Spatial-Temporal Beam Prediction: A Deep Learning-Based Approach With Compressed Multi-Beam Measurements

Aided with artificial intelligence, a great deal of beam management methods aiming to reduce the training overhead have emerged. However, these codebook-based approaches are confined to decrease the number of measured beam pairs and ignore the compressibility of the measured beam data. In this paper, we propose a novel beam prediction framework that leverages compressed multi-beam measurement as the model input and jointly optimize the multi-beam compression and reconstruction process via deep learning. Specifically, we build a spatial-temporal beam prediction based on compressed sensing, and design a neural network to infer the optimal beam pair for transceivers with reduced training overhead. Simulation results show that the proposed beam prediction approach outperforms the existing AI-based solutions even at a high moving speed with almost negligible increase in model storage overhead as well as computational complexity.