A Dual-Channel XGBoost–Transformer–LSTM Fusion Method for GNSS Ionospheric Scintillation Forecasting

Low-latitude ionospheric irregularities cause rapid fluctuations in GNSS signal amplitude and phase, threatening receiver tracking and services. We propose a dual-channel XGBoost–Transformer–LSTM fusion model to address rare events, missed-detection costs, and the difficulty of combining physics-informed and profile-structure information. Data from the HKOH station and geospace products spanning 1 January 2020–21 March 2024 were used. Space-weather variables, temporal factors, previous-day scintillation status, and a TEC120 latitude–UT grid at 06:00–11:00 UT were used to predict GPS L1 amplitude scintillation during 11:00–14:00 UT. The XGBoost channel used the raw TEC grid, first-order latitudinal and hourly TEC differences, and background variables, whereas the Transformer–LSTM channel extracted within-day structure. Channel probabilities were combined by soft voting, with the weight and threshold selected on the validation set. Chronological test results were recall = 0.8788, F1-score = 0.5949, ROC-AUC = 0.8263, PR-AUC = 0.6097, CSI = 0.4234, TSS = 0.4511, and MCC = 0.4096. Fusion raised recall from 0.8030 to 0.8788 but reduced precision from 0.4609 to 0.4496. Paired bootstrap analysis supported a stable improvement only in recall. The model therefore suits warning applications prioritizing detection over minimizing false alarms.