, giving additional emphasis to the Korean Peninsula. Results show that the ConvLSTM outperforms both comparison models, achieving a root mean square error (RMSE) of 5.09 TECU compared with 6.23 TECU for the 24‐hr‐lag persistence model and 8.37 TECU for International Reference Ionosphere‐2016. Adding quiet‐day data to the training did not improve storm‐time performance, suggesting that ionospheric responses during geomagnetic storms are independent of prior‐day conditions. However, model's performance improved in proportion to the number of storm events used for training. This result indicates that the availability of storm data is a key factor in accurately predicting storm‐time ionospheric plasma density.