AckermannTh. (ed) Wind Power in Power Systems

Подождите немного. Документ загружается.

//INTEGRAS/KCG/P AGIN ATION/ WILEY /WPS /FINALS_1 4-12- 04/0470855088_ 18_CHA17 .3D – 371 – [365–382/18]

17.12.2004 10:39PM

at a higher resolution (i.e. with smaller grid cells), and the physics and data incorporated

into the models are specifically configured for high-resolution simulations for wind

forecasting applications. The current eWind

configuration uses a single numerical

model [the Mesoscale Atmospheric Simulation System (MASS) model] to generate the

forecasts.

The statistical models are a set of empirical relationships between the output of the

physics-based atmospheric models and specific parameters that are to be forecasted for a

particular location. In this application, the specific parameters are the wind speed and

direction and air density at the location of the wind turbines that provide power to each

of the five substations. The role of the statistical model is to adjust the output of the physics-

based model in order to account for subgrid scale and other processes that cannot be

resolved or otherwise adequately simulated on the grids that are used by the physical model.

The third component of the system is a wind turbine outp ut model. This model is a

relationship between the atmospheric variables and the wind turbine output. The wind

turbine output can be either a fixed relationship that applies to a particular wind turbine

configuration or it can be an empirical statistical relationship derived from recent (e.g.

30-day) atmospheric data and wind turbine output data. The final piece is the forecast

delivery system. The user has the option of receiving the forecast information via email,

an FTP transmission, a faxed page or on a web page display.

For the SCE day-ahead power generation, the system is configured to deliver two

forecasts per day: an afternoon forecast at 5 p.m. PT, and a morning forecast at 5:30

a.m. PT. Each cycle consists of the following three parts (Bailey, Brower and Zack,

2000; Brower Bailey and Zack, 2001):

execution of the physical model;

reconstruction of the statistical model equations based on the previous 30 days of

physical model forecasts and measured data;

evaluation of these statistical equations for each forecast hour in the current cycle.

17.3.6 SIPREO

LICO

SIPREO

LICO is a statistics-based prediction tool developed by the University Carlos

III, Madrid, Spain, and the transmission system operator Red Ele

ctrica de Espan

(REE). It is presently a prototype that provides hourly predictions with a prediction

horizon of up to 36 hours. They are generated by using meteorological forecasts, as well

as online power measurements, as input data to time series analysis algorithms. REE

already uses these predictions for its online system operation. First, SIPREO

LICO

produces predictions for single wind farms. Once there are predictions for each wind

farm, they are aggregated in zones. Finally, the production forecast for the whole of

Spain is generated. For a given wind farm, SIPREO

LICO uses four types of input: the

characteristics of the wind farm; historical records of incoming wind and output power,

to arrive at a real power curve; online measurements of power output; and meteoro-

logical predictions provided by HIRLAM. The algorithms that SIPREO

LICO uses to

generate the predictions depend on what type of input is available. Input data may be

basic, additional or complete. Basic data are those that are available for every wi nd

Wind Power in Power Systems 371

//INTEGRAS/KCG/P AGIN ATION/ WILEY /WPS /FINALS_1 4-12- 04/0470855088_ 18_CHA17 .3D – 372 – [365–382/18]

17.12.2004 10:39PM

farm, and consist of the standard power curve and meteorological forecasts. Additional

data are the basic data plus the real power curve. Complete data are the basic data plus

online measurements of generated energy.

If only basic data are available, a prediction is based on the wind speed predictions

and the standard power curve which aggregates the single power curves of the existing

turbines. In the case of additional data, there are historical records of incoming wind

and output power for a wind farm. A real wind farm power curve can be produced, and

the predictions will be more accurate. In this case, wind direction forecasts as well as

wind speed can be taken into account. If online measurements are available, a statistical

time series analysis is performed. This method is the most accurate and it is the main

part of SIPREO

LICO. Currently, more than 80 % of the wind farms connected to the

grid supply online information. That means that there are accurate predictions for most

of the wind farms. The performance of SIPREO

LICO in the Andalucı

a area, Southern

Spain, is above average. This outstanding performance is attributable to the quality of

the HIRLAM data for that area. In contrast, the performance of SIPREO

LICO in the

region of Navarra is below average. The wind farms in Navarra are located far away

from the coast and the terrain is very complex, too. This is probably the reason for the

large differences between wind speeds that were measured by the wind farms’ anemo-

meters and the HIRLAM predictions. The location of the wind farms results in poor

forecasts for prediction horizons that exceed 12 hours. The performance for the whole of

Spain is satisfactory, though. The aggregation of individual predictions has a positive

effect on the performance of SIPREO

LICO (Sa

nchez et al., 2002)

17.3.7 Advanced Wind Power Prediction Tool

17.3.7.1 Online monitoring of wind power generation

AWPT is a part of ISET’s Wind Power M anagement System (WPMS) that includes an

online monitoring system, a short-term prediction system (1 to 8 hours ahead) and a

day-ahead forecast. First, we will give a brief overview of the WPMS, starting with the

online monitoring system. The most precise method for obtaining data for the genera-

tion schedule and grid balance is to monitor online the power output of each wi nd

turbine in a certain supply area. In Germany, however, wind turbines are scattered

throughout the country, and it is hardly realistic to equip all wind turbines with an

online monitoring system (Ensslin et al., 1999).

Online monitoring requires an evaluation model, which makes it possible to extra-

polate the total power fed in by the wind turbines of a larger grid area or control area

from the observed time series of power output of representative wind farms (see Figure

17.3). In cooperation with the German TSO E.ON Netz GmbH, ISET has successfully

developed an online monitoring system that is able to provide data on the wind power

generation of about 6 GW from all the wind turbines of the entire utility supply area

(EWEA, 2000). This model transforms the measured power output from 25 representa-

tive wind farms with a total capacity of 1 GW into an aggregated wind power produc-

tion for approximately 6 GW of installed wind power capacity. The selection of the

sampled wind farms and the development of the transformation algorithms are based on

the long-term experience of the ‘250 MW Wind’ programme and on its extensive

372 The German and Danish Networks

//INTEGRAS/KCG/P AGIN ATION/ WILEY /WPS /FINALS_1 4-12- 04/0470855088_ 18_CHA17 .3D – 373 – [365–382/18]

17.12.2004 10:39PM

measurement data and evaluation records (IRENE, 2001). The current wind power

production is calculated by extensive equation systems and parameters, which consider

various conditions, such as the spatial distribution of the wind turbines and environ-

mental influences. The data collected at the selected wind farms are transmitted online

to the control centre.

The online model is a basic part of ISET’s WPMS, which consists of three levels. The

second and third level (the day-ahead forecast and the short-term prediction) are based

on ISET’s wind power prediction tool AWPT.

17.3.7.2 Short-term prediction

In cooperation with E.ON Netz, Vattenfall Europe Transmission, Aktiv Technology

and the German Weather Service, ISET has developed a new wind power prediction

model, the AWPT. This model uses a mixture of three successful approaches:

accurate numerical weather prediction, provided by the German Weather Service;

the determination of the wind farm power output, based on ANNs;

the online model, used to extrapolate the total power that is fed into the utilities’ grid

from the predicted power.

The meteorological component of the prediction tool is based on operational weather

forecasting. For this purpose, the routine updates from the NWP model for the

respective area (i.e. the Lokal-Modell of the German Weather Service) are used. The

Figure 17.3 Representative wind farms providing input data for the online model in the Wind

Power Management System of the Institut fu

r Solare Energieversorgungstechnik

Wind Power in Power Systems 373

//INTEGRAS/KCG/P AGIN ATION/ WILEY /WPS /FINALS_1 4-12- 04/0470855088_ 18_CHA17 .3D – 374 – [365–382/18]

17.12.2004 10:39PM

Lokal-Modell is the latest generation model of the German Weather Service and is

specifically designed for handling the typical small-scale circulation patterns in German

inland areas. It provides results with a spatial resolution of 77km

and in an hourly

sequence. The updates are calculated twice a day. The following output data of the

Lokal-Modell are used for wind power prediction (see also Figure 17.4):

wind speed at 30 m above ground;

wind direction;

air pressure and temperature;

humidity;

cloud coverage;

temps (plot of air temprature in different altitudes), which are used for the determin-

ation of the atmospheric stability class.

For selected, representative wind farm sites, the routine forecast updates are evaluated

and the corresponding power output of the wind farm is calculated by ANN. The medium

points of these sites may be used as the grid points of the Lokal-Modell, for instance.

Several institutes had alread y examined whether and how ANNs could be used to

predict the power output of wind turbines (Tande and Landbe rg, 1993). Unlike previous

projects, ISET not only uses measur ement da ta from the past few hours for its power

predictions but also applies ANN modules to understan d the relationship between the

meteorological data and the wind power output, using past wind and power data. By

comparing the computed results with the observed power data, the optimal configura-

tion of the ANN modules is determined. The advantage of this approach is that it uses

Figure 17.4 The numerical weather prediction (NWP) model. Note: ANN ¼ artificial neural

network

374 The German and Danish Networks

//INTEGRAS/KCG/P AGIN ATION/ WILEY /WPS /FINALS_1 4-12- 04/0470855088_ 18_CHA17 .3D – 375 – [365–382/18]

17.12.2004 10:39PM

observed data to determine the coherence between meteorological da ta and wind power

output. It is not really possible to describe sufficiently the real relationship between

meteorological data and wind farm power output by physical models. Moreover, incor-

poration of further parameters does not require expensive modifications of the model.

These trained networks compute the predicted wind power output of the representative

wind farms. The predicted power output is then used as the input to the transformation

algorithm of the online model. The online model requires only a limited number of

locations with predicted wind speed in order to predict the total wind power generated

in large utility supply areas. This tool represents the second tier of the WPMS and provides

the wind power output for the control area or for selected subareas. The resolution is

1 hour and the prediction schedule is 72 hours. The NPW at the German Weather Service

runs twice a day, starting at midnight and at noon and finishing at 8 a.m. and 8 p.m.,

respectively. The day-ahead forecast (typically used for the load management) is based on

the results of the morning run of the NWP and provides the wind power generation for the

entire next day, without any update. The forecast horizon is therefore between 24 hours

and 48 hours. For this forecast period, the average error [root mean square error (RMSE)]

between predicted and observed power is about 9.6 % of the installed capacity.

Figure 17.5 shows the frequency distribution of the forecast error for the same time

period. Large errors, over 20 %, occur only 3 % of the time; 86 % of the time the

forecast errors range from 10 % to þ10 %.

The most frequent cause of forecast errors is attributable to the wrong timing of

significant, large variations of weather situations. Wind power prediction models that

are based only on operational weather forecasts are not able to correct these deviations.

Therefore, another module, the third tier, uses the predicted wind farm power output,

computed by the second tier in combination with measured wind farm power output of

Prediction error (%)

–50 –40 –30 –20 –10 0 10 20 30 40 50

Frequency (%)

Figure 17.5 Frequency distribution of the forecast error (as a percentage of installed power) of

the day-ahead forecast (prediction horizon of between 24 and 48 hours)

Wind Power in Power Systems 375

//INTEGRAS/KCG/P AGIN ATION/ WILEY /WPS /FINALS_1 4-12- 04/0470855088_ 18_CHA17 .3D – 376 – [365–382/18]

17.12.2004 10:39PM

the recent past, in order to provide topical updates and adjustments of the predictions

computed by the second layer. These updates and adjustments of the predicted power

output for the following 6 hours are also computed by ANN and can be carried out at

any time. Table 17.1 shows the accuracy of the 3–6 hour forecasts in comparison to the

persistence model.

The ad vantages of this model can be summarised as follows:

the model architecture and the combination of online monitoring and prediction

modelling make it universally applicable;

it provides high precision with minimum co mputation time;

it is easily adaptable to other renewable energy sources.

E.ON has been using the model since July 2001 and, in 2003, the model was installed

at two other German TSOs, RWE Net and Vattenfall Europe Transmission GmbH. It is

also used to calculate the horizontal wind energy exchange between the TSOs in

Germany. Although already in operation, the model is continuously being improved.

The next important feature is to provide a confidence interval for each prediction

(Ensslin et al., 2003).

17.3.8 HONEYMOON project

HONEYMOON (http://www.honeymo on-windpower.net) is being developed by

University College Cork, Ireland and is funded by the European Commission. It is

scheduled to be completed by 2004. Its approach differs from other win d power fore-

casting in that a wind power module is directly coupled to an NWP model in order to

obtain the wind power from the NWP. Offshore winds and waves can be predicted, too,

by coupling an atmospheric model to wave and ocean models. Offshore wind farms not

only requir e forecasts that include the ocean model and take into acco unt the direct

impact of the sea on wind power prediction but also require predictions of waves and

currents for maintenance purposes. With this coupling, customers may be provided with

offshore predictions for the maintenance of offshore turbines and towers.

HONEYMOON uses ensemble predictions to supply an uncertainty estimate of the

power predictions.

Table 17.1 Forecast errors RMSE and correlation between measurement and prediction

Forecast time (h) ISET AWPT 3 Persistence

RMSE (%) Correlation RMSE (%) Correlation

3 5.2 0.95 6.5 0.92

4 5.7 0.94 8.0 0.88

5 6.1 0.93 9.4 0.84

6 6.3 0.93 10.5 0.80

Note:AWPT¼ Advanced Wind Power Prediction Tool; ISET ¼ Institu

tfu

r Solare

Energieversorgungstechnik; RMSE ¼ root mean square error.

376 The German and Danish Networks

//INTEGRAS/KCG/P AGIN ATION/ WILEY /WPS /FINALS_1 4-12- 04/0470855088_ 18_CHA17 .3D – 377 – [365–382/18]

17.12.2004 10:39PM

17.4 Conclusions and Outlook

17.4.1 Conclusions

The described methods can be roughly divided into two classes: those that apply NWP

and use, as far as possible, physical equations in order to calculate the power output of

wind farms, and those that use statistical methods to obtain the power output from

NWP results. In addition to NWP data, the statistica l systems incorporate online

measurements to optimise short-time predictions of a few hours. However, the physical

systems use some model output statistic (MOS) modules to correct errors systematically.

In general, the results show that NWP-based forecasts are better at supplying forecasts

for longer periods of time, between 6 to 48 hours, with systems providing very short-

term predictions (0.5 h to 6 h) needing additional measurements to provide accurate

results. The first version of WPPT, for instance, used only measurement input data. It

provided satisfactory predictions for short horizons of up to 8 hours, but was not useful

for longer prediction horizons. Its second version soon followed and also used NWP

data as input data. This version gave much better results for longer prediction times.

Prediktor and Previento use only NWP data. The systems require measur ement data

to optimise the systems with MOS, but the data do not have to be collected online. This

makes the data collection much easier.

Zephyr, SIPREO

LICO and AWPT use NWP and online measurements as input da ta.

These systems combine the advantages of both approaches: the high accuracy of the

NWP for a longer prediction horizon (up to 48 hours) and the advantage of online

measurements for forecasts in the short-term range. However, the required online

measurements cause additional costs. The measured wind farms have to be selected

carefully in order to be representative of the whole area, and the equipment has to be

very reliable.

For calculating the prediction error, we first need the measured wind power. These

data do not have to be available online, but they ha ve to have a high accuracy. A single

wind farm is easily observed, but to compute the sum power of a large supply area

usually involves some upscaling algorithm. This depends, however, on the structure of

the wind farms. In Denmark and Germany, for example, there are many small wind

farms as well as single wind turbines. It is therefore difficult to arrive at the sum power

through measurements. In Spain and the USA, there are fewer but larger wind farms,

which makes it easier to obtain real measurement data without any upscaling. It is

evident that the higher the ratio between measured and total power, the higher the

upscaling accuracy, on the one hand, but the more costly the data collection, on the

other hand. The Zephyr system is planned to obtain measurement data from all wind

farms in the area, whereas Prediktor and WPPT use only a couple of dozen large wind

farms in the area. AWTP uses 25 wind farms in the E.ON area and about 100 measure-

ments from single turbines in the ‘250 MW Programme’. Previento includes about 50

measurements of single turbines for confirming the prediction for the whole of Germany.

Various error functions are used to calculate the prediction error. The most common

is the RMSE normalised to the installed wind power. The correlation coefficient is very

useful for this purpose, too. (For a summary of the prediction systems described in this

chapter, see Table 17.2).

Wind Power in Power Systems 377

//INTEGRAS/KCG/P AGIN ATION/ WILEY /WPS /FINALS_1 4-12- 04/0470855088_ 18_CHA17 .3D – 378 – [365–382/18]

17.12.2004 10:39PM

Table 17.2 Overview of prediction systems

Name Developer Method Input data In operation by July 2003?

Physical Statistical NWP Measurement

Prediktor Risø National Laboratory,

Denmark

3 X 3 X Yes, for several wind farms

(approximately 2 GW) in

Denmark

Wind Power

Prediction Tool

IMM, and University of

Copenhagen, Denmark

X 33 3Yes, in Denmark

Zephyr Risø National Laboratory and

IMM, Denmark

333 3Under development

Previento Oldenburg University

Germany

3 X 3 X Yes, for several wind farms in

Germany

e Wind

True Wind Solutions, USA 3 X 3 X Yes, for several wind farms (about

1 GW)

SIPREO

LICO University Carlos III, Madrid,

and Red Ele

ctrica de Espan

X 33 3Yes, in Spain (about 4 GW)

Advanced Wind

Power Prediction

Tool

ISET, Germany X 3

(ANN)

33Yes, in Germany (about 12 GW)

HONEYMOON University College Cork,

Ireland

3 X 3 X Under development

Note: ANN ¼ artificial neural network; IMM ¼ Department of Informatics and Mathematical Modelling, Technical University of Denmark;

ISET ¼ Institut fu

r Solare Energieversorgungstechnik.

//INTEGRAS/KCG/P AGIN ATION/ WILEY /WPS /FINALS_1 4-12- 04/0470855088_ 18_CHA17 .3D – 379 – [365–382/18]

17.12.2004 10:39PM

The utilities have other criteria to evaluate the quality of predictions. A too optimistic

forecast of wind power, for examples, causes a lack of power in the system and the utility

has to buy expensive regulation power. As opposed to this, a too pessimistic prediction

is not as expensive for the utility. However, unless the predicted and the true wind power

exceeds a certain amount (a couple of hundred megawatts in a TSO area) the error is

negligible. The goal for the development of forecast systems is to achieve a minimum

RMSE and a maximum correlation. Optimisation of the systems for the economic needs

of the utilities is then a subsequent step.

All the systems first calculate the power output of single wind farms. They then apply

an upscaling algorithm to obtain the sum power of the entire supply area. Regardless of

the applied system, the RMSE for a single wind farm is between 10 % and 20 %. After

upscaling to the sum power, the RMSE drops under 10 % because of the smoothing

effects produced by adding different signals. The larger the area the better the overall

prediction (for further information, see Holttinen, Nielsen and Giebel, 2002).

The systems are difficult to compare because they are applied in different areas. The

Danish systems run in a mostly flat terrain, which makes obtaining an accurate NWP

much easier, There are fewer slack periods, though, and the average wind power output

is higher, which increases the overall error. In Germany, it is much more difficult to

arrive at an accurate NWP, especially in the lower mountain areas, but periods of low

wind power, especially in the summer, decrease the overall RMSE.

It is impossible to decide which of the described systems is best, unless they are

compared under the same circumstances. It all depends on the user’s need s and his or

her abilities of providing data. Zephyr, AWPT and SIPREO

LICO are probably super-

ior regarding the entire prediction horizon from 1 to 48 hours. However, the costs

related to setting up the system are high, especi ally for collecting online data.

Systems using physical equations, such as Prediktor or Previento, require the exact

location and environment of the wind farms they predict. They also need a long

computation time to transform the geostrophic wind speed down to the hub heights

of the wind turbines, with WAsP or similar programs. Statistical models, such as WPPT

and AWPT, however, need time to learn the correlation between wind and power when

they are being set up, but they require only a minimum of computing time. The ANEMOS

project, funded by the European Commission, compares different approaches of

forecasting and is developing a new system that is expected to outperform the existing

systems (http://anemos.cma.fr).

All prediction systems have produced unsatisfactory forecasts in the following

situations:

during very-fast-changing weather conditions;

in very high wind situations (storms), when wind turbines cut out for safety

reasons;

when very local weather changes (thunderstorms) cause problems to the upscaling;

when bad measurements are produced that are not automatically detected (only for

systems using online data);

when weather service forecasts are delayed or missing;

when meteorological forecast data are poor (very often, a time shift between forecast

and real data causes a large error).

Wind Power in Power Systems 379

//INTEGRAS/KCG/P AGIN ATION/ WILEY /WPS /FINALS_1 4-12- 04/0470855088_ 18_CHA17 .3D – 380 – [365–382/18]

17.12.2004 10:39PM

17.4.2 Outlook

The largest potential for achieving better prediction accuracy lies in an improved NWP

model. A wind power model cannot be better than its weather input data. Running the

NWP models repeatedly with slightly different input data (ensemble forecasting) will

give the most probable result. Furthermore, the spread of the single results of the

ensemble will provide an estimation of the uncertainty, which is also very helpful for

users of the model.

The systems using online data can improve the NWP-based forecast, but only for a

small pre diction horizon of a few hours ahead. Nevertheless, a lot of research will be

dedicated to these forecasts, because they are very interesting from a technical point of

view. Many of the new power plants are fast adjustable, such as gas turbines or small

combined heat and power plants with heat storage. With a very accurate and reliable

forecast for the next few hours, these plants do not have to idle but can be switched off if

there is sufficient wind. This will help to save fossil fuel and achieve maximum benefits

from wind power regarding energy saving.

Another major field of future development is prediction for offshore wind farms.

Although weather prediction for offshore sites seems at first sight to be easy because of

the flat orography, the interaction between wind and waves and coastal effects requires

further work.

In addition to short-ter m predictions, power plant operators have an increased

interest in seasonal forecasts in order to be able to plan their fuel stock accu rately.

References

[1] Bailey, B., Brower, M., Zack, J. (2000) ‘Wind Forecast: Development and Applications of a Mesoscale

Model’, in Wind Forecasting Techniques: 33 Meeting of Experts, Technical Report from the International

Energy Agency, R&D Wind, Ed. S.-E. Thor, FFA, Sweden, pp. 93–116.

[2] Beyer, H. G., Heinemann, D., Mellinghof, H., Mo

nnich, K., Waldl, H.-P. (1999) ‘Forecast of Regional

Power Output of Wind Turbines’, presented at the 1999 European Wind Energy Conference and Exhibi-

tion, Nice, France, March 1999.

[3] Brower, M., Bailey, B., Zack, J. (2001) ‘Applications and Validations of the Mesomap Wind Mapping

System in Different Climatic Regimes’, presented at Windpower 2001, Washington, DC, June 2001.

[4] Dispower (2002) ‘Progress Report for DISPOWER’, reporting period 1 January 2002 to 31 December

2002; project founded by the European Commission and the 5th (EC) RTD Framework Programme

(1998–2002) within the thematic programme ‘Energy, Environment and Sustainable Development’,

Contract ENK5-CT-2001-00522.

[5] Ensslin, C., Ernst, B., Hoppe-Kilpper, M., Kleinkauf, W., Rohrig, K. (1999) ‘Online Monitoring of

1700 MW Wind Capacity in a Utility Supply Area’, presented at the European Wind Power Conference

and Exhibition, Nice, France, March 1999.

[6] Ensslin, C., Ernst, B., Rohrig, K., Schlo

gl, F. (2003) ‘Online Monitoring and Prediction of Wind Power in

German Transmission System Operation Centers’, presented at the European Wind Power Conference and

Exhibition, Madrid, Spain, 2003.

[7] EWEA (European Wind Energy Association) (2000) ‘Press Release 24’, January 2000.

[8] Focken, U., Lange, M., Waldl, H.-P. (2001) ‘Previento – A Wind Power Prediction System with an

Innovative Upscaling Algorithm’, presented at the 2001 European Wind Energy Conference and Exhibi-

tion, Bella Center, Copenhagen, Denmark, 2–6 July 2001.

[9] Holttinen, H., Nielsen, T.S., Giebel, G. (2002) ‘Wind Energy in the Liberalised Market – Forecast Errors in

a Day Ahead Market Compared to a More Flexible Market Mechanism’, presented at the 2nd Inter-

380 The German and Danish Networks