This paper proposes an advanced non-terrestrial communication architecture that integrates Rate-Splitting Multiple Access (RSMA) with a Beyond-Diagonal Active Reconfigurable Intelligent Surface (BD-ARIS) mounted on a UAV under the coverage of a Low Earth Orbit (LEO) satellite. The BD-ARIS adopts a group-connected structure to enhance signal amplification and adaptability, while RSMA enables efficient multi-user access by dividing messages into common and private components. The system jointly optimizes satellite beamforming, UAV positioning, power allocation, and rate-splitting ratios to maximize the overall energy efficiency (EE). To solve the resulting non-convex and high-dimensional problem, we employ three state-of-the-art deep reinforcement learning (DRL) algorithms: Trust Region Policy Optimization (TRPO), Twin Delayed Deep Deterministic Policy Gradient (TD3), and Asynchronous Advantage Actor-Critic (A3C). Moreover, realistic models for the power consumption of both the UAV and the BD-ARIS are considered. Simulation results reveal that TRPO consistently achieves the best performance in terms of EE and sum rate, especially under high transmit powers and challenging deployment scenarios. TD3 converges faster and performs competitively in moderate settings, while A3C suffers from instability due to its high variance. Additionally, the robustness of each algorithm under channel state information (CSI) uncertainty is evaluated, confirming TRPO resilience to imperfect observations. Overall, the proposed RSMA-BD-ARIS framework significantly outperforms conventional RIS-assisted designs and provides a scalable, energy-efficient solution for 6G and massive IoT applications in non-terrestrial networks.