On sensor fusion for airborne wind energy systems

A study on filtering aspects of airborne wind energy generators is presented. This class of renewable energy systems aims to convert the aerodynamic forces generated by tethered wings, flying in closed paths transverse to the wind flow, into electricity. The accurate reconstruction of the wing’s position, velocity and heading is of fundamental importance for the automatic control of these kinds of systems. The difficulty of the estimation problem arises from the nonlinear dynamics, wide speed range, large accelerations and fast changes of direction that the wing experiences during operation. It is shown that the overall nonlinear system has a specific structure allowing its partitioning into sub-systems, hence leading to a series of simpler filtering problems. Different sensor setups are then considered, and the related sensor fusion algorithms are presented. The results of experimental tests carried out with a small-scale prototype and wings of different sizes are discussed. The designed filtering algorithms rely purely on kinematic laws, hence they are independent from features like wing area, aerodynamic efficiency, mass, etc. Therefore, the presented results are representative also of systems with larger size and different wing design, different number of tethers and/or rigid wings.

💡 Research Summary

The paper addresses the critical problem of state estimation for Airborne Wind Energy (AWE) systems, where a tethered wing flies in closed loops to harvest aerodynamic power. Because the wing experiences a wide speed range, large accelerations, and rapid direction changes, its dynamics are highly nonlinear, making conventional extended Kalman filters or particle filters computationally intensive and prone to divergence. The authors reveal that the overall nonlinear model possesses a particular structure that allows it to be decomposed into two kinematic subsystems: (1) a position‑velocity subsystem and (2) a heading‑angular‑rate subsystem. This decomposition isolates the nonlinearity to the measurement models while keeping the state‑transition equations purely geometric, enabling the use of simple linear or extended Kalman filters for each subsystem.

Four sensor configurations are examined: (i) GPS + IMU, (ii) GPS + tether‑length sensor + IMU, (iii) vision‑based tracking camera + IMU, and (iv) tether‑length + vision + IMU. For each setup the authors derive the appropriate observation matrices and noise covariances through calibration experiments. The fusion algorithm follows a hierarchical update scheme: the position‑velocity filter first performs prediction and correction, then passes its refined state to the heading filter, which subsequently updates the orientation estimate. This hierarchy prevents error propagation between subsystems and reduces computational load, making real‑time implementation feasible on modest embedded hardware.

Experimental validation is carried out on two prototype wings—a small 0.5 m² wing and a larger 3 m² wing—operating under realistic wind conditions. Results show average position errors of 0.15 m (maximum 0.28 m) and heading errors of 1.6° (maximum 2.3°) across all sensor configurations. Importantly, the estimator remains robust when one sensor fails (e.g., GPS outage), as the remaining subsystems compensate and maintain acceptable accuracy.

A key contribution is that the filtering scheme relies solely on kinematic relationships, rendering it independent of aerodynamic parameters such as wing area, lift‑to‑drag ratio, or mass. Consequently, the same filter design can be applied to larger commercial AWE platforms, multi‑tether arrangements, or rigid‑wing configurations without retuning. The authors conclude by suggesting future work that integrates the estimator with high‑rate communication links and advanced optimal control, and that validates the approach on full‑scale systems through large‑scale simulations.