THE ZEPHYR-PROJECTTHE NEXT GENERATION PREDICTION SYSTEM

摘要

High penetration of wind energy like the more than 20% seen in some areas ofEurope are only possible with some kind of short-term forecasting. This paper presents theresults of a three-year research project from Denmark, which improved forecasting accuracyand was finally installed operationally in most major Danish utilities and TSOs.The aim of Zephyr was to integrate the separate tools Prediktor and WPPT (Wind PowerPrediction Tool) from Ris? National Laboratory and the Institute for Mathematical Modellingof the Danish Technical University. Both tools had run operationally for many years inDenmark, and allowed system operators to achieve 27% penetration in Western Denmark.The model built into Zephyr takes input from a numerical weather prediction model(HIRLAM of the Danish Meteorological Institute in the present implementation), from onlinemeter readings from selected wind farms and from monthly reported meter readings froma wider range of wind farms. As a consequence of the diverse input, the model in Zephyr isdual stringed.A new concept for predicting the total wind power production in a larger region wasdeveloped, based on a combination of on-line measurements of power production fromselected wind farms, offline power measurements for nearly all wind turbines in the area andnumerical weather predictions of wind speed and direction. Zephyr employs statisticalmodels to describe the relationship between power production and the numerical weatherpredictions. The statistical models belong to the class of conditional parametric models – amodel class particular useful for estimating non-linear relationships on-line. The estimation isfurthermore made adaptively in order to allow for slow changes in the system eg caused bythe annual variations of the climate.Improvements for the predictions of single farms were made using conditional parametricmethods for the estimation of a power curve. In most cases, the prediction improves slightly,while especially in difficult to predict cases like Spain, the new method was able to improvethe prediction significantly.