Using Transformers to Forecast Incredibly Rare Solar Flares

2026-05-11 · Source: Towards Data Science · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Space Science & Astronomy · Depth: Advanced, long

Summary

Machine learning forecasting for rare events, such as solar flares, requires a fundamental shift in modeling approach, moving beyond standard metrics like accuracy. The 2003 Halloween storms, an X-45 class solar flare 450 times stronger than an M-1 flare, exemplify the impact of such events, causing satellite malfunctions, GPS disruptions, and power outages like the Malmö Blackout. Predicting these rare, high-impact events is challenging due to their infrequency, necessitating specialized techniques. Data for solar flare prediction is collected from the Sun's photosphere by NASA's Solar Dynamics Observatory (SDO) using the Helioseismic and Magnetic Imager (HMI), despite flares occurring in the corona and chromosphere. This data undergoes localization and feature engineering to derive magnetic parameters like flux, current, twist, and helicity, which describe solar and magnetic structures. The model targets the probability of an M-1 class event within 24 hours, using the True Skill Statistic (TSS) as a more appropriate metric than accuracy for imbalanced datasets. A "tail model" is developed to focus on extreme events, using the Generalized Pareto Distribution (GPD) to model exceedances beyond a defined threshold. This tail model is then combined with a full distribution model using a transformer with multiple output heads to predict both flare occurrence and severity.

Key takeaway

For Machine Learning Engineers developing models for rare, high-impact events like solar flares, relying solely on accuracy is insufficient and misleading. You should prioritize metrics like the True Skill Statistic (TSS) and implement a "tail model" using distributions such as the Generalized Pareto Distribution (GPD) to specifically capture extreme event behavior. This approach, integrated with a multi-head transformer, will improve both prediction accuracy for rare occurrences and the estimation of event severity, leading to more robust and actionable forecasts.

Key insights

Predicting rare, high-impact events requires specialized ML techniques, including tail modeling and appropriate metrics.

Principles

• Rare event forecasting demands tailored metrics.
• Focus data collection on likely event locations.
• Combine models for full and tail distributions.

Method

The proposed method involves feature engineering from magnetic data, defining a precise target (M-1 class event probability), using TSS for evaluation, building a tail model with GPD for exceedances, and integrating these via a transformer with multiple heads.

In practice

• Use True Skill Statistic (TSS) for rare event classification.
• Engineer features from magnetic field data for solar flares.
• Employ Generalized Pareto Distribution for tail modeling.

Topics

• Rare Event Forecasting
• Solar Flare Prediction
• Transformer Models
• True Skill Statistic
• Generalized Pareto Distribution

Best for: AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• Neural Negative Binomial Regression for Weekly Seismicity Forecasting: Per-Cell Dispersion Estimation and Tail Risk Assessment — Seismic count data's overdispersion requires per-cell Negative Binomial modeling for accurate extreme event risk assessment.
• 5 Scipy.stats Tricks for Simulating ‘What If’ Scenarios — scipy.stats offers powerful, vectorized tools for robust probabilistic "what-if" scenario modeling, moving beyond static point estimates.
• Step-adaptive multimodal fusion network with multi-scale cloud feature learning for ultra-short-term solar irradiance forecasting — A multi-source fusion model with adaptive compensation improves ultra-short-term solar irradiance forecasting accuracy.
• Customer Churn Prediction on Structured Data Using FT-Transformer and Stacking Ensembles — Hybrid models combining transformers and tree ensembles offer superior, calibrated churn prediction on tabular data.
• Why a 12-year-old forecasting paper has stood the test of time — Effective forecasting of complex events relies on robust evaluation and intelligent fusion of diverse data sources.
• GPU-Accelerated Deep Learning for Heatwave Prediction and Urban Heat Risk Assessment — GPU-accelerated ConvLSTM models predict next-day urban thermal conditions and heat risk by integrating satellite and meteorological data.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Towards Data Science.