Tension Control using Reinforcement Learning for Airborne Wind Energy Systems

As society continues to expand and evolve, the energy demand grows to sustain living standards while supporting technological advancement and ecological sustainability. Renewable energy will play a pivotal role in meeting future energy requirements and achieving energy independence. However, the deployment of current renewable energy systems is often constrained by geographic suitability and environmental considerations. Despite its potential, existing wind energy infrastructure is hindered by limitations such as restricted mobility, substantial production and maintenance costs, and operational sensitivity to wind conditions. Airborne Wind Energy (AWE) systems offer a compelling alternative, enabling higher power generation at lower cost than traditional wind turbines. This paper presents a model-free deep reinforcement learning (DRL) framework to maximize reel-out energy in a kite-based AWE system that follows a predefined figure-eight orbit. The controller maps an 84-dimensional state vector (the kite’s position, velocity, attitude, angular rates, tether tension, and tether length at six orbit waypoints) to discrete tension-force commands without requiring explicit wind measurements. Deep Q-Network (DQN) and Advantage Actor--Critic (A2C) policies are trained in a MATLAB simulation environment under wind magnitudes from 6 to 12~m/s using replay buffers generated via exhaustive behavior-policy exploration. Results show that both methods recover optimal actions and power at low winds (6--7 m/s); at higher winds, A2C provides relatively stable near-optimal performance with small power reductions (about 2--6%), while DQN achieves exact optimality at several wind speeds but with greater variability and a worst-case reduction of about 10% at 10 m/s. Overall, the learned policies closely match the exhaustive-search optimum across operating conditions, demonstrating the feasibility of wind-agnostic, data-driven control for AWE energy maximization.