Energy-Efficient Wireless Power Transfer for Sustainable Federated Learning

Federated Learning (FL) has attracted great attention in recent years and is considered as an enabling technology in future smart wireless networks. Nevertheless, this learning paradigm faces a severe challenge in its implementation procedure, i.e., energy shortage issue. Different from the traditional centralized training paradigm, the training procedure of FL is carried out on mobile devices. Generally, the training tasks are computation-intensive and may involve several communication rounds for transmitting large-sized machine learning models, which indicates that they are high energy-consuming. This characteristic increases burden on mobile devices with limited battery capacity. In this paper, we employ the Radio Frequency (RF)-based Wireless Power Transfer (WPT) technology and time switching energy harvesting architecture to realize a sustainable FL framework, and then design a resource optimization strategy based on the dual and line search methods to minimize the amount of Transferred Energy (TE) required for completing the learning. Moreover, we interpret the Karush–Kuhn–Tucker (KKT) conditions of the formulated problem and obtain some engineering insights. Simulation results verify the convergence of the proposed resource optimization strategy and demonstrate the advantage of the proposed framework over the existing work in terms of TE.