Green Power Control for Multiuser Uplink Communications: Cnn and Dqn Perspectives

Renewable energy harvesting (EH) has been recognized as an effective means to realize perpetual wireless communications in the era of Internet of Thing (IoT) with low-powered wireless devices. The design of EH communications is adversely affected by the uncertainty and causality of energy arrivals in the battery replenishment, and it is thus imperative to harmonize the harvested energy for data transmissions over the time horizon. To achieve the optimal performance, the existing power control schemes that utilize the convex optimization require future knowledge of EH information (EHI) and channel state information (CSI), which is hard to be acquired in reality. The goal of this project is to investigate the green power control for multiuser uplink communications that rely on solar power as an energy source to recharge the battery. By means of the optimization capability of convex optimization as well as the prediction capability of deep learning, we aim at studying the multiuser power control to maximize the “long-term” uplink sum rate but only with the past knowledge of EHI and CSI, under the EH and storage constraints in single-cell and multi-cell scenarios. Specifically, in the single-cell scenario, offline power control is investigated based on convolutional neural networks by integrating the geographical location, solar EH, battery, and channel information into radio resource maps. In addition, online power control is designed based on deep reinforcement learning by taking various system states like solar, channel and battery into account, and applying multilayer perceptron for long-term sum rate prediction. As an extension to the multi-cell scenario, we further include the past inter-cell interference-related information into the power control designs, which implicitly reflects the EH and spending course of other cells and enables each cell to intelligently manage the harvested energy for data transmissions. Computer simulations will be conducted to rigorously evaluate the system performance of the proposed offline and online green power control schemes.