Energy-Efficient Resource Management for Federated Learning in LEO Satellite IoT
Zhou Ai, Ying Wang, Qiuyang Zhang · 2024
Federated learning (FL) is a paradigm that enables model training across various devices while keeping the data localized. However, for battery-powered passive devices in the satellite Internet of Things (IoT), the continuous update and transmission of the local model result in heightened energy consumption on the device side. To address this challenge, an FL framework with partial device participating is proposed. In this framework, the on-board controller strategically selects a subset of devices to upload local model parameters, effectively mitigating the overall energy consumption on the device side. Constrained by transmission power and transmission delay, a resource allocation problem is formulated. This problem jointly optimizes the uploading strategy and transmission power, aiming to minimize the utility function that combines the global model loss and energy consumption over multiple rounds of FL. Simulation results demonstrate that, compared with other benchmark schemes (DDPG, PPO), the proposed algorithm achieves energy efficiency in FL.