Robust Forecasting and Physical Interpretability of Geomagnetic Storms Using XGBoost and SHAP

Abstract Accurate forecasting of the disturbance storm time (Dst) index is critical for timely space weather alerts and for understanding solar wind–magnetosphere interactions. In this study, we develop a robust and interpretable Dst prediction model based on the XGBoost algorithm, optimized through Bayesian tuning and designed to handle missing data and class imbalance in continuous solar wind records. The resulting model predicts the Dst index 1 hr ahead with high accuracy (global mean absolute error (MAE) ≈ 2.21 nT, storm-time MAE ≈ 5.62 nT, R 2 ≈ 0.96), significantly outperforming a persistence baseline. Leveraging Shapley Additive Explanations, we demonstrate the physical consistency of the model, showing how feature contributions dynamically shift across 34 storms of varying intensities. For moderate events (minimum Dst > –121 nT), forecasts are driven primarily by autoregressive Dst memory, supplemented by solar wind coupling indicators, southward interplanetary magnetic field trends, and solar wind velocity fluctuations. Conversely, for extreme storms (minimum Dst ≤ –121 nT), cumulative energy injection and prolonged southward interplanetary magnetic field conditions become markedly more influential, reflecting intensified magnetospheric energy injection during severe geomagnetic events. These findings highlight the dynamic redistribution of driver relevance with storm intensity and demonstrate the model’s capacity to capture physically meaningful relationships underlying geomagnetic variability.