Short-term forecast of ionospheric sporadic E intensity fusing physical parameters via deep learning

Ionospheric sporadic E (Es) layers are thin-layer irregularities with abnormaly sharp enhanced densities of electrons and metal ions. Es layers affect the global navigation satellite system (GNSS) navigation and positioning by causing scintillations or loss of signals, while the timely and accurate forecast of Es intensity is difficult due to its complex spatiotemporal characteristics and various influencing factors. We propose a novel convolutional tensor-train long short-term memory-based model fusing physical parameters (Conv-TT-LSTM-Phy) for short-term forecasting of global Es intensity. Conv-TT-LSTM-Phy is trained using Es intensity maps derived from Constellation Observing System for the Meteorology, Ionosphere, and Climate GNSS radio occultation data during 2007–2014, and data of the global distributed vertical ion convergence driven by the wind shear are for the first time introduced into the forecast model. With the ground truth of Es intensity maps as reference, Conv-TT-LSTM-Phy is validated to outperform the other two deep learning models without fusing physical parameters and an empirical model in predicting Es intensity maps, with the reductions in ME/MAE/RMSE reaching 99.5%/80.5%/71.3% compared with empirical model. As to the prediction accuracy, which generally degrades with the increased prediction step, Conv-TT-LSTM-Phy performs best when the prediction time is within 6 h. Conv-TT-LSTM-Phy also reconstructs the morphology of Es intensity accurately. The proposed model can provide valuable information for ionospheric irregularities monitoring and space weather forecasting. Moreover, using Conv-TT-LSTM-Phy as the constraint for PPP can effectively both improve positioning performance and reduce convergence time.