AI-Based Space Weather Prediction for Satellite Protection

Abstract. Satellite operations and other space assets are gravely impaired by space weather disturbances, but existing forecasting systems often lack accuracy and integrated multi-objective prediction functionality. This paper suggests a unified AI-based multi-objective system for space weather prediction to enhance protection of satellites from solar-terrestrial disturbances. Five prediction tasks are addressed by the architecture: 1. solar active region classification; 2. solar flare prediction; 3. prediction of coronal mass ejection travel time; 4. Kp-index-based prediction of geomagnetic storms; and 5. satellite danger level classification. Multi-source inputs such as video sequences, heliophysical observations, magnetograms, and solar images were applied in developing and evaluating expert machine learning and deep learning models. The CNN+LSTM model for predicting flares had 0.67 accuracy and 0.47 F1-score, with good recall for quiet-class events and poor recall for flare events because of class imbalance. The system achieved 0.97 accuracy and 0.96 F1-score for active region classification. LightGBM performed better than XGBoost for CME trip time, with an R² of 0.86 and RMSE of 1.52 hours. XGBoost outperformed LightGBM in the prediction of Kp index (R² = 0.82, RMSE = 0.63), while LightGBM showed lower performance (R² = 0.67). The XGBoost classifier delivered strong multi-class performance with 0.999 accuracy and 0.87 F1-score for satellite risk level classification. The models demonstrated stability and robustness through satisfactory generalization over Solar Cycles 23 and 24. The five models are integrated for real-time application with a Gradio-based interface. Focus will be on enhancing flare detection and addressing class imbalance.