A Novel Neural Network Approach to Proactive 3-D Beamforming

This study explores three-dimensional proactive beamforming at millimeter wave frequencies using transformer neural networks (TNNs), long short-term memory networks (LSTMs) and gated-recurrent units (GRUs). The proposed scheme aims to reduce beamforming latency by predicting future directions of arrival (DoAs) based on past observations, allowing the system to prepare beamforming weights proactively. We simulate an urban environment using OpenStreetMap data to generate realistic movement paths, creating a comprehensive dataset for training and evaluation. Our focus is on the predictive capacity of TNNs, LSTMs and GRUs to anticipate future DoAs, even in non-line-of-sight scenarios influenced by urban infrastructure. We detail the environment simulation setup, the ray-tracing mechanism as well as the movement generation process for pedestrians and vehicles. A statistical analysis on the prediction accuracy and response time is presented to assess the most accurate model and discuss the trade-offs between the architectures. In addition, an end-to-end AI-based proactive beamforming scenario is examined where zero-forcing is applied on moving users. This is to further demonstrate and evaluate the capabilities and the performance of each model. Our findings suggest that proactive beamforming can significantly enhance performance in dynamically changing urban landscapes, offering a promising avenue for future research and development in adaptive communication systems.