Lightweight Channel Prediction for Low-Altitude UAV Communications via Knowledge Distillation

Abstract

The high mobility of uncrewed aerial vehicles (UAVs) leads to severe channel aging, which fundamentally limits beam tracking and link reliability in massive multiple-input multiple-output (MIMO) systems. To address these challenges, the knowledge-distillation (KD)-based lightweight channel prediction framework is proposed in this article for UAV communications. First, we design an Informer-based teacher network to efficiently capture long-range temporal dependencies of air–ground channels. Then, we construct a lightweight 1-D convolutional neural network (1D-CNN) as the student model and introduce a hybrid loss that fuses hard supervision from true channel state information (CSI) labels with soft targets from the teacher to transfer temporal knowledge. Finally, to demonstrate the generality of the framework, we further extend MobileNetV3 as an alternative student backbone to demonstrate the generalization capability and robustness of the proposed framework. Simulation results show that the proposed KD-based students achieve a 1.2-dB normalized mean square error (NMSE) gain while the inference latency is reduced to one quarter of that of the nondistilled baselines.