Task-Oriented Multiobjective Computation Offloading in LEO Mega-Constellation Edge Computing Network

Abstract

The low earth orbit (LEO) mega-constellation network, with its extensive coverage and low-latency characteristics, offers new opportunities to meet the demands of computation-intensive and latency-sensitive applications in remote areas. However, with the increasing complexity of task offloading demands and the limited availability of satellite resources, resource management and scheduling face significant challenges. To tackle these challenges, we propose a satellite-terrestrial integrated LEO mega-constellation edge computing network (LMCECN) management architecture, which enables satellite-terrestrial resource allocation and task offloading through the cooperative scheduling of primary and secondary satellites. Based on this architecture, we design a deep reinforcement learning-based task-oriented mega-constellation edge offloading (TOMEO) scheme, which significantly improves task offloading efficiency by incorporating task sorting and resource clustering preprocessing mechanisms. Furthermore, a multiobjective double dueling noisy deep Q-network (DDNDQN) algorithm is introduced, which comprehensively considers multiple optimization objectives, including task completion rate, load balancing degree, task delay, and energy consumption, further enhancing task offloading efficiency. The experimental results demonstrate that the proposed offloading scheme outperforms the baseline schemes across all optimization objectives and improves the task offloading performance.