📖 HelioGuard Glossary

Area Under the ROC Curve measures how well the model distinguishes between flare and non-flare events. 0.5 = random, 1.0 = perfect. Our model achieves 0.912, indicating excellent discrimination.

True Skill Score = Recall + Specificity – 1. A score of 0 means no skill, 1 means perfect. Our model achieves 0.676, well above the accepted good threshold (0.5).

Brier Score measures the calibration of probabilities (mean squared error). 0 is perfect, 0.25 is uninformative. Our model scores 0.015, meaning probabilities are highly reliable.

Our 95% CI is computed via Bootstrap resampling (200 iterations). It gives the range within which the true probability likely lies. For example, if probability is 60% with CI [45%, 75%], there's 95% confidence the real value is between 45% and 75%.

Combines 4 models: Surya (6h), Jinwu (12h), HelioGuard (24h), MagPy (48h). Each has different strengths; the ensemble provides a more robust estimate.

Global geomagnetic activity scale from 0 to 9. Higher Kp means stronger geomagnetic storms and more aurora. HelioGuard uses Kp for aurora predictions and photography tips.

Space-weather HMI Active Region Patches — 18 magnetic field parameters from SDO/HMI, including magnetic flux, current helicity, shear angle, etc. We add time-evolution features (deltas) to capture flare build-up.

Tailored recommendations for Satellite Operators, Grid Managers, Drone Fleets, Radio Operators, Navigation Specialists, and Researchers. Each role gets specific, actionable steps based on the risk level.

All metrics are computed on an independent time-split test set (last 10% of data, 2010–2024). No data leakage. Performance is stable across different solar cycles.

Predictions are available for 6, 12, 24, and 48 hours ahead. Shorter windows (6h) are more uncertain but react faster; longer windows (48h) provide more stable estimates.

High Recall: catches almost all flares, but may have more false alarms.
Balanced: optimal trade-off between recall and precision.
High Precision: minimizes false alarms, but may miss some flares.

A sudden, intense burst of radiation from the Sun's surface. Flares are classified by X-ray flux: X (strongest), M, C, B, A (weakest). They can disrupt communications, GPS, and power grids.

A massive burst of plasma and magnetic field from the Sun's corona. CMEs can travel at speeds up to 3000 km/s and, when directed at Earth, cause geomagnetic storms and auroras.

A disturbance in Earth's magnetosphere caused by solar wind and CMEs. Strong storms (Kp ≥ 7) can induce currents in power grids, damage satellites, and produce spectacular auroras.

A stream of charged particles (plasma) released from the Sun's upper atmosphere. It travels at 300–800 km/s and interacts with Earth's magnetic field, causing auroras and space weather.

The region around Earth dominated by its magnetic field. It protects us from solar wind particles. Geomagnetic storms occur when solar wind compresses or disturbs the magnetosphere.

Dark, cooler areas on the Sun's surface caused by intense magnetic fields. Sunspots are often the source of solar flares and CMEs. The more sunspots, the more active the Sun.

An approximately 11-year cycle of solar activity, measured by sunspot numbers. During solar maximum, flares and CMEs are more frequent; during solar minimum, the Sun is quieter.

A region on the Sun where the corona is cooler and less dense, allowing solar wind to escape at high speeds. Coronal holes are a major source of fast solar wind and recurrent geomagnetic storms.

An area on the Sun where magnetic fields are concentrated, often associated with sunspots and solar flares. Active regions are numbered by NOAA and tracked for space weather forecasting.

Techniques like Platt Scaling and Isotonic Regression adjust raw model outputs to make probabilities more accurate. Our model's Brier Score of 0.015 indicates near-perfect calibration.

Specificity = TN / (TN + FP). It measures how well the model identifies non-flare events. High specificity means few false alarms. Our 24h model achieves 83.9%.

Precision = TP / (TP + FP). It measures how many of the predicted flare events actually occurred. High precision means alerts are trustworthy. Our model achieves 8.3% (low but expected for 0.1% event rate).

Recall = TP / (TP + FN). It measures how many actual flares were caught by the model. High recall means few missed flares. Our 24h model achieves 83.7%.

F1 Score is the harmonic mean of precision and recall. It balances both metrics. Our model achieves 0.15, which is low due to the extreme class imbalance, but TSS is a better metric for this problem.

A table that summarizes prediction results: True Positives (TP), True Negatives (TN), False Positives (FP), False Negatives (FN).

Solar Dynamics Observatory / Helioseismic and Magnetic Imager – NASA satellite that provides high-resolution magnetic field data of the Sun. Our SHARP parameters are derived from HMI observations.

HelioGuard's backend is built with Flask, using SQLite for database, XGBoost for predictions, and JavaScript for interactive frontend. The system fetches real-time Kp indices, cloud data, and processes user requests.

Probability calibration ensures that a predicted 70% chance actually occurs about 70% of the time. We use Platt Scaling and Isotonic Regression to calibrate our XGBoost outputs, achieving a Brier Score of 0.015.

A method that fits a logistic regression model to the raw outputs of a classifier to produce calibrated probabilities. Used in our 6h and 12h models.

A non-parametric calibration method that fits a monotonic increasing function to the raw probabilities. Used in our 24h and 48h models, yielding slightly better calibration.

A resampling technique (200 iterations) used to estimate the uncertainty of our predictions. The 95% CI is the range within which the true probability likely lies, with 95% confidence.

The probability value above which we issue a "High Risk" alert. Our default threshold is 0.3 (30%), but you can adjust it in Pro mode via "High Recall/Balanced/High Precision" options.

The Receiver Operating Characteristic curve plots true positive rate vs false positive rate. AUC (Area Under the Curve) summarizes the model's ability to distinguish between classes. Our AUC is 0.912, indicating excellent discrimination.

We compute 6h and 12h differences (deltas) of key parameters like USFLUX, TOTUSJH, SAVNCPP, MEANSHR, and SHRGT45 to capture the rate of change. These help predict sudden flares.