A multi-scale prediction model(referred to as C-TCN-A) based on complete ensemble empirical mode decomposition with adaptive noise(CEEMDAN) and temporal convolutional network-attention mechanism(TCN-Attention) algorithms is proposed, which can be effectively applied to missing data imputation and future trend prediction of GNSS elevation time series. The model first employs CEEMDAN for multi-scale decomposition of time series, then utilizes TCN-Attention for prediction and reconstruction of different scale components to obtain final results. To validate the model's performance, 12 observation stations were selected for 1-day and 5-day predictions, with comparisons made against multiple other models. Results demonstrate that in 1-day prediction, C-TCN-A reduces RMSE and MAE by 35%-40% and 36%-41% respectively while improving correlation coefficient R by 25%-29%. For 5-day prediction, it achieves reductions of 20%-26% in RMSE and 20%-28% in MAE, with R increasing by 26%-33%. To verify the model's universality, C-TCN-A was applied to 99 stations from the crustal movement observation network of China(CMONOC) for 1-day and 5-day predictions. Results indicate generally favorable RMSE and MAE metrics with concentrated error distribution, where most errors remain below 4 mm. Spatial analysis reveals regional performance differences, with optimal results achieved in the northwestern region.