Learning-Based RIS Element Allocation and SWIPT Optimization for Active RIS-Aided UAV Networks

In this paper, we address the challenge of energy-aware operation in unmanned aerial vehicle (UAV)-assisted communication systems equipped with active reconfigurable intelligent surface (RIS). We propose a dual-domain control model that jointly optimizes the energy harvesting (EH) duration, user transmit power, RIS-user scheduling, phase shifts, and amplification factors. To eliminate the need for conventional power splitters, the RIS is spatially split such that a subset of elements harvest energy while others forward signals, enabling simultaneous wireless information and power transfer (SWIPT) directly at the RIS level. The harvested energy assists the UAV’s onboard battery in powering the active RIS electronics, thereby enhancing system endurance and ensuring coverage. The joint optimization problem is formulated as a constrained Markov decision process (CMDP) with a hybrid discrete–continuous action space. To solve it, we develop a framework based on the softmax deep double-deterministic policy gradient (SD3) algorithm, incorporating critic ensemble learning and soft Bellman updates for improved training stability. The proposed algorithm enables real-time resource control while satisfying QoS, power, and scheduling constraints. Simulation results show that the proposed scheme consistently outperforms baseline DRL methods such as DDPG, PPO, and TD3. In particular, it achieves up to 65% EH efficiency under active RIS settings, with robust performance across different RIS scales and user densities.