Authors: Jihyeon Son (Korea Astronomy & Space Science Institute, South Korea), Yong-Jae Moon (Kyung Hee University, South Korea), Young-Sil Kwak (Korea Astronomy & Space Science Institute, South Korea)

In this study, we propose a deep learning model to forecast the SYM-H index up to 12 hours ahead under strong southward interplanetary magnetic field (IMF) Bz conditions (Bz -3nT for at least 6 hours), which is well-known as the main driver of geomagnetic storms. The input data consist of 5-min resolution solar wind parameters (IMF components, solar wind speed, density, dynamic pressure, and electric field) from OMNIWeb, averaged over 30-minute intervals, with a 12-hour lookback window. Our proposed model adopts a multi-layer perceptron-based encoder-decoder architecture following the Time-series Dense Encoder (TiDE) framework. Our results are as follows. First, our model achieves comparable or better performance in terms of RMSE and correlation coefficient relative to previous Dst and SYM-H prediction models. Second, our model successfully predicts the minimum SYM-H value and its timing within the 12-hour forecast window. Third, we demonstrate that the model produces reliable SYM-H forecast profiles during storm periods or even beyond the strict southward Bz criterion used for training. To our knowledge, this is the first study to forecast a geomagnetic index under strong IMF southward conditions, which is under an explicit condition on the state of the primary storm driver, and we expect that this approach can serve as a useful tool for space weather forecasting.