Accurate forecasting of high-latitude ionospheric total electron content (TEC) is essential for mitigating space-weather impacts on Global Navigation Satellite System (GNSS) performance. This study benchmarks three machine-learning models—random forest (RF), support vector machine, and long short-term memory networks—against a three-point moving average MA(3), the International Reference Ionosphere (IRI-2020), and a Persistence baseline. Using GNSS-derived TEC from four high-latitude stations during the geomagnetically active year 2023, we evaluate model skill using root mean squared error (RMSE), MAE, R2, and error distributions. The results reveal a clear, physically interpretable regime dependence. In the auroral zone, RF reduces RMSE by up to 30% relative to persistence, demonstrating superior skill under storm-driven, intermittent variability. In the polar cap, MA(3) outperforms all machine-learning approaches, achieving R2=0.86–0.87 and reducing RMSE by ∼ 26%, reflecting the smooth, strongly autocorrelated nature of convection-driven TEC. Persistence provides the weakest forecasts but remains an essential lower benchmark. Feature relevance analysis shows that short-term lagged TEC and IRI climatology dominate prediction skill, with geomagnetic drivers exerting stronger influence in auroral regions. By linking forecast performance to plasma regime, this work establishes a physically grounded framework for TEC predictability and clarifies why nonlinear ensemble methods excel in auroral conditions while smoothing-based approaches dominate in the polar cap.
Uga et al. (Thu,) studied this question.