再生能源獵取已被公認為是在物聯網時代使用低功率無線設備實現永久無線通訊的有效方式。然而，能量獵取通訊的設計受到電池充電時能量到達的不確定性和時間因果關係的不利影響，因此必須協調在一段時間範圍內用於資料傳輸所獵取的能量。為了獲得最佳效能，利用凸優化的現有功率控制方式需要知道未來的能量獵取訊息和通道狀態訊息，而這在現實中很難獲得。本計畫的目標是研究多用戶上鏈通訊的綠能功率控制，而這些通訊節點依賴太陽能作為能源為電池充電。藉由凸優化的優化能力以及深度學習的預測能力，我們旨在研究單細胞及多細胞環境下之多用戶上鏈通訊功率控制，在滿足能量獵取時間因果關係與電池儲存能量約束下及僅需要知道過去的能量獵取訊息和通道狀態訊息條件下，最大化上鏈通訊的“長期”總資料傳輸速率和。具體而言，在單細胞情境中，通過將地理位置、太陽能能量獵取、電池存量和通道訊息整合到無線電資源地圖中，本計畫將基於捲積神經網絡研究離線功率控制。此外，在線功率控制是基於深度強化學習而設計的，其中考慮了太陽能、通道和電池等各種系統狀態，並將多層感知器應用於長期總資料傳輸速率預測。同時將設計延伸至多細胞情境中，吾人進一步將過去的細胞間干擾資訊用於功率控制設計中，此資訊隱含地反映了其他細胞的能量獵取和使用過程，並使每個細胞能夠智能地管理所獵取到的能量，以用於資料傳輸。最後將進行電腦模擬，以嚴格評估所提出的離線和在線綠能功率控制方法之系統效能。 ;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.
