Energy–Reliability Tradeoff in Vehicular LEO Multi-Connectivity: A Deep Q-Network Framework

Abstract

This letter studies reliable and energy-efficient uplink communication in vehicular satellite networks with time-varying line-of-sight (LOS) or non-line-of-sight (NLOS) conditions and satellite congestion, which challenge joint connectivity and hybrid automatic repeat request (HARQ) control. In this letter, we propose a deep Q-network (DQN) framework that jointly optimizes satellite association, HARQ retransmission levels, and transmission energy with adaptive multi-connectivity (MC) for vehicular users. A capacity-aware load model is incorporated to capture the impact of background traffic on link availability. Simulation results under two-line element (TLE) derived orbital positions and vehicular mobility show that energy-efficient heuristics achieve lower energy consumption by sacrificing the reliability requirements. In contrast, the proposed DQN effectively enforces the reliability requirement as a penalty- based (soft-constrained) formulation that empirically satisfies the reliability target, achieving reliability above 0.995 while reducing the average energy consumption by approximately 57% in all evaluated scenarios, approaching the Pareto-efficient frontier. These results demonstrate that load-aware reinforcement learning enables reliable and energy-efficient adaptive MC in vehicular low Earth orbit (LEO) networks.