GPS-Based Beam Prediction Using Lightweight Deep Learning Models for mmWave Networks

Millimeter-wave (mmWave) communication systems use narrow, directional beams, but exhaustive beam training is costly in dynamic settings. Prior location-assisted methods often rely on synthetic data that ignores the noise of real-world GPS measurements. We propose a unified framework that first denoises GPS trajectories with Gaussian-process regression and then predicts beams using a bidirectional long short-term memory network with an attention mechanism. Across nine real-world scenarios, our approach improves Top-k accuracy by up to 36%, reduces received-power loss by more than 1 dB, and cuts beam-training overhead by up to 95%. These results highlight the effectiveness of the proposed framework in bridging the gap between simulation-driven research and real-world mmWave beam alignment.