Sanket Sanjay Diwale PhD Researcher EPFL STI IGM LA3 ME C2 396 (Building ME) Station 9 1015 Lausanne Switzerland sanket.diwale@epfl.ch la.epfl.ch
Predictive Control of an Airborne Wind Energy System
using Gaussian Process Models
Sanket Sanjay Diwale, Ioannis Lymperopoulos, Colin N. Jones Department of Mechanical Engineering, Automatic Control Laboratory, EPFL The performance of an Airborne Wind Energy (AWE)
sys-tem in crosswind operation depends on the shape and characteristics of the flight path. A control scheme that updates the path online, requires a component responsi-ble for following the requested path despite disturbances and unknown dynamics. We present simulation studies for path following using data-based models.
We demonstrate an online data-based method to iden-tify the dynamics of an AWE kite system using Gaussian processes. A model available from literature [1] is used as a priori information to train a Gaussian process from observed data and to select the state space representing the dynamics. We create a map from the steering input of the kite and states to the future states and their predic-tion uncertainty. We then formulate a stochastic model predictive control (SMPC) scheme that takes into consid-eration uncertainty in the learned model. Virtual states and a virtual input are used to define a second order in-tegrator with the states parametrising and selecting the part of the reference path to be followed in the MPC hori-zon. With this framework we correct model-observation mismatch and perform path following. The virtual input allows the algorithm to choose the speed at which fol-lowing the path is feasible for the kite. We consider input box constraints and input ramp constraints only, however more general state-input constraints can also be incorpo-rated in the MPC scheme.
To test our method we generate optimal and several sub-optimal feasible trajectories. We then train our model
us-ing an expert that tries to track a path before automatic operation. Once the model is learned with sufficient ac-curacy we commence the automatic path following. The figure shows an instance of a generated feasible trajec-tory. The method is tested to track a different part of the path than on which it was trained to ensure that the learning generalises well. The algorithm continues to in-corporate information from the new path to better cap-ture the model dynamics while continuously solving the SMPC problem. In general the approach leads to a non-linear programming (NLP) problem and additional work is required to make the NLP suitable for real time imple-mentation, which is a topic under investigation.
Path tracking using GP-MPC on a 500 m radius reference sphere: red line - path to be tracked, red dot - kite position, green line - past 10 s of kite trajectory.
References:
[1] Erhard M., Strauch H.: Control of Towing Kites for Seagoing Ves-sels. IEEE Transactions on Control Systems Technology, Vol. 21, pp. 1629–1640 (2013)