GAIE: Real-Time Geomagnetic Storm Prediction Engine Based on NASA Satellite Data

GAIE (Geomagnetic AI Engine) is a core component of the HELIOS Space Intelligence Platform. It uses the XGBoost model and SHAP interpretability technology, combined with NASA/NOAA real-time satellite data, to achieve 97% R² accuracy (KP index prediction) and 98% F1 classification score (G grade prediction). Its goal is to shift space weather monitoring from passive response to predictive defense, protecting critical infrastructure from geomagnetic storms.

Geomagnetic storms are caused by the interaction between solar wind and Earth's magnetosphere, posing a huge threat to modern society's critical infrastructure. Historical events include:

1. The 1989 Quebec blackout (6 million people lost power for 9 hours, with $2 billion in losses);
2. The 2003 Halloween storm (30 satellites damaged, global high-frequency communication interrupted);
3. The 1859 Carrington Event (if repeated, losses would range from $0.6 to $2.6 trillion). Affected fields: communication satellites, GPS, power grids, polar aviation communication, astronaut safety.

The HELIOS platform integrates NASA/NOAA satellite data and includes 5 modules: orbital launch schedule, solar event monitoring, satellite tracking, AI prediction (GAIE), and solar energy optimization. GAIE solves the core problem: predicting the intensity of geomagnetic disturbances (KP index, G grade) in the next few hours based on solar wind data from the DSCOVR satellite (L1 point). Data sources are all government public APIs: NOAA SWPC's solar wind magnetic field/plasma/KP index, and NASA DONKI's flare/storm events.

The dataset contains 11249 records (9749 real DSCOVR data + 1500 synthetic data to supplement extreme storm samples). Feature engineering designed 20 features with clear physical meanings, such as bz_negativo (southward Bz component, magnetic reconnection channel), newell_coupling (energy transfer rate), pressao_dinamica (magnetosphere compression pressure), etc. Preprocessing steps: deduplication, outlier handling, time alignment, stratified division, and robust standardization.

Regression task (KP index): XGBoost performed best (RMSE=0.2768, MAE=0.1704, R²=0.9678), outperforming Ridge (R²=0.82) and Random Forest. Classification task (G grade): XGBoost achieved an accuracy of 0.9787 and a weighted F1 score of 0.9772. Key insight: The relationship between solar wind and geomagnetic activity is nonlinear, and gradient boosting can better capture complex interactions.

SHAP analysis was done using TreeExplainer, and the feature importance ranking is:

1. bz_negativo (southward Bz, magnetic reconnection trigger);
2. newell_coupling (energy transfer efficiency);
3. CME (main cause of G3-G5 storms);
4. Wind speed;
5. Dynamic pressure. The results are consistent with plasma physics research, proving that the model captures real physical phenomena rather than statistical artifacts.

GAIE is deployed as a Streamlit application (link: https://globalsolutiongenerativeai-gkw5rmitemjc8d7ue7mvub.streamlit.app). Its functional modules include real-time prediction, SHAP explanation, model metrics, and project introduction. Simulated extreme storm: Bz=-30nT + wind speed 750km/s + CME + M-class flare → KP7-8, G3-G4 grade. Alignment with UN SDGs: SDG9 (Protect Infrastructure), SDG13 (Climate Action), SDG11 (Sustainable Cities).

GAIE combines machine learning and space physics knowledge to solve practical social value problems. The 97% R² and 98% F1 scores mean satellite operators, grid managers, etc., can get early warnings hours in advance and take protective measures to avoid billions of dollars in losses. Its technical highlights include end-to-end ML engineering, physics-informed ML, explainable AI, robust data strategy, and production-level deployment.