Prediction of Geomagnetic Storms through Artificial Intelligence: A Review, Methodological Framework, and Simulated Case Study

Geomagnetic storms, driven primarily by coronal mass ejections (CMEs) and high-speed solar wind streams interacting with the Earth's magnetosphere, pose significant risks to satellite operations, power grid infrastructure, aviation, and communication systems. Traditional empirical and physics-based models of geomagnetic indices such as Dst and Kp, while physically interpretable, often struggle to capture the highly nonlinear and non-stationary dynamics of solar wind-magnetosphere coupling, particularly during the storm recovery phase. Artificial Intelligence (AI) techniques, including artificial neural networks, ensemble tree-based methods, and multilayer perceptron (MLP) models, have demonstrated considerable promise for improving the accuracy and lead time of geomagnetic storm forecasts. This paper (i) reviews the evolution of AI-based approaches to Dst/Kp prediction; (ii) proposes a consolidated methodological framework for building and validating AI-driven space-weather forecasting systems; and (iii) reports a complete, reproducible experimental case study in which a physically-grounded Burton-McPherron-Russell (BMR) ring-current simulator was used to generate a multi-year hourly solar-wind-driven Dst dataset, on which Random Forest, Gradient Boosting, Multilayer Perceptron, and linear baseline models were trained and evaluated under a chronologically leakage-aware train/validation/test protocol at 1-, 3-, and 6-hour forecast horizons. The best-performing model (MLP) achieved RMSE = 3.82 nT and R² = 0.956 at the 1-hour horizon, degrading to RMSE = 10.27 nT and R² = 0.684 at 6 hours, consistent with the general skill-decay pattern reported in the literature for physically driven Dst forecasting. Feature importance analysis confirmed that the most recent lagged Dst value and the 6-hour rolling mean of the southward IMF component (Bz) dominate predictive skill. Key challenges relating to data leakage, class imbalance for extreme events, multimodal data fusion, and model interpretability are discussed, along with directions for future research including physics-informed machine learning and probabilistic forecasting. Keywords: artificial intelligence; machine learning; neural networks; Random Forest; geomagnetic storms; Dst index; Burton-McPherron-Russell model; space weather