On the Estimation of Vector Wind Profiles Using Aircraft-Derived Data and Gaussian Process Regression

Abstract

This work addresses the problem of vertical wind profile online estimation at a given location. Specifically, the north and east components of the wind are continuously estimated as functions of time and altitude at two waypoints used for landing on the Adolfo Suarez Madrid-Barajas airport. A continuous nowcast of the wind profile is performed in which wind observations are derived from the aircraft states and assimilated into the model. It is well known that wind is one of the utmost contributors to uncertainties in the current and future paradigm of Air Traffic Management. Accurate wind information is key in continuous climb and descent operations, spacing, four dimensional trajectory-based operations, and aircraft performance studies, among others. In this work, wind data are obtained indirectly from the aircraft’s states broadcast by the Mode S and ADS-B aircraft surveillance systems. The Gaussian process regression is adapted to this framework and used to solve the problem. The presented method allows to construct a complete vector wind profile at any specific position that is continuous in time and altitude; namely, there is no need for grid points and time discretisation. The Gaussian process regression is a very flexible estimator which is statistically consistent under general conditions, meaning that it converges to the underground truth when more and more data are dispensed. In addition, the Gaussian process regression approach provides the whole probability distribution of any particular estimation, allowing confidence intervals to be computed naturally. In the case study presented in this paper, in which the wind is constantly estimated, the Gaussian process regression model is iteratively updated every 15 min to capture possible changes in the wind behaviour and give an estimation of the wind profile every half a minute. The method has been validated using a test dataset, achieving a reduction of 50% of the prediction uncertainty in comparison to a baseline model. Moreover, two popular wind profile estimators based on the Kalman filter are also implemented for the sake of comparison. The Kalman filter outperforms the baseline model, but it does not outperform the Gaussian process regression with errors higher by around 35%, in comparison. The obtained results show that the Gaussian process regression of aircraft-derived data reliably nowcast the wind state, which is key in Air Traffic Management.