Detecting non-linear structure and predicting turbulent gusts in surface wind velocities

We present a new approach, based on a locally low-dimensional prediction scheme, which was originally developed for deterministic chaotic signals, to detect the deterministic structure of local surface wind velocities. By exploiting properties of the signal which mimic the deterministic structure created by non-linear coherent effects of a turbulent fluid and by resolving the strong inherent non-stationarity as a sequence of different dynamical regimes, short time predictions of future wind velocities can be performed. Although on average no reduction of the prediction error can be achieved using a non-linear model instead of a linear stochastic one, intermittent gusts can be predicted with significantly higher accuracy. An asymmetry in the predictability is observed, which suggests that an increase of the velocity is rather deterministic, whereas the subsequent decrease is dominated by a stochastic behaviour.