Time series prediction for local wind characterization

Wind speed prediction could have an important role to play in the context of wind turbine control. The development of an accurate scheme for seconds-ahead predictions could drastically improve the performances of the controller. However, fewer models of this type are developed and most of them are quite basic. In this report, several AI-based models are developed and compared together for different spatial scales. At the rotor scale, ARIMA, MLP and GRU models are developed and evaluated. An autoencoder network is developed for predictions at the scale of a sector of the rotor. Models are trained and evaluated using data from the numerical simulator BigFlow at different levels of turbulence. Data for the training and the evaluation are independent. The report shows the good performances of the autoencoder model and the ability of some AI-based networks to outperform commonly used models. It also analyzes the influence of a data filtering process on the MLP network performance.