SAD-CP: Spatially Anchored Dual-Distilled Continual Prediction for Dynamic CSI Tracking

Abstract

Accurate channel state information (CSI) prediction is becoming increasingly critical for reliable and efficient 6G wireless communications. CSI prediction at the edge cannot be treated as a fixed-distribution problem, since user mobility, trajectory change, and link-quality variations continuously shift the CSI patterns. The performance of offline-trained predictors may degrade after deployment, while sequential fine-tuning often causes forgetting. To address this issue, we propose SAD-CP, a Spatially Anchored Dual-distilled Continual Prediction framework for edge-side dynamic CSI tracking. The key idea is to preserve relatively stable spatial propagation priors while adapting task-sensitive dynamic components. Specifically, the radio-map backbone is frozen during edge adaptation, while the temporal encoder, trajectory encoder, spatial mapping head, and decoder are updated. Output distillation, spatial feature distillation, and signal-quality-aware weighting are further introduced to mitigate forgetting. Experiments on three NVIDIA Sionnabased continual task streams show that SAD-CP achieves average NMSEs of 0.0848, 0.1432, and 0.0914, respectively, on trajectoryshift, SNR-shift, and mixed-shift streams. Compared with naive fine-tuning, our scheme reduces the average NMSE by 34.82%, $\mathbf{3 5. 0 6 \%}$, and $\mathbf{4 0. 5 9 \%}$ on the three streams, respectively.