<![CDATA[Geomagnetic K-index characterizes local disturbance intensity on a quasi-logarithmic scale and serves as a critical input for space weather early warning systems. Predicting this index is challenging due to its non-stationary, chaotic nature and severely imbalanced class distribution. This study develops and compares two univariate hybrid deep learning models, VMD-CNN 1D and VMD-LSTM, for K-index prediction using data from BMKG Tuntungan Observatory from January 2020 to June 2025 (16,064 samples, 3-hour resolution). The preprocessing pipeline applies Variational Mode Decomposition (VMD) with six intrinsic mode functions, followed by an eight-timestep sliding window. An ablation study confirms that VMD substantially contributes to predictive performance, reducing RMSE by 55.4% and 52.4% for CNN 1D and LSTM respectively compared to their non-VMD counterparts with comparable parameter capacity. Both models were evaluated using RMSE, MAE, and R², with statistical significance confirmed via Wilcoxon signed-rank tests. VMD-CNN 1D achieved superior overall performance (RMSE 0.3608, R² 0.8846) compared to VMD-LSTM (RMSE 0.3889, R² 0.8660), and both substantially outperformed a Persistence baseline (RMSE 0.9643). However, under storm conditions (K ≥ 5, p = 0.024), VMD-LSTM outperformed VMD-CNN 1D (R² 0.6632 versus 0.5430) and achieved higher storm-detection recall (0.891 versus 0.812). These findings indicate that architecture choice should reflect the target operational context, with VMD-CNN 1D suited for routine monitoring and VMD-LSTM for storm-period forecasting.]]>
Copyrights © 2026
