A Deep Learning Model for Optimizing Downlink Parameters of Wireless Information and Power Transfer in Massive MIMO Networks

Massive multiple-input multiple-output (MIMO) technology with a base station (BS) equipped with a large number of transmit antennas is a promising application for wireless information and power transfer (WIPT) in energy-constrained wireless networks. Recently, a transmit time switching (TS) protocol, where the information and energy are transferred over a different fractions of a time-slot, has received much attention to enable WIPT. The optimal beamforming design and time-fraction allocation for information and energy transfer are important to maximize the worst-case throughput while fulfilling the energy harvesting requirement. However, the existing solution in literature is computationally demanding due to its dependence on off-the-shelf solvers such as CVX. In this work, a deep learning model is sought to predict the optimum resource allocation parameters in WIPT-based massive MIMO systems. The model is composed of three fully connected networks, each of which is trained to predict certain optimization variables. The proposed network enables the system to quickly update the required optimization parameters as the network condition changes without revisiting the original optimization problem. Moreover, it is found that user distances are sufficient input features to train the networks, as training using channel state information (CSI) of the users renders the proposed solution time-inefficient, as a channel estimation technique is required. It is also found that training using distances yields more accurate results than training using CSI, which suggests that the deep neural network (DNN) extracts higher quality features from spatial information than the features extracted given the corresponding CSI.