Beamforming optimization can enhance the next-generation wireless networks. However, finding the optimal beamforming in real-time is hard due to the need for large beam training overhead. The problem can be more challenging in dynamic environments with small coherence channel time. In this paper, we propose an accurate beam prediction solution using light detection and ranging (LiDAR)-assisted radio frequency (RF) system. More specifically, we propose a deep-learning model based on long-sort term memory (LSTM) to predict future beam indices from a set of pre-defined beam steering codebook. In addition to solving the beamforming overhead problem, the proposed deep learning approach is a model-free approach that can be applied with no required knowledge of the channel state information. We compare the proposed model with the cross-validation results in addition to other benchmarks using the top-k accuracy metric. The numerical results show that the proposed scheme has achieved a top-1 accuracy of 84.6% compared to an 80.3% for cross-validation while the other benchmarks have achieved a 57.5% using the same data set.


Electrical and Computer Engineering

Keywords and Phrases

beamforming; deep-learning; light detection and ranging (Li-DAR); long-sort term memory (LSTM)

International Standard Book Number (ISBN)


Document Type

Article - Conference proceedings

Document Version


File Type





© 2023 Institute of Electrical and Electronics Engineers, All rights reserved.

Publication Date

01 Jan 2023