# Conditional weighted combination of wind power forecasts

**ABSTRACT** The classical regression model for combined forecasting was reformulated in order to impose restrictions on the combination weights. This restricted linear combination model was then extended to the case where the weights were allowed to be a non-parametric function of some meteorological variables, yielding the so-called conditional weighted combination method. The weight functions are estimated with local regression techniques. The conditional weighted combination method was applied to a test case of a wind farm over a period of 10 months. Various combinations of two forecasts out of the three available ones were considered. Analysis of the data suggested that meteorological forecasts of air density and turbulent kinetic energy may be considered as relevant external variables in the combination scheme. A performance comparison showed that the conditional weighted combination method introduced in this paper outperformed the least-squares combination method for almost all prediction horizons, especially for larger ones. This indicates that further developments based on conditional combination methods, including adaptivity of the weight functions estimation, may significantly enhance forecast accuracy and dampen the risk of large prediction errors. Copyright © 2010 John Wiley & Sons, Ltd.

**0**Bookmarks

**·**

**109**Views

- [Show abstract] [Hide abstract]

**ABSTRACT:**Following the growing share of wind energy in electric power systems, several wind power forecasting techniques have been reported in the literature in recent years. In this paper, a wind power forecasting strategy composed of a feature selection component and a forecasting engine is proposed. The feature selection component applies an irrelevancy filter and a redundancy filter to the set of candidate inputs. The forecasting engine includes a new enhanced particle swarm optimization component and a hybrid neural network. The proposed wind power forecasting strategy is applied to real-life data from wind power producers in Alberta, Canada and Oklahoma, U.S. The presented numerical results demonstrate the efficiency of the proposed strategy, compared to some other existing wind power forecasting methods.IEEE Transactions on Sustainable Energy 08/2011; · 3.84 Impact Factor - SourceAvailable from: Asifullah Khan[Show abstract] [Hide abstract]

**ABSTRACT:**Depletion of conventional resources has led to the exploration of renewable energy resources. In this regard, wind power is taking significant importance, worldwide. However, to acquire consistent power generation from wind, the expected wind power is required in advance. Consequently, various prediction models have been reported for wind power prediction. However, we observe that Support Vector Regression (SVR), and specially, a hybrid learning model based on SVR offer better performance and generalization compared to multiple linear regression (MLR) and is thus quite suitable for the development of short-term wind power prediction system. To this end, a new methodology ML-STWP namely Machine Learning based Short Term Wind Power Prediction is proposed for short-term wind power prediction. This approach utilizes a combination of machine learning (ML) techniques for feature selection and regression. The proposed methodology is thus a hybrid ML model, which makes use of feature selection through irrelevancy and redundancy filters, and then employs SVR for auxiliary prediction. Finally, the wind power is predicted using enhanced particle swarm optimization and a hybrid neural network. The wind power dataset on which the model is tuned and tested consists of real-time daily values of wind speed, relative humidity, temperature, and wind power. The obtained results demonstrate that the proposed prediction model performs better as compared to the existing methods and demonstrates the efficacy of the proposed intelligent system in accurately predicting wind power on daily basis.Computers & Electrical Engineering 08/2014; · 0.99 Impact Factor

Page 1

Submitted to Wind Energy December 10, 2008.

Conditional Weighted Combination of Wind Power

Forecasts

Fannar¨O. Thordarson*1, Henrik Madsen1, Henrik Aa. Nielsen2, Pierre Pinson1

1DTU Informatics, Technical University of Denmark;

Richard Petersens Plads (bg. 321), DK-2800 Kgs. Lyngby, Denmark.

2ENFOR A/S;

Lyngsø All´ e 3, DK-2970 Hørsholm, Denmark.

Abstract

The classical regression model for combined forecasting is reformulated in order to im-

pose restrictions on the combination weights. This restricted linear combination model

is then extended to the case where the weights are allowed to be a nonparametric func-

tion of some meteorological variables, yielding the so-called conditional weighted com-

bination method. The weight functions are estimated with local regression techniques.

The conditional weighted combination method is applied to a test case of a wind farm

over a period of 10 months. Various combinations of two forecasts out of the three avail-

able ones are considered. Analysis of the data suggests that meteorological forecasts of

air density and turbulent kinetic energy may be considered as relevant external vari-

ables in the combination scheme. A performance comparison shows that the conditional

weighted combination method introduced in this paper outperforms the least-squares

combination method for almost all prediction horizons, especially for larger ones. This

indicates that further developments based on conditional combination methods, includ-

ing adaptivity of the weight functions estimation, may significantly enhance forecast

accuracy and dampen the risk of large prediction errors.

Key words: wind power, forecasting, linear combination, local regression, conditional parametric

models, conditional weighted combination

*Corresponding author:

F. Thordarson,

DTU Informatics, Richard Petersens Plads (bg. 321), DK-2800 Kgs. Lyngby, Denmark.

Tel: +45 4525 3349, fax: +45 4588 2673,

E-mail: ft@imm.dtu.dk, Webpage: www.imm.dtu.dk

Page 2

Introduction

The wind power industry is growing rapidly, while the share of wind generation in the

electricity mix of a large number of countries is getting larger every year. Research and de-

velopment have improved the technology, with advances in performance as well as reduced

production costs, and made wind energy competitive on the energy market. Expectations in

the medium-term are that 10-20 percent of world electricity consumption will be generated

by wind power [1]. However, the variability and low-predictability of wind generation are

commonly seen as penalizing, since they translate to e.g. an increase in reserve require-

ments for Transmission System Operators (TSOs), or to a diminution of potential gains on

the electricity market for the case of utilities or independent wind power producers. This

makes forecasting wind power generation for time-scales ranging from a few minutes to

a few days ahead paramount. Forecasting tools are helping to secure the value of wind

generation on the market, while contributing to increasing security of supply. A review of

state of the art wind power forecasting has been published by Giebel et al. [2].

The loss function of wind power forecast users, i.e. the mathematical representation of

their sensitivity to consequences of forecast errors, is usually not linear. This translates

to saying that their sensitivity to forecast errors changes with their range of magnitude.

Indeed, errors perceived as small are easily accepted, while forecasters will always be re-

minded about their errors of very large magnitude and that yielded significant financial

or technical consequences. Today, users of wind power predictions are increasingly pro-

vided with several forecasts for the coming period, generated from various methodologies

and/or based on different meteorological input. In such a case, it can be attractive proce-

dure to combine these forecasts. By combining the independent information included in

each individual forecast, more accurate prediction can be accomplished, and the risk of

large prediction errors significantly reduced [3]. This was originally suggested by Bates &

Granger [4] for two individual forecasts, and since then a number of procedures have been

suggested, see e.g. Clemen [5] or de Menezes et al. [6] for a literature review.

Combining wind power forecasts is especially appealing if several meteorological fore-

casts from different meteorological offices are available. Since meteorological forecasts

directly input wind power prediction methods (for look-ahead times ranging from 6 to 72

hours ahead), employing different meteorological forecasts will yield power forecasts with

different characteristics. Combination is traditionally performed through a linear regres-

sion model, for which the parameters represent the weights of corresponding individual

forecasts in the combination. Common linear methods for forecast combination include

the simple plain average [7], the Bayesian approach [8] and regression-based methods [9].

Initially the present paper employs a linear model, with restrictions on the combination

weights. The main originality of our approach is that the parameters in the linear model

(i.e. the combination weights) are further allowed to be dependent upon some external

variables, which then leads to a nonlinear model. In other words, the combination scheme

can be made conditional to the predicted meteorological conditions, if summarized by a set

of predicted meteorological variables. The block diagram in Figure 1 illustrates the pro-

posed approach to forecast combination, and more specifically how information from me-

teorological forecasts impact weight estimation. Owing to its specific characteristics, our

newly introduced combination scheme is referred to as Conditional Weighted Combination

(CWC). This proposal comprises an alternative approach to the situation-specific combina-

1

Page 3

tion approach introduced by Lange et al. [10] and Lange & Focken [11], for which forecast

combination is proposed to be made dependent upon the prevailing synoptic weather sit-

uation. Classification of weather situations in a number of weather ’modes’ is performed

through a method combining principal component analysis and cluster analyses. Within

each weather class the accuracy of various prediction models is evaluated and an optimal

combination determined. Such a method may be cumbersome, as it requires analysis of a

huge amount of data in order to determine prevailing weather situations. Another draw-

back is that such a method involves the possibly demanding task of predicting, to which

weather class the coming weather situation belongs to. In contrast, the situation-specific

combination method originally introduced here only necessitates a limited quantity of fore-

cast information at the level of the wind power production site considered.

————— Figure 1 is approx. here —————

The paper begins with a description of the mathematical framework for forecast combi-

nation by considering combination as a restricted linear model with the individual fore-

casts as explanatory variables. Then the CWC scheme is introduced by replacing the

weights in the linear model with smooth, but otherwise unknown, functions of some me-

teorological variables. A method for estimating such functions is briefly presented. These

methods are then evaluated and compared in a least-squares minimization fashion [9].

The test case is that of real wind farm located in the north of Jutland (Denmark), for

which three different types of power forecasts are available (with three different meteoro-

logical forecast inputs), along with corresponding wind power measurements for a period

of 10 months. The main evaluation criterion is the Root Mean Square Error (RMSE),

which gives more weight to large prediction errors. Its use is thus consistent with the idea

that forecast users may be more sensitive to consequences of large forecast errors [12]. In

parallel, employing the RMSE as an error measure is also consistent with the idea that

wind power point forecasts commonly inform on the mean potential generation for each

look-ahead time, or in more mathematical terms, give the conditional expectation of the

wind power generation random variable for every forecast horizon. The expectation is the

summary statistics of a random variable that minimizes a quadratic loss function, while

the RMSE is the corresponding quadratic error measure. The paper concludes with some

discussion and perspectives regarding further developments related to conditional combi-

nation schemes.

Mathematical Framework for Combining Wind Power Fore-

casts

Since the issue of combining forecasts was first introduced, many methods have been de-

veloped, sometimes even based on advanced complex techniques. However, it is still not

obvious which combination method will always perform best [6]. In this paper, the linear

combination model is adopted first because of its inherent simplicity and its performance

often deemed as highly acceptable. Later, however, an extension is proposed in order to

allow for the coefficients to be functions of some exogenous variables, and this makes the

model nonlinear.

2

Page 4

Formulation of a Restricted Linear Combination Model

For the wind power application, the variable of interest to be predicted is the wind power

production at time t+h, which is denoted as yt+h. In parallel, let ? yi,t+h|tbe the ithindividual

forecast for the wind power at time t+h (out of the k different forecasts available)calculated

using the actual information available at time t, with h being the forecast horizon, i.e. the

number of hours between the moment at which the forecast is issued (time t) and the time

point for which the prediction is made. The ithprediction error, that is, the difference

between the measured power production and the ithcompeting prediction is ei,h. For the

combination of k individual forecasts, a linear model of the forecasts is formulated as

? yc,t+h|t= µh+

k

?

i=1

wi,h? yi,t+h|t

(1)

where wi,h= wiis the weight given to the ithof the k forecasts. In the following, the forecast

horizon h is omitted from the weight terms of the model in order to simplify notations.

In the present study, a linear model is written for each horizon independently, and the

horizon of interest can easily be deduced from notations employed for the various forecast

variables. The term µ represents the constant term in the linear model, which is provided

if the individual forecasts are biased, especially when the coefficients for the individual

forecasts are restricted to sum to one [13, 14]. Considering an additional noise in the

power production process, the combined forecast corresponds to the conditional expectation

of power production at time t + h, given the information available at time t,

yt+h= ? yc,t+h|t+ ec,t+h

where {ec,t+h} is a noise sequence, which in the present case consists of a set of identically

distributed realizations of a random variable with finite variance.

(2)

The parameters of interest in (1) are the weights which indicate the influence of each

individualforecast in the combined forecast. The weights in the model are restricted to sum

to one, since each weight defines the fraction of information provided by each constituent

prediction in the combination, i.e.

k

?

i=1

wi= 1

(3)

By applying this restriction alone, negative weights may be obtained as there is no con-

straint on a minimum value for each weight. Here, weights are allowed to be negative and

no consideration is given to a potential lower limit, although, according to Gunter [15],

a non-negativity restriction may be assumed to be more effective and robust, in addition

to being more easily interpretable from a physical point of view. The constant term in the

linear combination model is included to detect and correct for any global discrepancy in the

combination, i.e. any systematic error still remaining after merging the forecast informa-

tion provided by each individual prediction. Any potential bias in the individual forecasts

is embraced in the constant term, with the same estimated weight as for the corresponding

individual forecast. This is then equivalent to correcting the bias for the combined forecast

as a weighted sum of bias terms originating from the individual forecasts.

3

Page 5

Inclusion of the additive restriction can be done in a straightforward manner: the con-

straint is added into the linear combination model (1) for the kthweight (by expressing this

weight as 1 minus the others, i.e. wk= 1−?k−1

This is a quite general manner to extend a combination scheme to more than two individ-

ual forecasts, as discussed in Nielsen et al. [16]. This yields a new formulation in the form

of a restricted linear combination model,

i=1wi), and the kthforecast is then subtracted.

y∗

t+h= µ +

k−1

?

i=1

wi? y∗

i,t+h|t+ ec,t+h

(4)

where ? y∗

ec,t+hnor the intercept µ are affected by this modification. The weights are not only con-

sistently estimated, but the restricted model in (4) also gives forecast errors for the linear

combination model in (1) and the corresponding intercept.

i,t+h|t= ? yi,t+h|t− ? yk,t+h|tand y∗

t+h= yt+h− ? yk,t+h|t. Neither the prediction error

The method applied in the following to estimate the weight coefficients of the restricted

linear model (4) is the least squares method [17].

The Least-Squares Method in a more General Regression Framework

By integrating the restriction into the most general form for forecast combination, i.e. in a

regression model, and considering the weights to be time-varying, we obtain

y∗

t+h= g(? y∗

t+h|t,t;wt+h) is a known mathematical function of the independent individual

forecasts ? y∗

are unknown. To properly estimate the constant µ in the following study is ? y∗

error ec,t+his as defined in the combination model (2). The general linear model is a special

case of the regression model where the response is defined as a linear function of individual

forecasts, i.e.

t+h|t,t;wt+h) + ec,t+h

(5)

where g(? y∗

t+h|t= (? y∗

0, ? y∗

1,t+h|t,..., ? y∗

k−1,t+h|t)⊤, but the weights wt+h= (µ,w1,t+h,...,wk−1,t+h)⊤

0= 1 kW. The

y∗

t+h= w⊤

t+h? y∗

t+h|t+ ec,t+h

(6)

In order to estimate the weights in (6), it can be chosen to minimize a loss function L,

where L represents the cost of deviation between combined forecast and measurement.

For the combined forecast to be competitive with the individual forecasts, its loss function

value has to be less than or equal to those of the individual forecasts, i.e. L(ec,t+h) ≤

mini{L(ei,t+h)}. The loss function that is the most commonly chosen for weight estimation

in (6) is the Least Squares (LS), where L(ec,t+h) = E[(ec,t+h)2].

The data set is used to estimate the unknown parameters, thus the design matrix is

a matrix of the constituent forecasts, including all values in the data set (?Y ), and the

measurements form a vector of observed values (y). The solution to the problem is an

estimator of the weights that minimizes the sum of squared residuals, i.e.

? w = (?Y⊤?Y )−1?Y⊤y

where the matrix?Y⊤?Y is assumed to be invertible [18].

(7)

4

Page 6

Conditional Weighted Combination

Let us consider a class of models that are linear to some regressors, and for which the

coefficients are assumed to be changing smoothly as an unknown function of some other

variables. These kinds of models are called varying-coefficient models [19], but when all

coefficients depend on the same variable, the model is referred to as a conditional para-

metric model. The formulation is related to the procedure of locally weighted regression

[20, 21, 22], where the function g in the regression model (5) is now a smooth function, esti-

mated by fitting a polynomial of the dependent variables to the response. By extending the

weights in the linear combination model in (6) to be function of some external variables, a

new method of combining forecasts is introduced, the so-called Conditional Weighted Com-

bination (CWC). In the following analysis the weight restriction in (3) is implemented in

the CWC model. Notations similar to those employed above are also used here. However,

the mathematical framework of CWC may be seen as generic and employed for forecast

combination in other applications than wind power prediction.

When using a conditional parametric model to describe the response y∗

variables are divided in two groups. In the case of combining forecasts, the group of indi-

vidual forecasts enter globally as in the linear regression model, while the weights depend

on some other group of variables, ut+h. In practise will ut+hdenote a group of forecasted

variables. The combination model in (6) hence becomes

t, the explanatory

y∗

t+h= w(ut+h)⊤? y∗

where w(ut+h) is a vector of coefficient functions to be estimated at time t + h, and ec,t+h

is as described in (2). The dimension of ? y∗

poses the dimension of ut+hmust be low [23]. The weight functions are estimated at a

number of distinct fitting points by locally approximating these functions using polyno-

mials, and then fitting the resulting linear model locally to each of these points. Let u

denote a particular fitting point and pd(ut+h) be a column vector of terms in the corre-

sponding dthorder polynomial evaluated at u. The order of the polynomial can be different

for the elements in the vector w(ut+h). Furthermore, we have the k−1 individual forecasts,

? y∗

z⊤

t+h=

? y∗

φ⊤(u) =

φ⊤

k−1(u)

t+h|t+ ec,t+h,

(8)

t+h|tcan be quite large, but for practical pur-

t+h|t= [? y∗

1,t+h|t,··· , ? y∗

?

?

k−1,t+h|t]⊤, then

1,t+h|tp⊤

d(ut+h),··· , ? y∗

1(u),··· ,φ⊤

k−1,t+h|tp⊤

?

d(ut+h)

?

,

(9)

(10)

where the terms in φ(u) are column vectors of local coefficients at u, corresponding to the

terms in z⊤

t+hφ(u) + ec,t+his then fitted locally at u using

weighted least squares:

t+h. The linear model y∗

t+h= z⊤

?φ(u) = argmin

φ(u)

N

?

t=1

βu(ut)

?

y∗

t− z⊤

tφ(u)

?2

(11)

with N being the number of available observations [20]. The weights βu(ut) are assigned

by using the tri-cube function [21, 23].

5

Page 7

The ithelement of the weighting vector w(u) is then estimated as

? wi(u) = p⊤

where?φi(u) is the weighted least squares estimate of φi(u). The procedure described above

is identified as an offline procedure, since a given set of data is used for the estimation.

The coefficients depend on the values of the meteorological forecasts exclusively where

no time-variation is considered. When considering potential long-term variations in the

coefficients, it is possible to track adaptively such coefficients with methods similar to that

proposed in Nielsen et al. [24]. This possibility is not, however, considered in this paper.

d(u)?φi(u)

(12)

Application Results

So far the mathematical framework for combining forecasts has been introduced, with the

introduction of a new approach allowing for conditional weighted combination (CWC) of

wind power production forecasts. In the following, CWC’s of two individual forecasts are

considered, for which the external variables are selected from the available predicted vari-

ables of one considered meteorological forecasting system. The test case and available data

are introduced in a first stage, followed by a detailed description of the study performed

to select the external variables to be used in the CWC scheme, and to identify the model.

Actually, the study on forecast combination of wind power predictions detailed in Nielsen

et al. [16] is based on the same data. In that study the combination weighted is not con-

ditional, but depends on the bias of the individual forecasts and the variance-covariance

matrix for the individual forecast errors.

Description of the Case Study

Power predictions are produced using WPPT (Wind Power Prediction Tool [25]) for the

Klim wind farm, which is a 21MW wind farm located in north-western Jutland, Denmark.

The data is available in two separate data sets: one data set includes the wind power

predictions and corresponding wind power measurements, and the second data set consists

of meteorological forecasts to be used as the external variables in the CWC procedure, cf.

Figure 1. A more detailed description of the data is given below:

1. Using three meteorological forecasts, WPPT is used to generate a forecast of the wind

power for each meteorological forecast. In the first data set the three wind power

forecasts are listed along with the corresponding measured power production. The

meteorological forecasts for WPPT are provided from three different global meteo-

rological systems. First are predictions from the meteorological system Deutscher

Wetterdienst, second are predictions from the HIRLAM model of the Danish Mete-

orological Institute [26], third there are predictions from WPPT when using meteo-

rological forecasts based on MM5 modelling [27] and GFS (Global Forecast System)

from the National Centers for Environmental Predictions, in the USA. Furthermore,

the MM5 forecasts were prepared for WPPT by CENER, the National Renewable En-

ergies Centre, in Spain, see for instance Villanueva and Mart´ ı [28]. In the following

study the individual wind power forecasts will be abbreviated DWD, HIR and MM5,

6

Page 8

respectively. Even though the three different sets of forecasts are based on different

meteorological data, they are highly correlated.

2. The second data set consists of meteorological forecasts from the Danish Meteorolog-

ical Institute forecasting system HIRLAM. The objective is to generate a conditional

weighted model for the combined forecast, where the weights depend on one or sev-

eral of the available forecast variables. For each of four different pressure levels in

the atmosphere, the variables included are wind speed, wind direction and turbulent

kinetic energy, along with air density, friction velocity and radiation. The wind speed

and wind direction at 10 meters above ground level also available as forecast vari-

ables, but are already used as input in WPPT for providing the power predictions

for HIRLAM and, therefore, of less importance in the analysis. Correlation is strong

among wind speed variables, and for the various wind direction and turbulent kinetic

energy variables. As a consequence, only one variable in each group is considered in

the study.

Both data sets contain forecasts provided at midnight, denoted as 00Z in the figures, with

an hourly temporal resolution up to 24 hours ahead. All variables span the period February

2nd2003 to December 2nd2003, which in total gives 7272 data points for each variable in

the data sets.

The individualwind power forecasts are investigated to see if the conclusion drawn from

the combination improves the individual predictions for certain. The RMSE values using

the competing meteorological forecasts for horizons up to 24 hours are depicted in Figure

2a. It is clear that the RMSE for all three cases increases with larger prediction horizons.

MM5 is the least accurate forecast, while HIR is the best prediction. DWD appears to be

somewhere in between, though closer to HIR. From the figure it could be assumed that

combining the two best performing forecasts would result in the most adequate combina-

tion. Figure 2b shows that the correlation between the individual forecasts is rather high,

around 0.6. Further, it can be seen that the correlation with the HIR forecast is roughly the

same for all prediction horizons, while the lowest correlation appears to be that between

the DWD and MM5 wind power forecasts.

————— Figure 2 is approx. here —————

Results and Discussion

With the available data for this study, the weights in the combined forecasts are consid-

ered as weight functions being dependent on possibly any of the two available meteorolog-

ical forecasts, then serving as explanatory variables [29]. Selection in Thordarson [29] is

performed through data analysis with coplots, employed in combination with conditionally

parametric fits [30]. By doing so, any relation of a weight in the combined forecast with any

two meteorological forecast variables can simultaneously be studied and demonstrated. An

available meteorological forecast is selected and partitioned in several bins, and for each

of the bins the weights in the combination are smoothed regarding some other available

meteorological forecast (with a conditionally parametric fit). The study reveals that of the

7

Page 9

available forecasts, the two most adequately influencing weights in the combination are

air density and turbulent kinetic energy.

In a first stage, Figure 3 depicts coplots with the DWD weights as the response variable

(w1), while air density (uad) and turbulent kinetic energy (utke) are seen as the explanatory

variables. For both Figures 3a and 3b the partitioned variable has its lowest values in

the lower left panel and, first, increases to the right and, then, upwards. The partitioning

is performed so that there is the same number of data points in each of the bins. For

each fraction of the partitioned variable, the panels show the distribution of a combination

weight as a function of the second external variable. The fitted line in each panel is the

local regression for the cloud of data points.

In Figure 3a utkeis partitioned. This figure shows how the slope of the local fit between

weights and uaddecreases with higher values in utke, as well as the variance of the weights.

For low values in utke(bottom row of panels) little or no change in slope occurs, but that

is due to the density of the variable for low values. When the turbulent kinetic energy

increases, first the intermediate panels show some shifting in slope of the local fit, around

uad=1270, and then the slope approaches zero as seen in the top row of Figure 3a. The dis-

tribution of the points in each panel also shows how the response spreads with increasing

uad, but the distribution is reduced with increasing utke. By partitioning the air density

and fitting the weights to the 4throot of utke(Figure 3b) the local regression appears to be

constant for almost all panels. Closer inspection of the fit reveals that the smoothed line

shifts upwards with increasing air density. The top row of panels also show that the fit

has a negative slope before it forms a concave curvature. The data points in the panels

verify the distribution of the weights with increasing air density, while the variance of w1

is constant with changes in utkewithin the panels.

————— Figure 3 is approx. here —————

In conclusion there seems to be coherency between forecasted values of uadand utke,

and the DWD weights. The uadand utkeforecasts from DWD are therefore used as external

variables for the CWC model. The combination forecast is then expressed as

? yc,t+h|t= µ(uad,utke) + w1(uad,utke)? y1,t+h|t+ w2(uad,utke)? y2,t+h|t

Using the constraint for forecast coefficients as introduced in (4), the above equation can

be rewritten as

(13)

? y∗

c,t+h|t= µ(uad,utke) + w1(uad,utke)? y∗

The scatterplot of uadand utkemakes a basis for the weight estimation in (13) but ob-

serving the uad-utkeplane in Figure 4a reveals that the meteorological data does not cover

the whole plane. The utkeforecasts are more densely distributed for low values, while the

uadforecasts are closer to being Gaussian distributed. This implies that for high values

of utkeand either low or high values of uad, very few values exist. The area where the

weights have some valid estimation on the uad-utkeplane is thus defined as the area below

the dashed lines in Figure 4a.

1,t+h|t

(14)

Figure 4b shows the time series for the two DWD meteorological forecast variables of

8

Page 10

interest. Air density is known to be inversely related to air temperature. Figure 4b is

indeed an illustration of this effect, since air density is high through the colder months of

the year but decreases during the summer period. In contrast, turbulent kinetic energy

is not found to be correlated with any of the other available meteorological variables. It

varies through the entire period, although less at both tails of the data set, indicating

reduced variation over the winter period. Turbulent kinetic energy is a variable used to

study turbulence and its evolution in the atmospheric boundary layer. When the boundary

layer becomes stable, turbulent kinetic energy is suppressed. The time series plot shows

that layers of air in the atmosphere are more stable over the winter period.

————— Figure 4 is approx. here —————

Weights for all three possible combinations of two individual forecasts, are examined on

the uad-utkeplane. Only one weight from each combination is displayed since the weights

are restricted to sum to one. The local polynomials for the two weight functions are consid-

ered to be of zero and of first order for µ and w1, respectively. From (12), the local models

for the weight functions are then

µ(uad,utke) = µ,

(15)

w1(uad,utke) = w10+ w11uad+ w12utke

(16)

where the parameters are estimated with the weighted least squares method in (11). In-

serting the local models directly into the restricted linear combination model in (14) would

then make our combination model equivalent to the following linear model

? y∗

c,t+h|t= µ + w10? y∗

Combining wind power forecasts as in (17) indicates that the weights are linearly de-

pendent on the forecasts of the two meteorological variables considered, i.e. air density

and turbulent kinetic energy. If such an equivalent linear model were then considered,

the parameters µ, w10, w11and w12would be estimated as constant terms, independent

from the meteorological variables. In contrast, for our proposed combination scheme, the

weights are allowed to be unknown general functions of the meteorological variables, and

by smoothing the weights over the uad-utkeplane, the surface plots in Figure 5 are obtained.

Since we only consider the combination of 2 forecasts, and the weights are constrained to

sum to one, the weight surfaces in this figure represent the relative importance given to

either one or the other constituent wind power forecast, depending on the forecast value of

air density and turbulent kinetic energy. Figure 5 shows that the weights (w1) on DWD,

when combined with HIR or MM5, appear to have similar surfaces. For low values of

utke, the DWD weights become larger as uadincreases, but for higher values of turbulent

kinetic energy the weights are reduced with progressing air density. Hence this defines

zones where wind power predictions from DWD meteorological forecasts are more or less

to be trusted in comparison with those based on HIR and MM5 meteorological forecasts.

In parallel, the behavior of the HIR weight when combined with MM5 is more challeng-

ing to interpret. There is some fluctuation for the low values of utke, but with increasing

turbulent kinetic energy the surfaces’ variation is reduced. It can also be seen that the

1,t+h|t+ w11uad? y∗

1,t+h|t+ w12utke? y∗

1,t+h|t

(17)

9

Page 11

share of the HIR forecasts is higher in the combination for larger values of turbulent ki-

netic energy. Finally, the surfaces for the intercepts are similar where low utkeimplies high

negative value for the intercept, but with increasing utkethe intercepts increase as well.

The intercepts depending on uadshow some kind of bell-shaped structure where the high

and low values of air density imply low value for the intercept, but around mean uadthe

intercept is at its maximum.

————— Figure 5 is approx. here —————

Considering two meteorological forecasts to fit the weights in the CWC model gives

rather smooth surfaces, especially for the weight in the DWD/HIR combination (Figure

5b). For minimum turbulent kinetic energy, the weight appears to increases linearly with

increase in air density, with some relative slope and intercept. However, with increasing

utkethis linear relation is maintained, but the slope is reduced and the intercept is shifted

upwards. Thus, to simplify the CWC model further the shape of the surface for w1in the

DWD/HIR combination is merged into the model. From now on, let us only focus on the

DWD weight in the DWD/HIR combination in order to study how the combination scheme

could be further enhanced.

From Figure 5b it can be concluded that there is a linear relationship between the

weight and airdensity, where both intercept and slope are functions of the turbulent kinetic

energy, i.e.

w1(uad,utke) = v10(utke) + v11(utke)uad,

(18)

where v10and v11are the intercept and the slope, respectively. The plot shows that for

increase in utkethe intercept increases as well, but the slope decreases and becomes neg-

ative for utkehigher than 8000. Figure 5a is a surface plot of corresponding intercept in

(14), but it shows a wave-like behavior for low values of utkearound the mean value of uad,

which is challenging to interpret. However, the influence of the variance of the intercept

can be appraised by comparing the terms of the CWC model in (14), i.e. the variance of the

intercept is compared to the product of the forecast weight and the forecast. A comparison

of the diagonal terms in the covariance matrix for the product terms in the model shows

that the second term in (14) is much greater than the first term. This indicates that the

intercept term does not have much influence in the combination [29].

By omitting the deep valley on the surface of the intercept, the relationship between

the weight and the air density appears to be a bell-shaped function which fades out with

increasing utke. Such a function is difficult to formulate, and its implementation in the

model would give a complicated interpretation. The naive assumption is that uaddoes not

affect the intercept but the intercept is a function of utke, such as

µ(uad,utke) = v0(utke).

(19)

This assumption might be a bit crude but will make it easier to integrate the proposed

10

Page 12

functions from (18) and (19) into the conditional weighted model. This finally gives

? y∗

c,t+h|t= v0(utke) + [v10(utke) + v11(utke)uad] ? y∗

= v0(utke) + v10(utke)? y∗

= v0(utke) + v10(utke)z1+ v11(utke)z2

1,t+h|t

1,t+h|t+ v11(utke)? y∗

1,t+h|tuad

(20)

where z1 = ? y∗

coefficients only depend on one unknown variable instead of two, namely the turbulent

kinetic energy. Figure 6 shows how the weights in (20) change with utke. The same valley

appears for the intercept, i.e. as in Figure 5a, and the same test is performed as before to

estimate sufficiency of the variance of the intercept. The diagonal elements of the matrix

reveal that the variance of the intercept is only a fraction of the other variances, and,

therefore, the valley at utkeof around 3000 can be neglected. The opposite behavior of the

parameters v10and v11is not surprising since they form the weight in the original model

(14). From Figure 6 it can be assumed that the parameters, v10and v11, are linear functions

of utkewhich changes slope when utkeis approximately 9000.

1,t+h|tand z2 = ? y∗

1,t+h|tuad. The model in (20) is a CWC model where the

————— Figure 6 is approx. here —————

In a last stage, the quality of the forecasts obtained from the combination methods

is investigated using the RMSE criterion. For a discussion on error measures in wind

power forecasting, we refer to Madsen et al. [12]. First; if comparing the forecast accuracy

of combined forecasts with that of the individual forecasts in Figure 2a, it is clear that

forecast combination leads to a significant decrease in RMSE. Notice that combinations

involving the least accurate forecasts (that is, MM5) are as good as those involving the most

accurate ones (DWD and HIR). This shows that one should not just pick the most accurate

predictions as input for combination, since their correlation may also have an impact on

the quality of the combined forecasts. This point is further discussed in Nielsen et al.

[16]. Second; in Figure 7a the CWC forecasts are compared to the LS-combination, for the

three combinations of the two forecasts considered. For the shorter prediction horizons the

CWC method performs similar to the LS method. However, for the intermediate and larger

prediction horizons the CWC method outperforms the LS predictions. The improvement of

applying the CWC is displayed in Figure 7b in percentages. It shows how the LS model

is performing better for the first five to nine horizons, but for the intermediate horizons

CWC outperforms LS by up six percent for the HIR/MM5 combination and four percent

for DWD/HIR and DWD/MM5. For the larger horizons the improvement is reduced for all

combinations, but still in favor of the CWC method. On average, the performance of CWC is

1% better then the LS combination, but considering only prediction horizons larger than 8

hours ahead this improvement is closer to 2%. By looking at each combination, the average

prediction horizon in DWD/HIR, DWD/MM5 and HIR/MM5 shows improvement of 0.9%,

2.1% and 2.5%, respectively. The performance of the alternative CWC proposed in (20) (i.e.

weights depending on only one external variable, but with an adapted model structure)

for the specific case of the DWD/HIR combination is also depicted in Figure 7 (denoted as

CWC.1 in the legend of the Figure). It appears to be approaching the performance of the

LS combination method (in terms of NRMSE) with a slight improvement.

11

Page 13

————— Figure 7 is approx. here —————

The results show that a first impression of the introduced CWC model is capable of

improving a linear model for combining forecasts in all three combinations. The advantage

of the new method over the classical approach is obvious, even though the results appear

to be marginal. Conditioning the weights on the meteorological forecasts establish a new

scheme for combining forecasts that can easily be extended, such that further improvement

is achieved.

Conclusions

The combination of wind power forecasts, when several meteorological forecasts are avail-

able, is known to lead to an increase in forecast accuracy. This has also been confirmed

by the analysis in the present paper. Various methods for combining forecasts have been

presented with the classical linear regression model as a basis. This particular model has

been thoroughly described, in its form including a constant term and a weight restriction,

and for both offline and online cases. It has been shown that not considering a constant

term to account for bias in the combination model may lead to a loss of forecast accuracy.

Therefore, our advice for future developments related to combining wind power forecasts

is to always account for this bias term.

In addition, an original proposal in this paper is to replace the constant weights tra-

ditionally found in combination schemes with weights that are nonparametric functions

of some external variables. This new combination scheme is referred to as Conditional

Weighted Combination (CWC). The weight functions are estimated using local polynomial

regression. For wind power prediction, the combination scheme can then be made con-

ditional on the evolution of some meteorological variables. This goes along with an idea

originally proposed in Lange et al. [10] which consists of a forecast combination, condi-

tioned to the prevailing weather situation. Here, our alternative proposal is to consider

that forecasts of certain meteorological variables may directly be used to determine the

weights in the combination of several wind power predictions.

A study of available forecasts of meteorological variables showed that forecasts of air

density and turbulent kinetic energy are relevant candidates for conditioning forecast com-

binations. It is not possible to generalize application of these two variables to other wind

farms, located in different terrains and subject to different wind climatologies. However,

the proposed methodology can be employed for other wind farms to determine which fore-

casts of meteorological variables may be considered for conditioning a wind power forecast

combination scheme. As a consequence, perspectives regarding future work must obviously

include applications of the proposed conditional weighted combination scheme to other test

cases, i.e. to wind farms located in regions with different climates, and for which a variety

of meteorological forecasts are used as input to wind power prediction. This may permit

validation of the importance of employing air density and turbulent kinetic energy as ex-

ternal variables in the combination scheme in general. In parallel, this may reveal the

interest of considering different external variables depending on local wind climatologies,

or on the type of meteorological forecasts employed.

From a mathematical point of view, the method proposed to estimate the weight func-

12

Page 14

tions in the CWC model should be made (time-) adaptive, by employing some recursive

formulation of the problem, and exponential forgetting of past observations. A method

similar to that described in Nielsen et al. [24] could be applied for this purpose. This

would certainly further improve the forecast accuracy of the combined forecasts. Finally,

emphasis should be put on providing uncertainty information along with combined fore-

casts. Even if forecast uncertainty is reduced by employing forecast combination schemes,

the characteristics of the uncertainty in the combined forecasts obtained is also becoming

more complex. This is due to the fact that overall uncertainty is then driven by an evolv-

ing weighting of the situation-specific uncertainty of each individual wind power forecasts

used as input. Relevant nonparametric methods taking the aspects into account will be

the focus of further studies.

Acknowledgements

This study was sponsored by the Danish Public Service Obligation (PSO) fund FU5766,

and we would like to express out gratitude for this sponsorship. We are also grateful to the

Danish Meteorological Institute for providing the meteorological data set used in the study.

The HIRLAM data was kindly provided by Elsam Kraft A/S (now part of DONG Energy),

the MM5 data was prepared by CENER, the Spanish National Renewable Energies Centre,

with data originally provided by the National Center for Environmental Prediction (USA),

and the DWD forecasts were obtained commercially via financial support from the Danish

PSO fund under contract FU2101.

References

[1] Thor S-E, Weis-Taylor P. Long-term research and development needs for wind energy for the

time frame 2000-2020. Wind Energy 2002; 5:73–75.

[2] Giebel G, Kariniotakis G, Brownsword R. The state-of-the-art in short-term prediction of

wind power. Technical Report, Anemos Project Deliverable Report D1.1, 2003. Available on-

line: http://anemos.cma.fr (accessed March 1st 2008).

[3] Hibon M, Evgeniou T. To combine or not to combine: selecting among forecasts and their

combinations. International Journal of Forecasting 2005; 21:15–24.

[4] Bates JM, Granger CWJ. The combination of forecasts. Operational Research Quarterly 1969;

20:451–468.

[5] Clemen RT. Combining forecasts: a review and annotated bibliography. International Journal

of Forecasting 1989; 5:559–589.

[6] de Menezes LM, Bunn DW, Taylor JW. Review of guidelines for the use of combined forecasts.

European Journal of Operational Research 2000; 120:190–204.

[7] Gunter SI. Theoretical justification of the efficiency of simple average combination. Working

paper, Temple University, Philadelphia, PA; 1990.

[8] Bunn DW. A bayesian approach to the linear combination of forecasts. Operational Research

Quarterly (1970-1977) 1975; 26:325–329.

[9] S´ anchez I. Adaptive combination of forecast with application to wind energy forecast. Interna-

tional Journal of Forecasting 2008; 24:679–693.

[10] Lange M, Focken U, Meyer R, Denhardt M, Ernst B, Berster F. Optimal combination of differ-

ent numerical weather models for improved wind power predictions. In: Proc. of the 6th Inter-

13

Page 15

national Workshop on Large-Scale Integration of Wind Power and Transmission Networks for

Offshore Wind Farms, Delft, The Netherlands, 2006.

[11] Lange M, Focken U. Physical Approach to Short-Term Wind Power Prediction. Springer Verlag:

Berlin, 2005.

[12] Madsen H, Pinson P, Kariniotakis G, Nielsen HAa, Nielsen TS. Standardizing the Performance

evaluation of short-term wind power prediction models. Wind Engineering 2005; 29:475–489.

[13] Granger CWJ, Ramanathan R. Improved methods of combining forecasts. Journal of Forecast-

ing 1984; 3:197–204.

[14] Holden K, Peel DA, Thomson JL. Economic Forecasting: An Introduction. Cambrigde Univer-

sity Press: Cambridge, 1990.

[15] Gunter SI. Non-negativity restricted least squares combinations. International Journal of

Forecasting 1992; 8:45–59.

[16] Nielsen HAa, Nielsen TS, Madsen H, San Isidro Pindado MJ, Mart´ ı Perez I. Optimal combi-

nation of wind power forecasts. Wind Energy 2007; 10:471–482.

[17] Madsen H. Time Series Analysis. Chapman & Hall/CRC: London, 2007.

[18] Basilevsky A. Applied Matrix Algebra in the Statistical Sciences. Dover Publications, Inc: Mi-

neola, New York, 2005.

[19] Hastie T, Tibshirani R. Varying coefficient models. Journal of the Royal Statistical Society.

Series B 1993; 55(4):757–796.

[20] Cleveland WS. Robust locally weighted regression and smoothing scatterplots. Journal of the

American Statistical Association 1979; 74:829–836.

[21] Cleveland WS, Devlin SJ. Locally weighted regression: an approach to regression analysis by

local fitting. Journal of the American Statistical Association 1988; 83:596–610.

[22] Hastie T, Loader C. Local regression: automatic kernel carpentry. Statistical Science 1993;

8:120–129.

[23] Hastie T, Tibshirani RJ. Generalized Additive Models. Chapman & Hall: London, 1990.

[24] Nielsen HAa, Nielsen TS, Joensen AK, Madsen H, Holst J. Tracking time-varying-coefficient

functions. International Journal of Adaptive Control and Signal Processing 2000; 14:813–828.

[25] Madsen H, Nielsen HAa, Nielsen TS. A tool for predicting the wind power production of off-

shore wind plants. In: Proc. of the Copenhagen Offshore Wind Conference & Exhibition, Copen-

hagen, Denmark, 2005.

[26] Feddersen H, Sattler K. Verification of wind forecasts for a set of experimental DMI-HIRLAM

ensemble experiments. DMI Scientific Report 05-01, 2005.

[27] Dudhia J, Gill D, Manning K, Wang W, Bruyere C. PSU/NCAR mesoscale modeling sys-

tem tutorial class notes and users’ guide (MM5 modeling system version 3). Pennsylva-

nia State University/National Center for Atmospheric Research numerical, January 2005.

http://www.mmm.ucar.edu/mm5/.

[28] Villanueva J, Mart´ ı I. Improving wind resource assessment and wind forecast by using a CFD

model developed in CENER. Proc. of the Conference: The Science of Making Torque from Wind,

Delft, The Netherlands, April 19–21, 2004.

[29] Thordarson F. Optimal combined wind power forecasts using exogeneous variables. Master

Thesis, Informatics and Mathematical Modelling, Technical University of Denmark, DTU,

2007.

[30] Cleveland WS. Coplots, nonparametric regression and conditionally parametric fits. In: Mul-

tivariate Analysis and its Applications (Anderson TW, Fang KT, Olkin I, ed.). IMS Lecture

Note-Monograph Series 1994; 24:21–36.

14