Real-Time Experimental Optimization of Closed-Loop Crosswind Flight of Airborne Wind Energy Systems via Recursive Gaussian Process-based Adaptive Control

Power generated by airborne wind energy (AWE) systems can be dramatically increased by moving the system in controlled crosswind flight patterns. In this paper, the crosswind flight pattern of an AWE system is optimized in real time using recursive Gaussian process (RGP)-based adaptive control. RGP-based adaptive control fuses machine learning tools with real-time adaptive control principles. Traditionally, Gaussian process (GP) modeling requires a complete database of all the previously tested data points and their associated performance values. By utilizing a recursive update law, the RGP-based modeling used here avoids the need to maintain a complete database. The RGP-based modeling estimates a predictive mean and variance model over the control parameter design space based on instantaneous performance feedback. The candidate design space is explored by selecting points at locations of maximum uncertainty. Design points that are determined to be statistically inferior to the perceived optimum are rejected from the candidate design space. In this work, the RGP-based adaptation is applied to a lab-scale platform for experimental crosswind flight of AWE systems. Experimental crosswind flight results presented here demonstrate a 60% increase in power augmentation over traditional stationary flight.