Turbulent Wind Flows over Crete: Linear Versus Nonlinear Forecasting Performance

M. Genius and G.P. Tsironis (Greece)


wind speed, prediction, AR, nonlinear


We present a preliminary, empirical analysis of wind data from the island of Crete using a linear autoregressive approach as well as a nonlinear phase space reconstruction approach to forecasting. The specific data we are using are characterized by large and varying turbulence intensity and gustiness that is typical of mountainous terrains in Crete. Although the data are clearly non-stationary, we can isolate contiguous segments where near stationarity is observed enabling thus the use of autoregressive and reconstruction methods. Our analysis with regard to data predictability shows that for short (approximately one to five steps) forecast horizons both autoregressive and reconstructing models give similar forecasts, although an AR that is re-estimated for each data segment used appears to be better. For forecast horizons larger than the first few steps the nonlinear method is generally superior, resulting in systematically lower averaged forecasting errors. The selection of the embedding dimension as well as the delay of the reconstruction is done through optimization based on the minimization of the root-mean square forecast error for the prediction horizon and the maximization of the correlations between data and forecasts.

Important Links:

Go Back