Data-Driven Koopman-based Modeling and Control of Quadcopter in a Turbulent Wind Field

We present a data-driven approach to model wind effects on quadcopter trajectory tracking and make use of Model Predictive Control (MPC) to improve the tracking performance in a turbulent wind field. To generate datasets for learning, we employ a quadcopter dynamic model enhanced with additional terms that incorporate wind effects on body and rotor dynamics. We propose a tracking error-based learning model based on the Koopman operator theory. We collect the data with an existing controller in the loop and design MPC controllers to augment the existing controller. The Koopman-based model captures the nonlinear dynamics in quadcopter trajectory tracking in a linear form. We design two different MPC controllers and evaluate their performance using Monte Carlo simulations. The controllers differ in how the wind is modeled and predicted. We demonstrate that these controllers improve tracking performance and reduce tracking errors better.