An HHT–ANN Framework for Short‐Term Kp Forecasting

Abstract

The geomagnetic activity index Kp is an important indicator of solar wind–magnetosphere coupling, and accurate 3‐hr‐ahead forecasting is important for space‐weather monitoring and warning. Because upstream solar wind and interplanetary magnetic field (IMF) signals are strongly nonlinear and nonstationary, methods based only on conventional statistical descriptors or simple regression may not adequately characterize their local temporal variability. To address this issue, we propose an HHT–ANN forecasting framework that combines Hilbert–Huang Transform (HHT)‐based feature extraction with an artificial neural network (ANN). Using upstream solar wind and IMF observations from the ACE and DSCOVR spacecraft at L1 together with Kp data released by GFZ, we construct supervised samples on 3‐hr Kp intervals and fuse statistical features with HHT‐derived adaptive time–frequency features for model prediction. On the test set, the proposed model achieves and RMSE = 0.613, indicating strong continuous forecast skill. For event‐based evaluation with an event threshold of , the model yields ROC–AUC = 0.956 and PR–AUC = 0.779, showing useful discrimination of disturbed intervals. Validation for representative years spanning different solar‐cycle phases, together with two geomagnetic storm case studies in November 2023 and May 2024, further confirms that the model captures rapid Kp variations during enhanced activity. Compared with several representative published studies, the proposed framework achieves competitive or improved performance while maintaining a relatively simple and physically interpretable structure.