Deep Learning-based Beamforming and Blockage Prediction for Sub-6GHz/mm Wave Mobile Networks

To meet the stringent demands of Beyond 5G applications, an optimized and seamless usage of sub-6 GHz and mmWave networks under high user mobility is essential. In particular, to alleviate the heavy burdens of mmWave channel state information feedback, we propose a Deep Learning-based scheduler at the base station that predicts the future mmWave blockage status and optimal beamforming vectors of the mobile user, solely based on sub-6 GHz channel knowledge, i.e., out-of-band information. The designed Deep Neural Network (DNN) comprises Long Short-Term Memory layers to extract the temporal correlation from the low-frequency channel knowledge, for enabling an accurate blockage and beam prediction. We investigate the influence of the available past channel information and of the user speed on the prediction accuracy and on the achievable data rates. Simulation results show that the proposed method significantly improves the prediction accuracy of optimal mmWave beamformers compared to the benchmark DNN with much reduced complexity. Furthermore, the proposed DNN demonstrates high robustness for different user speeds, while approaching optimal rates achieved through exhaustive search and perfect blockage knowledge.