Periodic and long range dependent models for high frequency wind speed data

The production of wind power as one source of renewable energy has a huge potential to serve the increasing demand for energy. Therefore, it is necessary to improve the accuracy of wind energy forecasts to increase the energy output. We focus on short-term wind speed forecasts. This article considers a comparison study of two different periodic regression models with autoregressive fractionally integrated moving average errors in the mean part and asymmetric power generalized autoregressive conditional heteroscedasticity within the conditional variance (ARFIMA-APARCH) and two different estimation methods. The first one is a two-step approach where the regression parameters are estimated first with the least-squares method. Hereafter, an ARFIMA-APARCH process is fitted to the stationary residuals. In contrast to this, the second approach estimates the entire model with a QML (quasi-maximum likelihood) estimator in one step. Both models include periodic explanatory variables, but one model lets them vary over time. Moreover, we distinguish between two main distributional assumptions for the residuals. Each model is used to create suitable short-term forecasts up to 6 h. Eventually, we come up with a recommendation which model is preferred.