| Popis: |
Very short-term wind farm power forecasts are essential for the Transmission System Operator in order to optimally operate reserves for the continuous balance of the power system, since wind power fluctuations at time scales of some minutes are those wich most seriously affect the balance in the power system. This aspect motivates our choice of 3-hours ahead forecasts and a time series resolution of 10 minutes. For such prediction horizons, it is generally accepted that statistical time series based models are more accurate than physical models. We focus on a non-linear approach based on varying-coefficient models by generalising linear Auto-Regressive with eXogenous input models. The main idea is to replace constant parameters of models with functions that take into account local observations such as wind speed and wind turbines power production and numerical weather predictions of wind speed and direction. To improve accuracy of very short-term wind power forecasts the relationship between the power production for each turbine of the farm are taken into account in order to minimize the impact of topographical particularities of the area, wind turbines interposition within the wind farm, atmospheric processes occurring at different scales, the wake shadowing effect generated by nearby wind turbines, etc. Furthermore, extension from point to probabilistic forecast is given since they are very useful in the sense that they provide us with a measure of the uncertainty associated with a point forecast. Comparison of two different families of varying-coefficient models (regime-switching and conditional parametric models) is carried out. The case of the on-shore wind farm of Danilo in Croatia has been considered. |
