Is this project an undergraduate, graduate, or faculty project?
Undergraduate
Project Type
individual
Campus
Daytona Beach
Authors' Class Standing
John Rendleman, Freshman
Lead Presenter's Name
John Rendleman
Lead Presenter's College
DB College of Engineering
Faculty Mentor Name
Claudia Ehringer Lucas
Abstract
STORMTRACK: A Regime-Aware Classifier-Router Architecture for Multi-Horizon Kp Index Forecasting Current algorithms in operational space weather face extreme difficultly predicting the Kp geomagnetic index beyond 24 hours, a lead time that is critical for protecting high-frequency communications and infrastructure. Most regression models are optimized for quiet conditions, which dominate the data, leading to systematic underpredictions of storm events that cripple space infrastructure. Probabilistic approaches and physics-based numerical models also falter due to the same class imbalance plaguing standard regressors at multi-day lead times. The ICARUS 6 architecture addresses this by splitting the forecasting component into quiet and storm regimes, which was discovered in testing that a single model cannot minimize error on the quiet majority and still classify rare storm events. Twelve LightGBM classifiers estimate storm events (P(Kp ≥ 4, 5, 6)) over four forecast windows up to 72 hours. A sigmoid blending function routes between quiet or storm regressors based on these probabilities. The quiet pathway uses gradient-boosted ensembles with Huber loss on the entire dataset, while the storm pathway uses an identical architecture on only storm-context windows with an asymmetric loss that penalizes underprediction. The IMF Bz autocorrelation drops below 0.5 within 12 hours (Lockwood et al., 2019), so both pathways must transition from raw temporal features to summary statistics beyond 24 hours. The ICARUS 6.2.1a model trains on 18 OMNI solar wind parameters from 1995 to 2025, expanded into over 500 engineered features, including coupling functions, rolling statistics, Bz persistence, and 27-day recurrence features. Gaps are linearly interpolated up to 3 hours, and the data follows an 85/15 split with a one-week temporal buffer. The model achieves a mean absolute error of 0.94 and a root mean square error of 1.20 across 3–72-hour horizons, demonstrating that the classifier-routed decomposition can reduce error for class imbalance challenges.
Did this research project receive funding support (Spark, SURF, Research Abroad, Student Internal Grants, Collaborative, Climbing, or Ignite Grants) from the Office of Undergraduate Research?
No
Included in
Artificial Intelligence and Robotics Commons, Atmospheric Sciences Commons, Data Science Commons
STORMTRACK: A Regime-Aware Classifier-Router Architecture for Multi-Horizon Kp Index Forecasting
STORMTRACK: A Regime-Aware Classifier-Router Architecture for Multi-Horizon Kp Index Forecasting Current algorithms in operational space weather face extreme difficultly predicting the Kp geomagnetic index beyond 24 hours, a lead time that is critical for protecting high-frequency communications and infrastructure. Most regression models are optimized for quiet conditions, which dominate the data, leading to systematic underpredictions of storm events that cripple space infrastructure. Probabilistic approaches and physics-based numerical models also falter due to the same class imbalance plaguing standard regressors at multi-day lead times. The ICARUS 6 architecture addresses this by splitting the forecasting component into quiet and storm regimes, which was discovered in testing that a single model cannot minimize error on the quiet majority and still classify rare storm events. Twelve LightGBM classifiers estimate storm events (P(Kp ≥ 4, 5, 6)) over four forecast windows up to 72 hours. A sigmoid blending function routes between quiet or storm regressors based on these probabilities. The quiet pathway uses gradient-boosted ensembles with Huber loss on the entire dataset, while the storm pathway uses an identical architecture on only storm-context windows with an asymmetric loss that penalizes underprediction. The IMF Bz autocorrelation drops below 0.5 within 12 hours (Lockwood et al., 2019), so both pathways must transition from raw temporal features to summary statistics beyond 24 hours. The ICARUS 6.2.1a model trains on 18 OMNI solar wind parameters from 1995 to 2025, expanded into over 500 engineered features, including coupling functions, rolling statistics, Bz persistence, and 27-day recurrence features. Gaps are linearly interpolated up to 3 hours, and the data follows an 85/15 split with a one-week temporal buffer. The model achieves a mean absolute error of 0.94 and a root mean square error of 1.20 across 3–72-hour horizons, demonstrating that the classifier-routed decomposition can reduce error for class imbalance challenges.