SwitchConvNet: Deep Learning Framework for Mobility-Aware Spatial Beam Prediction in Beyond 5G and 6G Systems

Abstract

Beam management (BM) plays an important role in maintaining reliable links in next-generation wireless systems, especially at millimeter-wave and terahertz frequencies where massive MIMO (mMIMO) is used. Conventional 5G beam selec- tion methods often suffer from acquisition delays and signaling overhead. In many cases, reducing latency also lowers throughput because the beam search becomes less granular. To address these issues, we propose SwitchConvNet, a deep learning (DL)- based spatial beam prediction framework designed for mobility- aware urban macro cell (UMa) scenarios. The proposed approach reduces beam acquisition time by more than 80% while also improving throughput compared with classical non-DL methods. It achieves correct top-1 beam prediction in 77% of cases and places the optimal beam within the top-3 predictions in nearly 97% of cases. In addition, the adaptive switching filter (ASF), which combines spline and Gaussian filtering, further improves prediction accuracy, increasing top-1 performance and reducing the RSRP estimation error by about 0.08 dB. Importantly, the framework remains effective under practical constraints such as quantized RSRP reporting and also reduces feedback and training overhead. These results show that SwitchConvNet is a scalable and practical option for future wireless networks.