State-of-the-Art in Wind Power Prediction in Germany and International Developments

Dr. Matthias Lange, Dr. Ulrich Focken energy & meteo systems GmbH Marie-Curie-Str. 1, D-26129 Oldenburg, Germany mail@energymeteo.de, www.energymeteo.de

Abstract
The state-of-the-art in wind power prediction is described where the focus lies upon practical implementations and experiences concerning the operational systems in Gemany. In addition, the current state of research and development (R&D) in this eld is reported in a European context. Shortterm wind power predictions are based on numerical weather forecasts and use either statistical and physical approaches to determine the anticipated wind power. As an example the physical system Previento is depited in detail. As accuracy is a decisive topic results from current evaluations of the Geman-wide predictions as well as single site forecasts are discussed. The most promising approaches for future improvenments are identied as further ne-tuning of the wind power prediction systems on the one hand and the intelligent use of multiple meterological input on the other hand. In terms of the ne-tuning approach the European research project ANEMOS is taken as an example to highlight the relevant R&D topics that are already investigated. The multiple input approach introduces ensemble predictions and the use of several weather models in parallel into wind power prediction.

In addition, precise wind power predictions play an important role in the allocation of balancing power for the next few hours to come. This intraday application requires reliable updates of the predictions in order to have increasingly better forecasts for decreasing look ahead times. In Germany the installed wind power has grown tremendously in recent years and reached about 17 GW. On stormy days the electricity production has already reached a peak of more than 13 GW (in March 2005). Thus, the contribution of wind energy to the total electricity production uctuates substantially such that wind power predictions play an important role in the daily planning procedures at utilities.

Introduction

Wind power prediction systems provide the information how much wind power can be expected at which point of time in the next few days. In countries having a high penetration of wind energy in the electrical grid, in particular Denmark, Spain and Germany, wind power prediction tools turned out to be indispensable for the energy industry. The forecasts are mainly used for dayahead scheduling of conventional power plants and trading of electricity on the spot market. Therefore, the time horizon of the forecast comprises at least 36 hours as the deadline to place bids on the market is typically at noon for the complete following day.

Figure 1: Wind power installation in Germany is mainly concentrated in the Northern control areas of E.ON and Vattenfall Europe. The immediate horizontal exchange of wind energy requires TSO to immediately exchange wind power between each other in proportion of the electricity consumption in their control area. The four large German transmission system operators (TSO) are responsible for balancing and distributing wind energy that is fed into the transmission grid either directly at high voltage or via the medium-voltage grid. Wind energy has priority and, hence, the TSO are obliged to incorporate the current production. Due to their geographical location the amount of wind power installed in the TSO areas and the electricity consumption are very

divergent(illustrated in g. 1). While e.g. the Vattenfall Europe area contains approx. 38% of wind power the relative electricity consumption is only 19%. Up to August 2004 each TSO was responsible for regulating the complete share of wind power within their area. An exchange of wind energy used to be based on daily or monthly schedules with constant power. But with the amendment of the German renewable energy act (EEG) in September 2004 a new mechanism was introduced which requires TSO to immediately exchange uctuating wind power between each other in proportion of the electricity consumption in their control area. This so-called immediate horizontal wind exchange distributes wind power including the uctuations among TSO in near real-time leading to a more even-handed procedure regarding the allocation of balancing power and the associated costs. Eectively, in terms of wind energy Germany consists of only one control area. Hence, the TSO are mainly interested in predictions for the aggregated output of all German wind farms.

State-of-the-art

Modern wind power prediction system providing forecasts for a time horizon of up to four or ve days in advance typically use numerical weather predictions (NWP). Hence, all the information about the future, in particular the expected evolution of the wind eld, is provided by the NWP. The national weather services but also private weather data providers oer a broad range of dierent NWP data which are suitable for wind power predictions. The key issue in wind power forecasting is to transform the given numerical weather data into the power output of a wind turbine. For this purpose two fundamentally dierent approaches, the statistical approach on the one hand and the physical approach on the other hand, have been developed in recent years. Both of them led to prediction systems which are scientically as well as commercially successful. A recent overview can be found in Giebel et al. [1].

Several dierent methods to determine the relation between forecast and power output have been develped. One very prominent example is the system WPPT by the Danish Technical University [2, 3] which has been working operational at the Danish TSO Eltra for many years. This system uses autoregressive statistical methods to determine the predictions. Under this approach the power production is described as a nonlinear and time-varying (and, hence, non stationary) stochastic process. Another example is the system developed by ISET from Germany [4] which provides forecasts for a number of German TSOs. The system works on articial neural networks (ANN) which are trained with either historical wind farm data or measurements from transformer stations where a number of wind farms is connected. In addition, the system provides an online estimation of the wind power that is currently fed into the electrical grid based on extrapolating measurements at representative wind farms. The advantage of statistical systems is clearly that the predictions are inherently adapted to the location of the wind farm such that systematical errors are automatically reduced. The disadvantage lies in the need for long-term measurement data and an additional eort for the training. Moreover, it is dicult for these systems to correctly predict rare atmospheric conditions if they appear too seldom during the training prediod. Unfortunately, a correct prediction of these rare situations is rather important and can otherwise lead to large forecast errors.

2.2

Physical Systems

2.1

Statistical Systems

The statistical approach is based on training with measurement data. The idea is to derive a statistical relation between the given input from the weather prediction and the measured power output of wind farms. Hence, these systems completely rely on data analysis ignoring the meteorological details.

Figure 2: A physical prediction system uses methods from boundary layer meteorololgy to extrapolate the wind speed at hub height from the given numerical weather prediction. Physical systems use parametrisations based on a detailed physical description of the lower atmosphere. The basic problem to be solved is the trans-

formation of the wind speed given by the weather service on a coarse numerical grid to the on-site conditions at the location of the wind farm. This envolves two important steps: the horizontal interpolation from the grid points to the co-ordinate of the turbine and the transformation of the wind speed from the height provided by the NWP, e.g. 10 m or 100 m, to the hub height as illustrated in gure 2. For this purpose methods from boundary layer meteorology are applied to calculate the vertical wind prole for individual forecast situations. The corrected wind speed is then plugged into the corresponding power curve to determine the power output. More details can be found in a recently published book on this subject [5]. A number of physical systems has been introduced in recent years. One of the rst systems was Prediktor [6, 7] developed by Risoe in Denmark. In Germany two physical systems are in operational use for the major TSOs: The system SOWIE by Eurowind GmbH and Previento by energy & meteo systems GmbH which was developed by the University of Oldenburg [8, 9].

single wind farm. In addition to the prediction values Previento calculates the individual uncertainty for each forecast time depending on the wind speed and the prevailing weather situation [12, 13]. Thus, the end-user is supplied with an estimation regarding the risk of trusting in the prediction. In practical use energy traders and TSOs require the combined power output of many spatially dispersed wind farms in a region instead of that of a single wind farm. Consequently, Previento contains an advanced up-scaling algorithm [8] that determines the expected power output of all wind farms in a certain area based on a number of representative sites selected in an appropriate manner such that spatial smoothing eects are properly taken into account [14]. If measurement data from the wind farms are available a statistical correction of systematic forecast errors is applied using linear regression techniques.

2.3

Operational Service

Figure 3: General scheme of Previento as it is operated by energy & meteo systems. As an example the physical prediction system Previento is discussed in detail. Figure 3 shows the basic scheme of Previento. To calculate the wind speed at hub height the thermal stratication of the atmosphere is modeled in detail [10, 11]. Then the wind speed is transfered to power output by the power curve where either the certied curve or a site-specic curve which has been determined at the location can be used. In a wind farm the shadowing eects among the wind turbines can lead to reductions in power output up to 20% such that these farm eects are also considered. As a result Previento provides the predicted power output of a

With the growing importance of wind power predictions a certain standard has been established regarding the operational service associated with the predictions. The requirements of the end-users are clear: highly accurate forecasts, reliable delivery before 7 UTC and compatibility with existing data formats of the energy management system. Hence, the predictions are electronically delivered by email or FTP in an established data format like ESS or KISS. To ensure a high availability the wind power predictions are run on several independent servers in order to manage a system failure of one of them. Apart from the delivery of the forecasts some providers of wind power predictions oer special trainings to educate the personnel, e.g. in the dispatching centres, how to eectively use the predictions. This additional service enables users to judge dicult weather situations. In the trainings carried out by energy & meteo systems GmbH for dispatchers at the major German TSO it turned out that having some basic meteorological background and an understanding of wind power predictions supports users a lot in coping with forecast errors.

2.4

Accuracy

The accuracy is an important topic as forecast errors are directly related to the use of balancing power and, thus, additional costs. A forecast error of the magnitude of 1 GW over several hours could produce costs of the order of one million Euros.

The accuracy of the wind power predictions is constantly evaluated by the TSO in Germany. The most popular error measure to assess the quality of the forecast is the root mean square error (rmse) based on the deviation between prediction and measurement normalised to the installed wind power. It has turned out that the rmse reects the corresponding cost function as it emphasises large forecast errors which are disporportionately expensive. A recent evaluation by the TSO EnBW of predictions of the aggregated power production of all German wind farms from several providers in Germany revealed a signicant improvement over the last year. The results of this evluation for the Previento prediction is shown in gure 4. The investigated period comprises November 2004 to June 2005, i.e. eight months including the stormy prediods in winter and spring. The accuracy is expressed in terms of the rmse normalised to the installed capacity. The rmse increases from 3.8% intraday (0 - 23 h), to 5.8% dayhead (24 - 47 h) up to 9.5% 2 days ahead.

Figure 5: Monthly evaluation of dayahead prediction of Previento for Germany. tage of spatial smoothing eects, i.e. errors at one location cancel out partly with other distant locations. For a single wind farm the power prediction is, therefore, associated on average with a relatively higher forecast error. In gure 6 the results of the Previento prediction for a wind farm with 17 MW installed capacity is shown. This site is typical for the locations in at terrain in the North of Germany.

Figure 4: Accuracy of prediction of aggregated production of all German wind farms. The Previento system has been evaluated from November 2004 to June 2005. The monthly evaluation of the German wide Previento prediction is illustrated in gure 5. Typically, forecast errors are higher in winter which is due to the generally higher level of wind speeds on the one hand and the larger uncertainty in the prediction of the stormy weather situations, in particular low pressure systems and their frontal zones. During summer the forecast errors are mainly dominated by eects of thermal stratication. Especially, in high pressure situations the wind speed at large hub heights strongly increases after sunset which is dicult to predict. The German wide prediction has a comparatively low forecast error because it can take full advanFigure 6: Accuracy of single wind farm with 17 MW installed power located in at terrain in Northern Germany.

Research and Development

The research and development (R&D) activities related to wind power prediction follow two main approaches: the improvement of the wind power prediction systems by enhanced statistical or physical methods on the one hand and the improvement of the meteorological input by using several numerical weather predictions in parallel on the other hand.

In addition, a number of R&D projects are devoted to a better integration of uctuating wind energy into the electricity supply system where wind power prediction plays, of course, an important role. Hence, these activities are concentrated on the question how to use the information generated by a wind power prediction system. This issue will touched in the outlook in section 4. In the following key aspects of on-going R&D projects are hightlighted in order to convey an impression of the technological progress that is being made. The focus will be on R&D directly related to wind power prediction. However, it is beyond the scope of this work to provide a complete list of all activities.

3.1

Improved wind power prediction systems

In order to improve wind power prediction systems, i.e. to obtain better results from a given input by a NWP, the current research concentrates on the major challenges faced by wind energy utilisation: Oshore The wind conditions over the ocean are only partly known. In particular, the vertical wind prole is dierent from onshore due to the large heat capacity of the ocean and the variable roughness of the water surface. Complex terrain For hilly and mountainous terrain, especially in connection with high solar irradiation like on the Iberian Peninsula, advanced numerical ow models are required to rene the weather data to the site specic conditions. The wind elds have to be calculated in detail by socalled meso-scale models. Upscaling Not only single wind farms but also regions with high wind power penetration have to be correctly predicted. For this purpose methods to calculate the aggregated power output of many wind farms based on a few representative sites are necessary. Prediction Uncertainty The uncertainty quanties the range of possible deviations from the given forecast value in individual forecast situations. This additional information allows users to assess the risk of trusting in the prediction. These topics are, for example, in the focus of the research project ANEMOS [15] where the major European research institutes in the eld of wind energy have been co-operating together intensively

since the year 2001. Major progress was made in the understanding of the marine boundary layer (oshore) [16, 17], predictions in complex terrain in Spain [18], in the foundations of regional upscaling [14], and in decribing the forecast uncertainties with a broad variety of methods [12, 19]. In addition, the project comprises a benchmarking of the approved wind power prediction systems in Europe with the major operational systems from Denmark, Spain and Germany. The comparison of the systems allowed to dene the state-of-the-art [1] as the initial point to the further development. The improved wind power prediction systems will be summarized under a common shell which allows a convenient management of the dierent systems. As wind power forecasts are already in operational use it is important for applied R&D projects to closely include the end-users, i.e. grid operators, power plant operators, energy traders, and wind farm operators. Their experiences and special requirements are indispensable for an ecient transfer of scientic results into practical application. In the case of ANEMOS a number of large European utilities such as ELSAM (DK), ESB (IRL), EdF (F), EHN (ES), EWE (D) are actively participating. Their task is to implement and test the ANEMOS shell in their daily work.

3.2

Multiple Meteorological Input

The majority of the existing wind power forecasting systems is based on numerical weather prediction models (NWP) which do not provide perfect predictions due to the fact that the atmosphere is a highly non-linear chaotic system [20]. Hence, it is reasonable to use more than one single NWP output to produce the wind power forecast and somehow combine the results. As the dierent forecasts have dierent properties the errors cancel out partly by superimposing the forecasts. Two major appraoches are under development: Ensemble predictions where one numerical model calculates a number of dierent predictions by using slightly dierent initial conditions. Combination of NWP model where the output of several dierent NWP models with dierent physical and numerical implementations is used.

Ensemble Predictions
In this approach the inherent properties of one NWP model are reected. For example gure 7

of physical and numerical implementations of the dynamics of the atmosphere inside the models.

Figure 8: The combination of multiple input from dierent weather services in Europe is used to improve wind power predictions. In a recently started project funded by the German Ministry for the Environment, Nature Conservation and Nuclear Safety (conducted by energy & meteo systems, the TSO RWE and the German Weather Service DWD) the NWP models of the major European weather services are used. Figure 8 shows an overview. The aim is to nd the optimal combination of the models to improve wind power predictions. Figure 7: Ensemble prediction (spaghetti plot) of American weather model GFS. Top: 12 hours after initialisation the dierent ensembles are still close together. Bottom: After 96 hours the ensemble members fall apart into dierent weather situations. shows ensembles of the American NWP GFS. Starting with slightly dierent initial conditions leads to increasingly divergent weather situations. The European weather service ECMWF also supplies ensemble predictions based on 51 dierent ensembles [21, 22]. In wind power prediction the ensembles have been investigated by Giebel et al. [23] and it was shown that averaging can lead to a signicant improvement.

Outlook

Today wind power predictions are used in Germany mainly for balancing the transmission grid, power plant scheduling and estimating the conventional energy that has to be purchased for the next days. However, knowing how the uctuating ressource wind energy will develop over the following hours and days is a valuable information that can be used for further applications in the future: Grid Management With higly localised wind power forecasts allows to predict possible bottlenecks in the local grid. Together with power ow calculations an active grid management can reduce or avoid the danger of grid congestions. Storage Management Storage is regarded as very helpful to compensate for uctuations of wind energy. Several techniques are already on the agenda for different applications, e.g. pumped water stor-

Combination of Weather Models

The combination of dierent NWP models takes advantage of the fact that dierent forecast models have dierent strengths and weaknesses in dierent weather situations. This is due to the broad variety

age, compressed air, or hydrogen. Here, wind power predictions enable strategies to anticipatory operate the storage, e.g. to empty the storage shortly before a storm. Integration into Energy Market In order to directly trade wind power on the energy market, e.g. on intraday or dayahead markets, the amount of wind power that can be oered can only be known with a reliable prediction. This also holds if schedules for wind power production have to be led with the grid operator.

J. Usaola. Sate-of-the-art on methods and software tools for short-term prediction of wind energy production. In Proc. Europ. Wind Energy Conf. EWEC, Madrid, 2003. [2] T.S. Nielsen, H. Madsen, and J. Tfting. Experiences with statistical methods for wind power prediction. In Proc. Europ. Wind Energy Conf. EWEC, pages 10661069, Nice, 1999. [3] T.S. Nielsen, H.A. Nielsen, and H. Madsen. Prediction of wind power using time-varying coecient-functions. In Proceedings of the XV IFAC World Congress, Barcelona, 2002. [4] B. Ernst, K. Rohrig, P. Schorn, and H. Regber. Managing 3000 MW Wind Power in a Transmission System Operation Centre. In Proc. Europ. Wind Energy Conf. EWEC, page 890, Copenhagen, 2001. [5] M. Lange and U. Focken. Physical Approach to Short-term Wind Power Prediction. Springer Verlag, Heidelberg, 2005. [6] L. Landberg. Short-term Prediction of Local Wind Conditions. Technical Report Ris-R702(EN), Ris National Laboratory, 1994. [7] L. Landberg. Short-term prediction of the power production from wind farms. J. Wind Eng. Ind. Aero. Dyn., 80:207, 1999. [8] U. Focken, M. Lange, and H.P. Waldl. Previento - a wind power prediction system with innovative upscaling algorithm. In Proc. Europ. Wind Energy Conf. EWEC, page 826, Copenhagen, 2001. [9] U. Focken, M. Lange, and H.P. Waldl. Previento - Regional wind power prediction with risk control. In Proc. Global Wind Power Conf., Paris, 2002. [10] U. Focken. Leistungsvorhersage r aumlich verteilter Windkraftanlagen unter besonderer Ber ucksichtigung der thermischen Schichtung der Atmosph are, volume 503 of FortschrittBerichte VDI Reihe 6. VDI Verlag, D usseldorf, 2003. [11] U. Focken and D. Heinemann. Inuence of Thermal Stratication on Wind Proles for Heights up to 130 m. In Proc. Europ. Wind Energy Conf. EWEC, Madrid, 2003. [12] M. Lange. Analysis of the Uncertainty of Wind Power Predictions. PhD thesis, University of Oldenburg, 2003. http://docserver.bis.unioldenburg.de/publikationen/dissertation/ 2003/lanana03/lanana03.html.

Conclusion

Wind power predictions are established as valueable tools to integrate wind energy into the electricity supply. In Germany the predictions of the total power output of all wind farms are mainly used for grid operation, power plant scheduling and trading. The important time horizons are intraday (023 hours), dayahead (24-47 hours), and from Friday to Monday (72-96 hours). Wind farm operators do not need forecasts so far due the prioritised feed-in under the current Renewable Energy Act (EEG). Several providers oer wind power predictions on a commercial basis. The accuracy of the predictions has signicantly improved over the last year. Customers typically use dierent forecasts in parallel to see the range of possible outputs. In this context the prediction uncertainty for individual weather situations is supplied by some prediction products. With specic training the personnel can be educated to work more eectively with wind power preditions and cope better with forecast errors. In order to further improve wind power forecasts intense research and development eorts are already on track. In a close co-operation between scientic community and energy industry the important topics of future wind energy utilisation are treated. This comprises in particular oshore, complex terrain, upscaling of representative sites to regions, and assessment of the forecast uncertainty. Moreover, the intelligent use of several numerical weather predictions as input to wind power prediction systems opens further possibilities to reduce the forecast error. On-going projects in this direction focus on ensemble predictions and combinations of dierent weather models.

References
[1] G. Giebel, L. Landberg, G. Kariniotakis, J. Halliday, R. Brownsword, H. Madsen, T.S. Nielsen, U. Focken, M. Lange, I. Marti,

[13] M. Lange and D. Heinemann. Relating the uncertainty of short-term wind speed predictions to meteorological situations with methods from synoptic climatology. In Proc. Europ. Wind Energy Conf. EWEC, Madrid, 2003. [14] U. Focken, M. Lange, K. M onnich, H.P. Waldl, H.G. Beyer, and A. Luig. Short-term prediction of the aggregated power output of wind farms - a statistical analysis of the reduction of the prediction error by spatial smoothing effects. J. Wind Eng. Ind. Aero. Dyn., 90:231, 2002. [15] ANEMOS Development of a Next Generation Wind Resource Forecasting System for the Large-Scale Integration of Onshore and Oshore Wind Farms. http://anemos.cma.fr [16] J. Tambke, M. Lange, and U. Focken. Forecasting Oshore Wind Speeds above the North Sea. Wind Energy Journal, 8:316, 2005. [17] R. Barthelmie, G. Giebel, and J. Badger. Short-term forecasting of wind speeds in the oshore environment. In Proc. Copenhagen Oshore Wind 2005 Conference, Copenhagen, 2005. [18] I. Marti, D. Cabezon, J. Villanueva, M.J. Sanisidro, Y. Loureiro, E. Cantero, and J. Sanz. LocalPred and RegioPred. Advanced tools for wind energy prediction in complex terrain. In Proc. Europ. Wind Energy Conf. EWEC, Madrid, 2003. [19] P. Pinson and G. Kariniotakis. On-line Assessment of Prediction Risk for Wind Power Production Forecasts. Wind Energy Journal, 7:119132, 2004. [20] T.N. Palmer. Predicting uncertainty in forecasts of weather and climate. Rep. Prog. Phys., 63:71116, 2000. [21] ECMWF European Centre for Medium Range Weather Forecasts. http://www.ecmwf.org [22] D. Hou, E. Kalnay, and K.K. Droegemeier. Objective Verication of the SAMEX 98 Ensemble Forecast. Monthly Weather Review, 129:7391, 2001. [23] G. Giebel, L. Landberg, J. Badger,K. Sattler, and H. Feddersen. Using Ensemble Forecasting for Wind Power. In Proc. Europ. Wind Energy Conf. EWEC, Madrid, 2003.