Accurate Ionospheric TEC Prediction With a Causal Attention Network at Northern EIA Crests

As the Sun approaches the peak of its 25th activity cycle, precise ionospheric forecasting has become increasingly challenging. Low-latitude regions have emerged as a persistent bottleneck for space weather operations. Extensive evidence reveals that existing AI models exhibit significant performance degradation in these regions, demonstrating remarkably higher error than mid-latitude predictions during solar maxima due to their inability to capture electrojet-driven spatiotemporal dependencies near equatorial ionization anomaly (EIA) crests. To address this limitation, we propose a region-specialized causal attention network (CAN) architecture tailored for low-latitude total electron content prediction. CAN integrates learnable joint embeddings that dynamically capture nonlinear couplings between spatial configurations and ionospheric gradient structures; Dual-stage attention enforces strict chronological dependence for temporal dynamics while enabling global feature interaction for spatial correlations. Then, a terminal state regression optimizing operational forecasting efficiency. Validated during the 2024 solar maximum at EIA crest longitudes (65°W and 115°E), CAN reduces RMSE by 24%–69% versus state-of-the-art models to about 2 TECU, maintains mean absolute percentage error below 9.5% during severe geomagnetic storms, and achieves R² > 95% for 7-day forecasts. These results signify that localized challenges necessitate localized solutions: universal models intrinsically fail to resolve equatorial electrodynamics, whereas region-optimized architectures such as CAN establish a new paradigm for physically constrained AI in space weather forecasting. By demonstrating that “local models for local problems” are essential, this work provides insights for enhancing global navigation satellite systems and augmentation system performance during extreme space weather events.