| 摘要: | 近年來人工智慧與物聯網(AIoT)在無線通訊應用蓬勃發展，無線通訊所造成的能源消耗及電量需求更是不可忽視，而能量獵取技術被提出此解決有限電池電量問題。為了最大化系統通道容量，傳統的凸優化方法受電池容量和能量獵取時間因果性限制必須假設未來所有時刻的能量獵取及通道增益等狀態資訊是完美已知才能實現傳輸功率控制，這在實際情況下難以達成且具有高計算複雜度。因此本論文提出深度神經網路(Deep Neural Network, DNN)學習過去能量獵取的傳輸模式來預測未來功率控制值，所提方法不需知道未來能量獵取及通道增益等狀態資訊，且有效降低複雜度以達到實際即時的綠能無線通訊應用。
在本篇論文中，具體地提出基於深度神經網路的功率控制方法，用於單用戶及多用戶綠能無線通訊系統之功率控制以最大化系統通道容量。在單用戶環境下，使用凸優化(Convex optimization)及歷年能量獵取資料產生最佳功率控制值，再讓多層感知器(Multilayer perceptron, MLP)學習其輸入/輸出的關係。在多用戶環境中，吾人研究集中式與分散式的功率控制技術，主要區別在於後者不需取得其他用戶的資訊。為了讓多層感知器進行學習，分別提出加權和最小均方誤差(Weighted-sum MMSE, WMMSE)和遞迴方向性注水演算法(Iterative Directional Water-filling, IDWF)產生參考功率控制值。模擬結果顯示本論文提出的方法不僅符合實際應用還能大幅降低計算複雜度並且在單用戶及多用戶應用上都能近似於參考功率控制值的系統通道容量。
;In recent years, the artificial intelligence and the Internet of Things have been widely developed in wireless communication applications. The increasing demand on energy consumption in wireless communications has stimulated the rapid development of energy harvesting technology to solve the problem of limited battery power at wireless nodes. The convex optimization is a common approach to control transmission power over a future finite time horizon for achieving the maximum system throughput. It, however, requires non-causal perfect knowledge of energy harvesting patterns and channel gains over the power control duration due to the battery storage and the energy causality constraints, resulting in high complexity and impractical in real applications. To overcome this problem, in this thesis, deep neural networks (DNNs), along with the use of past energy harvesting patterns, are proposed to predict the future power control values, which can effectively reduce the computation complexity for real-time applications.
Specifically, the DNN-based power control schemes are proposed for maximizing the system throughput in two scenarios of energy harvesting communications: single-user and multi-user setups. In the single-user scenario, the convex optimization is utilized to generate the optimal power control values based on the historic energy harvesting data, and the results are later used in the multilayer perceptron (MLP) to learn its input/output relationships for power control prediction. In the multi-user scenario, centralized and distributed MLP-based power control schemes are investigated, and the main difference between them lies that the former scheme requires the past energy harvesting conditions and channel gains of all users, while each user only requires the past energy harvesting conditions, channel gains and interference power values related to itself in the later scheme. For the learning of MLPs in the multi-user scenario, a weighted-sum minimum mean-square-error (MMSE) approach and an iterative directional water-filling approach are respectively proposed to generate the reference power control solutions in the centralized and distributed designs. The simulation results show that the proposed method can greatly reduce the computational complexity and approach the system throughput of the reference solutions very well in both single-user and multi-user application. |
