• Proposes ITP-ST, a two-phase model for incomplete time-series traffic prediction in LEO satellite networks, integrating missing data imputation and traffic forecasting.
• Introduces IDAE-MDI, an improved denoising autoencoder that leverages Gramian angular summation field and spatio-temporal correlation for accurate missing data imputation.
• Develops TP-CACNN, a multi-channel attention convolutional neural network that effectively combines spatio-temporal traffic correlations for prediction.
• Demonstrates through experiments that ITP-ST consistently outperforms baseline models in traffic prediction accuracy across various data missing rates.