Comparative Benchmarking Study of Deep Learning Models for Geomagnetic Storms Prediction

Abstract

Geomagnetic storms have been known to have a substantial impact on the performance of satellites, communication systems, navigation systems, and other space weather-based systems. This research is concerned with developing a data-driven forecast of geomagnetic activity using deep learning and traditional machine learning models. This study focuses on comparing deep learning and machine learning models for predicting the Disturbance Storm Time (Dst) index through solar wind and geomagnetic parameters. For the analysis, an hourly dataset for 2000-2024 has been utilized resulting 219,168 observations. From these, eight parameters including magnetic field components, solar wind speed, geomagnetic indices, and past Dst values have been selected. Additionally, Random Forest, XGBoost, LightGBM, and persistence were employed as traditional baseline models. The performance measures considered included RMSE, MAE, R², and Pearson correlation, alongside training time, inference time, number of parameters, and a feature ablation study. Among the evaluated deep-learning architectures, LSTM achieved the best average performance across the five independent seeds. The RMSE of LightGBM and persistence baselines was 3.77459 nT and 4.63338 nT, respectively. The ablation analysis shows that AE, Kp10, and solar-wind variables substantially influence forecasting performance, whereas historical Dst and a longer 48-hour window do not necessarily improve the reported error. Overall, the results have shown that it is possible to model nonlinear temporal dependencies in geomagnetic observations using recurrent architectures. However, the use of traditional baselines along with repeated-seeds and ablation tests makes the evaluation more thorough.