Wind Time Series Modeling for Power Turbine Forecasting

This paper addresses the problem of predicting the average wind speed at different prediction horizons ranging from 6-hours to 1-day based on wind velocity recorded at a point. The problem is relevant in several application fields and recently appears of particular interest for operators of electrical wind turbine plants and/or for optimisation of conventional power plants. Exogenous inputs are not taken into account in this preliminary work, so that the problem is set as pure time series identification and carried out by considering NAR (Non-linear Auto Regressive) models. Therefore model performance are solely related to the degree of autocorrelation of the considered time series and to some extent on the kind of non-linear approximation basis function taken into account. Different data set were considered for training and validation purposes in order to assess the model generalization capabilities. The role of model order was evaluated on the space of representative performance indices. Results show that while the forecasting performances are remarkable for the 6 and 12 hours prediction horizons, they look no so good for the 1-day prediction horizon. Work is still in progress in order to overcome these shortcomings. 