Continuous-Time mmWave Beam Prediction With ODE-LSTM Learning Architecture
Ke Ma, Fan Zhang, Wenqiang Tian, Zhaocheng Wang
Abstract
In order to avoid frequent millimeter-wave (mmWave) beam training in high-speed scenarios, in this letter, we propose to exploit the neural ordinary differential equation (ODE) to predict the arbitrary-instant optimal beam between the current and next beam training instants. Specifically, long short-term memory (LSTM) network is utilized to model the beam dynamics based on the received signals of previous periodical beam training, and the ODE solver is adopted to learn the derivative of beam variations, so that the optimal beam at arbitrary instant can be predicted by integrating the derivatives. Simulation results demonstrate that our proposed ODE-LSTM assisted methodology could achieve higher beamforming gain over its state-of-the-art counterparts.