Neural-Network-Based Power Control Prediction for Solar-Powered Energy Harvesting Communications

In this article, we design neural network (NN)-based transmit power control prediction for solar-powered energy harvesting (EH) communications under single-user (SU) and multiuser (MU) scenarios with real solar data. Although the directional water filling (DWF) is known as the optimal scheme for the SU case, it necessitates the full (past and future) channel state information (CSI) and energy state information (ESI) in realizing the optimal solution over a time period. To conquer this impracticality, an SU-GreenPCNet, which only requires the past short-term CSI and ESI for predicting the SU transmit power, is proposed and trained with the historical solar data. For the MU case, two iterative algorithms, namely, weighted-sum minimum mean-square error (WMMSE) and iterative DWF (IDWF), are investigated to tackle the original nonconvex power control problem when the full state knowledge is assumed to be known in advance. The solutions with the historical solar data are then served as benchmarks for designing two MU-GreenPCNets. As an extension of the SU-GreenPCNet, a centralized scheme is proposed at the central controller for jointly determining the MU transmit power values based on the past short-term state knowledge of all users. A distributed scheme is further investigated to reduce the signaling overhead, in which each transmitter merely utilizes the past short-term CSI, ESI, and MU interference (MUI) knowledge associated with its user pair. The simulation results show that the proposed power control prediction schemes can pragmatically achieve satisfied sum rate performance in both SU and MU scenarios, as compared with the benchmark schemes.