Aurora prediction models
Pro Research Analysisby 
Searched over 200M research papers
Machine Learning and Neural Network Aurora Prediction Models
Recent advances in aurora prediction have leveraged machine learning and neural networks to improve forecast accuracy. Models using algorithms such as K-Nearest Neighbors (KNN), Random Forest (RF), and especially XGBoost have shown strong performance in predicting the location of the auroral oval based on geomagnetic indices like Kp, magnetic local time, and magnetic latitude. The XGBoost model, in particular, outperformed traditional empirical models, especially during geomagnetically disturbed conditions (Kp = 5–6), and demonstrated a linear relationship between increasing Kp and the equatorward movement of the auroral oval boundary. However, these models are more effective for nightside auroras and may require additional parameters for improved dayside predictions .
Neural network-based models, including generalized regression neural networks (GRNN) and conditional generative adversarial networks (CGAN), have also been developed using ultraviolet imager data. These models use solar wind parameters, interplanetary magnetic field, and geomagnetic indices as inputs to predict the spatial distribution of auroral intensity. Both GRNN and CGAN models achieved good similarity with observed auroral images, and their performance improved with larger training datasets .
Empirical and Statistical Aurora Forecast Models
Empirical models like OVATION-Prime and its updated versions (e.g., OVATION-Prime 2013) remain widely used for operational aurora forecasting. These models use statistical relationships between solar wind parameters, interplanetary magnetic field, and auroral particle fluxes to predict the probability and intensity of auroras. The OVATION-Prime 2013 model, for example, incorporates optimized solar wind-magnetosphere coupling functions and accounts for seasonal variations and different aurora types. It performs well in predicting the equatorward extent of the auroral oval, especially as geomagnetic activity increases, but is less reliable for the poleward boundary and during high geomagnetic activity (Kp ≥ 3) 7810.
Interactive computer models based on OVATION-Prime have been implemented as web-based systems, providing short-term (30–70 minutes) forecasts and visualizations of aurora probability. These systems have demonstrated high agreement (86%) between predicted and observed auroras, supporting both real-time forecasting and retrospective analysis .
Feature Tracking and Image-Based Aurora Models
Feature tracking models, such as the Feature Tracking of Aurora (FTA) model, use satellite ultraviolet imager data to track the boundaries and energy flux of auroral precipitation across magnetic local time sectors. These models capture the widening and narrowing of the auroral oval in response to geomagnetic activity and have shown better agreement with satellite measurements during high activity levels compared to other models .
Image-based deep learning models, such as those using Convolutional Long Short-Term Memory (ConvLSTM) networks, have been developed to predict the evolution of aurora during substorm expansion phases. These models can estimate both the intensity and movement of auroral boundaries with reasonable accuracy, as validated by similarity metrics between predicted and observed images .
Evaluation and Comparison of Aurora Prediction Models
Comparative studies have evaluated the performance of various aurora prediction models against satellite and ground-based observations. Models like OVATION-Prime have been shown to provide accurate forecasts for the general public, with correct predictions of visible aurora occurrence 77–86% of the time for Kp ≥ 3 . However, their performance decreases in dayside sectors and during periods of high geomagnetic activity, where they tend to underpredict or overpredict auroral occurrence probabilities 710.
Electron transport models used to predict auroral emissions from precipitating electrons have also been compared. Updated models with improved chemical and radiative rates show better agreement with in situ electron measurements and ground-based imaging, confirming their utility for deriving electron and ionospheric parameters from optical data .
Conclusion
Aurora prediction models have evolved from empirical and statistical approaches to advanced machine learning and deep learning techniques. While traditional models like OVATION-Prime remain operational standards, machine learning and neural network models are enhancing forecast accuracy, especially for complex and rapidly changing auroral dynamics. Ongoing improvements in data quality, model inputs, and evaluation methods continue to advance the reliability and utility of aurora prediction for both scientific and public applications 1234+6 MORE.
Sources and full results
Most relevant research papers on this topic