Super substorms are extreme disturbances in Earth’s magnetic field that can disrupt satellites, power systems, and communication networks. Predicting them is difficult because most auroral activity is weak, and these quiet times dominate the data. This imbalance causes traditional models to focus on small events and underestimate the rare, extreme ones.

We developed IRANNA (Imbalanced Regression Artificial Neural Network for the Auroral electrojet index): a model designed to overcome this challenge by giving more weight to large events during training. Using only solar wind measurements as input, IRANNA predicts the strength of auroral disturbances without relying on past index values.

Tests show that IRANNA can successfully predict the peak strength of many strong-to-extreme events, a first for this kind of modeling. Some challenges remain, such as handling localized events and uncertainties in solar wind data. Our results show that imbalanced regression techniques can greatly improve extreme space weather prediction and deepen our understanding of how the magnetosphere responds to solar activity.