在物聯網時代，從外在環境獵取能量技術被視為一項維持可充電式低功耗無線裝置運作的重要解決方案。即使可獵取的能量是長久且源源不絕，使用能量獵取技術於無線通訊的主要挑戰在於電池充電中能量獲得的未知性及時間因果性造成時間上的能量使用排程更加不易。現存方法使用凸優化於功率控制設計，但不合理地假設無線裝置能完美地知道未來能量獲取資訊。本研究計畫的目的在於設計基於太陽能獵取之可充電式綠能無線通訊功率控制。吾人將使用過去太陽能大數據資料於凸優化及深度學習，在滿足有限電池容量及能量獲取時間因果性限制下，研究不需知道未來能量獲取資訊的單用戶及多用戶綠能無線通訊系統之功率控制預測以最大化系統容量。具體而言，在單用戶通訊環境下，吾人將研究多層感知器與長短期記憶神經網路，以及搭配不同的損失函數設計來智能地預測及最佳化獵取能量的使用。吾人進一步延伸單用戶設計至功率控制預測更為關鍵的多用戶干擾通訊環境，依據多用戶與共同終端數據交換量的多寡，研究適用於多用戶的集中式與分散式功率控制技術。最後吾人將藉由電腦模擬來評估所提基於深度學習之功率控制方法的系統效能。 ;Energy harvesting from ambient environments has been considered as a potential remedy for providing perpetual energy to low-powered wireless devices, especially in the era of Internet of Things (IoT). While the harvested energy is unlimited for a long time, a major concern for applying energy harvesting in wireless communications lies in that the uncertainty and causality of energy arrivals in battery replenishment make it difficult to schedule energy usage over a finite time horizon. The existing schemes utilize convex optimization for power control but unreasonably assuming that the harvested energy profiles in the future are perfectly known to the wireless devices. The goal of this project is to design the power control of green wireless communications that rely on solar power as an energy source to replenish the battery. By means of historical real solar data in deep learning and convex optimization, we investigate the prediction of power control, which does not require the knowledge of prospective energy arrivals, for maximizing the system capacity, subject to battery storage and energy harvesting constraints, in both single-user and multi-user scenarios. Specifically, we resort to the multilayer perceptron (MLP) and the long short-term memory network (LSTM) with various designed loss functions for intelligently optimizing and predicting energy usage in the single-user scenario. As an extension to the multi-user scenario, in which power control prediction becomes much more critical under multi-user interference, we study centralized and distributed power control schemes for multi-users with different amounts of information exchange between the users and the center. Computer simulations will be conducted to rigorously evaluate the system performance of the proposed deep learning-based power control schemes.
