The use of different ensemble forecasting systems for wind power prediction on a real case in the South of Italy

Abstract

Short-term forecasting applied to wind energy is becoming increasingly important due to the constant growth of this renewable source, whose uncertainty requires a constant effort to meet the needs of the national electrical systems and their operators. Regarding to this, the probabilistic approach applied to wind power forecasting (WPF) is showing an increasingly interest in terms of the possibility to reduce forecast errors, giving also a useful information on the accuracy of a forecast and a reliable estimation of its uncertainty; in fact, the prediction accuracy is not constant and often depends on the location of a certain wind farm, as well as on the atmospheric conditions of the site and the forecast horizon used. According to previous studies of the same authors, the ECMWF Ensemble Prediction System (EPS) can be used as an indicator of a three-days ahead deterministic WPF accuracy. A statistical calibration performed on the wind speed EPS members allows an improvement from an over-confident situation observable from the rank histograms (in which the measurements fell quite always outside the bounds of the probability distribution) to a consistent ensemble spread. After that it is possible to convert the data to wind energy: the spread calculated on wind power can then be used as an accuracy predictor due to its level of correlation with the deterministic WPF error. In this presentation we investigate the performances for both wind power and accuracy prediction of the new EPS used at the ECMWF, whose horizontal resolution was increased on January 2010 from 60 km to 32 km, on a complex terrain area already used in previous studies and located in Southern Italy. The work consists in the use of the ECMWF deterministic model in a WPF approach followed by a recursive feed-forward Neural Networks (NN) and finally by the application and verification of the EPS in order to estimate the forecast accuracy. We also preliminary compare these performances with the results obtainable from the application of other ensemble prediction systems with higher resolution. Analyzing the results it can be seen that EPS calibration is a fundamental requirement in order to extract usable information from data; after an adequate calibration method, the ensemble spread calculated on wind power seems to have enough correlation with the deterministic forecast error in order to be used as a predictor of accuracy, at least until the three days ahead forecast horizon.