@TECHREPORT\{IMM2007-05539,
    author       = "F. Ö. Thordarson and H. Madsen and H. A. Nielsen",
    title        = "Optimal combined wind power forecasts using exogenous variables (PSO2004/FU5766 - Improved wind power prediction)",
    year         = "2007",
    number       = "",
    series       = "IMM-Technical Report-2007-17",
    institution  = "Informatics and Mathematical Modelling, Technical University of Denmark, {DTU}",
    address      = "Richard Petersens Plads, Building 321, {DK-}2800 Kgs. Lyngby",
    type         = "",
    url          = "http://www2.compute.dtu.dk/pubdb/pubs/5539-full.html",
    abstract     = "The aim of combining forecasts is to reduce variation fromobserved values by composite two or more forecasts, which predict for the same event at the same time. Many methods have developed since the problem was presented, ranging from a method of equal weights to more complex methods, e.g. state space methods. Despite this complexity a linear model of the combination appears to be most acquired where the parameters of the forecasts are summing to one. The parameters, also called weights, are unknown and need to be estimated to get optimal combined forecast. In this report the problem of combining forecasts is addressed by (i) estimating weights by local regression and comparing with recursive least squares and minimum variance methods, which are well known procedures within combining, and (ii) using information from meteorological forecasts to estimate the forecast weights with local regression.

The methods are applied to the Klim wind farm using three {WPPT} forecasts based on different weather forecasting systems. It is shown how the prediction is improved when the forecasts are combined by using locally fitted linear model and that it outperforms the {RLS} estimation which is also considered. Furthermore, the meteorological forecasts from {DMI-}HIRLAMare inspected and the air density and the turbulent kinetic energy at pressure level 38 are found to be optimal regressors for locally fitting the weights into of linear combination model.

The results in this report show that using the meteorological information to estimate the weights gives a resonable fit compared to the referencemodels, which can be elevated by further analysis."
}