| |
 |
UNIVERSITY OF BUCHAREST
FACULTY OF PHYSICS Guest
2026-06-11 23:58 |
 |
|
|
|
Conference: Bucharest University Faculty of Physics 2026 Meeting
Section: Atomic and Molecular Physics. Astrophysics. Applications. Optics, Spectroscopy, Plasma and Lasers
Title:
Operational Dst Forecasting Across Solar Cycle Using Physics-Motivated Gradient Boosting
Authors:
Ionut-Catalin SANDU (1), Diana BESLIU-IONESCU (1)
Affiliation:
1) Astronomical Institute of the Romanian Academy
E-mail
supportcatalin.sandu@aira.astro.ro
Keywords:
dst, forecasting, space-weather
Abstract:
A physics-motivated machine learning framework is developed for forecasting the hourly Dst index at lead times of 4 to 12 hours. An expanded feature space (e.g. convolutional filters, rolling statistics, energy accumulation) is constructed from solar wind and interplanetary magnetic field measurements, going beyond usual lagged values. An XGBoost gradient boosting model is trained on solar cycles (SC) 20--24 and evaluated on the entirely independent SC 25 (August 2017 -- November 2025, including the minimum phase between SC24 and SC25), comprising 73,056 hourly values. At a forecast horizon of 4 hours, the model achieves MAE $= 5.52$ nT, RMSE $= 8.36$ nT, and $R^2 = 0.79$. A threshold-based classification evaluation is also introduced as a complement to standard regression metrics, directly assessing storm detection performance at operationally relevant Dst thresholds. F1-scores of 0.75, 0.71, and 0.77 are obtained for moderate ($leqslant -50$ nT), intense ($leqslant -100$ nT), and severe ($leqslant -150$ nT) storm thresholds respectively. Interpretability analysis through SHAP values and Individual Conditional Expectation (ICE) plots reveal a context-dependent shift in dominant predictors: recent Dst history governs predictions under quiet conditions, while the solar wind electric field and convolution-derived IMF coherence features become dominant during active storm periods. The trained model is able to switch between relevant features, depending on the forecast horizon. This regime-dependent behavior is consistent with theoretical work on sustained solar wind energy transfer and emerges without explicit storm-phase conditioning, indicating that physics-motivated feature engineering guides the model towards physically meaningful predictions.
|
|
|
|