Speaker
Description
Solar flares pose risks to infrastructure both on Earth and in space, from induced currents in power grids to satellite damages. Most operational flare forecasting models treat the problem as a binary classification task (flare vs. non-flare) based on a fixed prediction horizon, e.g., 24 hours.
We recast solar-flare forecasting as a continuous time-to-event problem, without the necessity of a fixed prediction boundary, and bring survival-analysis techniques – widely adopted in medicine and reliability engineering – into the realm of solar flare prediction. Using time series of photospheric vector-magnetogram features, we compare several survival models, including Cox proportional- and non-proportional-hazard networks, random survival forests, neural additive models, and universal function approximators derived from the Kolmogorov-Arnold representation theorem.
Our calculations show that these models not only predict when a flare is likely to occur, but also yield time-dependent feature importances, revealing the contributions of individual magnetic parameters to imminent flare risks. We analyze the advantages and disadvantages of each model to find the best suited method in the application on solar flare forecasting.
By providing time-continuous risk estimates and interpretable drivers of flare imminence, survival-analysis frameworks continue to provide a promising avenue towards real-time, probabilistic space-weather forecasting.
| Do you plan to attend in-person or online? | In-person |
|---|
