Ionospheric TEC Forecasting
CNN-DDL architecture for real-time solar wind drivers and extreme geomagnetic superstorms.
Project Overview
Co-authored a CNN-DDL architecture featuring a dynamic gate that conditions real-time solar wind drivers to scale non-linear corrections over physical persistence baselines.
Key Achievements
- Zero-Shot Testing: Evaluated via zero-shot cross-solar-cycle testing (SC25 to SC24), achieving SOTA storm-time error rates (2.30 RMSE) and outperforming deep sequential baselines and IRI-2020.
- Uncertainty Quantification: Integrated Marginal Split Conformal Prediction to establish distribution-free uncertainty quantification, providing a mathematically reliable 90% empirical coverage bound during extreme geomagnetic superstorms.
Technologies Used: PyTorch, Deep Delta Learning, Conformal Prediction, CNNs