A Prediction Method for LEO Satellite Channels Under Channel Aging

To address channel state information (CSI) aging in low-Earth-orbit (LEO) links caused by propagation delay and platform/user mobility, we study CSI prediction with carrier-to-noise ratio (C/N0) as the target. Building on the 3GPP Non-Terrestrial Network - Tapped Delay Line C (NTN-TDL-C) channel model, we construct a multi-UE LEO scenario and derive features including effective isotropic radiated power (EIRP), free-space path loss (FSPL), received isotropic power (RIP), link margin, elevation angle, energy-per-bit to noise power spectral density ratio (Eb/N0), and geographic coordinates. On this basis, a hybrid neural network is subsequently designed, combining a Graph Convolutional Network (GCN) with a Gated Recurrent Unit (GRU): a GraphSAGE-based GCN encodes cross-UE dependencies in a star topology within each time window, and a GRU models the temporal evolution of the target UE. In contrast to timeseries predictors that use only a single UE's history or methods that reconstruct only the current CSIT, the proposed model leverages concurrent observations from neighboring UEs to forecast the target UE's subsequent downlink quality. On a MATLAB NTN-TDL-C simulation platform with 10 UEs-covering dense-urban and mountainous settings and multiple orbital altitudes-the approach reduces Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), and Sum of Squared Errors (SSE) by 32.9%, 31.9%, 19.8%, and 44.3%, respectively, relative to GRU-single, GRU-multi, GCN-single, GCN-multi, Long Short-Term Memory (LSTM) baselines and Transformer-Single baselines; across 2-5 ms prediction horizons, Normalized Mean Squared Error (NMSE) improves by about 6 dB. The framework provides a multi-UE-aware spatiotemporal predictor that maintains accuracy and stability across environments and horizons, and the resulting channel forecasts can support proactive scheduling and multibeam beamforming. Data and code are available for reproduction.