Based on the strong spatiotemporal characteristics of zenith tropospheric delays(ZTD),a multi-site ZTD combination prediction model with an improved attention mechanism based on convolutional neural networks(CNN-ATT)is proposed.The model integrates multiple data sources,including daily estimation accuracy,day of the year(DOY),and three-dimensional coor-dinates,for the first time in ZTD prediction tasks.A study is conducted using observation data from 5 reference stations(CORS)in Nanning and 14 International GNSS Service(IGS)stations,spanning a total of 1 501 DOY.Traditional back propagation(BP)models,global pressure and temperature 2wet(GPT2w)models,and ATT models are selected as baseline models for compar-ative analysis.The prediction results demonstrate that in terms of prediction accuracy,the im-proved CNN-ATT model outperforms traditional BP neural network models,with a reduction in mean squared error(MSE)and mean absolute error(MAE)by 5.5 mm and 4.4 mm respectively,leading to an improvement in prediction accuracy by 41.4％and 67.8％.Compared to the ATT model,the improved CNN-ATT model also shows reductions in MSE and MAE by 4.6 mm and 2.1 mm,respectively,resulting in a 36.2％and 50.0％enhancement in prediction accuracy.Re-garding positional accuracy,the improved CNN-ATT model outperforms the SAAS,GPT2w,BP,and ATT model.Furthermore,when compared to the traditional SAAS tropospheric model,the CNN-ATT model achieves noteworthy accuracy improvements in the N,E and U directions,with enhancements of 18.2％,12.6％and 31.0％respectively.Additionally,the research unveils that the CNN-ATT model exhibits a more stable performance in extended prediction time steps,making it particularly suitable for multi-station prediction tasks.Moreover,it manifests a faster convergence rate in precise point positioning(PPP)applications.
维普
万方数据