A CNN-iTransformer Spatiotemporal Fusion Model for Long-Term eLoran Propagation Prediction

The accurate prediction of propagation delay is imperative for enhancing the timing precision of eLoran systems. In this letter, a hybrid deep learning framework that integrates spatially encoded environmental features and temporal modelling to predict eLoran propagation delay was proposed. The model employs a one-dimensional convolutional neural network (1DCNN) to intelligently encode and compress high-dimensional environmental data (meteorological and soil parameters) along the signal path into low-dimensional feature representations. Then, an iTransformer module is utilized to capture long-term temporal dependencies from encoded features and historical eLoran data. The model was evaluated using multi-year, fieldcollected dataset. The results show the superior performance of the model, achieving a Mean Absolute Error (MAE) of 8.0 ns and a Root Mean Square Error (RMSE) of 10.1 ns, significantly outperforming convolutional neural networks and the standard Transformer model. Furthermore, the model displays graceful performance degradation over prediction horizons of up to 720 hours, a trait not exhibited by non-temporal baseline models. These results validate the efficacy of our spatiotemporal fusion approach and its potential to advance resilient PNT systems.