Ionospheric TEC Forecasting (CNN-DDL)
Zero-shot cross-solar-cycle generalization using Deep Delta Learning and Conformal Prediction.
Theoretical Formulation
Accurate forecasting of ionospheric Total Electron Content (TEC) during severe geomagnetic storms is critical for maintaining GNSS integrity. Standard sequential models suffer from chronological overfitting, memorizing epoch-specific plasma backgrounds.
To solve this, we architected the Convolutional Neural Network with Deep Delta Learning (CNN-DDL). Instead of predicting absolute TEC, the model learns the perturbation delta over a physical persistence baseline:
\[\hat{T}_{t+1} = T_{t} + \hat{\Delta}_{t}\]The Dynamic Beta-Gate
We replaced standard additive Transformer residuals with an adaptive rank-1 update. A dynamic gate ($\beta$) conditions the magnitude of the predicted correction on real-time solar wind variables:
\[\beta = \sigma(W_{\beta}^{(2)}ReLU(W_{\beta}^{(1)}x))\]This allows the network to amplify corrections during active conditions ($K_p \ge 5$) while suppressing noise during quiet periods.
Zero-Shot Cross-Cycle Evaluation
Trained exclusively on the extreme May 2024 superstorm (Solar Cycle 25), the model was evaluated zero-shot on the March 2015 and September 2017 storms (Solar Cycle 24).
- Performance: Achieved SOTA storm-time RMSE of 2.30 TECU (May 2024), outperforming BiLSTM, GRU, and SpatioTemporal ConvLSTM.
- Uncertainty Quantification: Integrated Marginal Split Conformal Prediction, providing a mathematically guaranteed 90% empirical coverage bound.