Data-Efficient Throughput Prediction on 28-GHz 5G via Geometry-Aware Preprocessing and Mobility-Diverse Pre-Training

Abstract

Millimeter-wave communications are highly vulnerable to line-of-sight (LOS) blockage, which can cause severe link quality (LQ) degradation. To enable proactive control for stable mmWave communication, deep neural network (DNN)-based LQ prediction using physical space information—such as the user equipment (UE) position, orientation, and velocity—has been widely studied. Since supervised learning requires large-scale datasets for each base station (BS), data-efficient models that generalize across BSs with limited training data are essential. This paper proposes a data-efficient throughput prediction framework integrating geometry-aware preprocessing of physical space information and mobility-diverse pre-training. The proposed preprocessing normalizes UE position and orientation relative to each BS, while the pre-training strategy leverages diverse mobility patterns to enhance adaptability across environments. Experiments conducted with a single UE setup using robot-controlled routes in two indoor BS environments of a commercial 5G network including 28-GHz channels demonstrate that, under the evaluated conditions, the framework reduces the throughput prediction root-mean-square error (RMSE) by up to 56.3% with 100 target-environment training samples, compared with environment-specific models trained from scratch. Furthermore, the geometry-aware preprocessing effectively mitigates negative transfer during adaptation to target scenarios, while mobility-diverse pre-training improves the detection rate of low-throughput states below 100 Mbps by up to 47.8 percentage points.