Wind Energy Utilization for UAVs Via Deep Reinforcement Learning

Improving the endurance and energy efficiency of unmanned aerial vehicles (UAVs) is essential for advancing green aviation. Inspired by the albatross’s ability to glide long distances without flapping, dynamic soaring enables UAVs to harvest wind energy in flight. However, traditional trajectory optimization methods rely on complete wind field knowledge and involve high computational cost, limiting onboard applicability. This study proposes a deep reinforcement learning (DRL) framework to achieve autonomous dynamic soaring from the perspective of AI. A six-degree-of-freedom UAV model and linear wind field are integrated into a TD3-based training system. To reduce training complexity and enhance generalization, different flight–wind angle scenarios are unified using a fixed northward wind and variable initial yaw angles. Results show that the trained agent can complete a 100 m unpowered flight while gaining up to 32.3% energy, with flight paths resembling those from optimal control. The agent also learns key energy-harvesting strategies, such as exploiting wind-relative heading and vertical transitions, without prior modeling. This work demonstrates the potential of DRL for real-time, energy-aware UAV control, offering a scalable solution for intelligent and sustainable flight systems.