Preprint from the EWEC in Madrid, June 16-20, 2003

Using Ensemble Forecasting for Power

1 1 1 Gregor Giebel , Lars Landberg , Jake Badger , Kai Sattler 2, Henrik Feddersen 2, Torben Skov Nielsen 3, Henrik Aalborg Nielsen 3 and Henrik Madsen 3 1 Risø National Laboratory, P.O. Box 49, DK-4000 Roskilde, risoe.dk Tel/Fax: +45 4677 5095 / 5970. E-mail: [email protected]. 2 Danish Meteorological Institute, DK-2100 København Ø, dmi.dk 3 Informatics and Mathematical Modelling, Technical University of Denmark, DK-2800 Lyngby, imm.dtu.dk

Abstract: Short-term prediction of wind power has a long tradition in Denmark. It is an essential tool for the operators to keep the grid from becoming unstable in a region like Jutland, where more than 27% of the electricity consumption comes from wind power. This means that the minimum load is already lower than the maximum production from wind energy alone. Danish utilities have therefore used short-term prediction of wind energy since the mid-90ies. However, the accuracy is still far from being sufficient in the eyes of the utilities (used to have load forecasts accurate to within 5% on a one-week horizon). The Ensemble project tries to alleviate the dependency of the forecast quality on one model by using multiple models, and also will investigate the possibilities of using the model spread of multiple models or of dedicated ensemble runs for a prediction of the uncertainty of the forecast. Usually, short-term forecasting works (especially for the horizon beyond 6 hours) by gathering input from a Numerical Prediction (NWP) model. This input data is used together with online data in statistical models (this is the case eg in Zephyr/WPPT) to yield the output of the wind farms or of a whole region for the next 48 hours (only limited by the NWP model horizon). For the accuracy of the final production forecast, the accuracy of the NWP prediction is paramount. While many efforts are underway to increase the accuracy of the NWP forecasts themselves (which ultimately are limited by the amount of computing power available, the lack of a tight observational network on the Atlantic and limited physics modelling), another approach is to use ensembles of different models or different model runs. This can be either an ensemble of different models output for the same area, using different schemes and different model physics, or a dedicated ensemble run by a large institution, where the same model is run with slight variations in initial conditions and/or parameterizations. Two of the large ensembles run this way are available from the European Centre for Medium-Range Weather Forecasts (ECMWF) in Reading, and from the National Center for Environmental Protection (NCEP) in the US. These are used to calculate the uncertainty of the prediction from the model spread. However, since the model domains are global, it is not certain that this approach will work, due to insufficient spread in Denmark. Additionally, we will try to establish an ensemble of members of DMIs forecasts together with forecasts from the . The project is funded by the Danish PSO funds under the reference no. ORDRE-101295 (FU 2101).

1 INTRODUCTION use a proper Numerical Weather Prediction (NWP) model, of the type used in the large meteorological centres. Actually, The high regional penetrations of wind energy can only be short-term prediction models based on NWP outperform time- integrated successfully using a short-term prediction model, to series models already for 4-6 hours lead-time. However, for a predict the wind power production for some hours ahead. How week-ahead prediction, even the best NWP models are not many hours look-ahead time one needs, depends on the quite good enough, and a month-ahead prediction is also application. For the scheduling of power plants, the most theoretically hardly conceivable due to the inherent chaos in important time scale is determined by the start-up times of the the atmosphere. For an introduction into short-term prediction other power plants in the grid, from 1 hour for gas turbines up models, refer to [1]. A much larger text on the topic is to 8 hours or more for the largest coal fired blocks. The underway, too [2]. markets determine the second time scale of interest. In The typical short-term prediction model uses NWP data from Denmark, the main market for electricity is the NordPool, one operational model, run by a meteorological service. Eg , followed by the German market. NordPool regulations the Danish Zephyr model uses the four daily runs of the mandate trading for the next 24-h day at noon the day before. HIRLAM model of the Danish Meteorological Institute (DMI) This means that the wind power forecasts have to be accurate as input. This is usually the local met. institute, since they for about 37 hours lead-time. A third time scale is involved in know the best parameterisations for the local conditions and maintenance planning, of power plants or the electrical grid. run the model with the highest resolution for the particular The ideal lead-time for this would be weeks or even months country. ahead. However, using data from more than one met. institute can The different lead-times can be served with different short- increase the resilience against errors, and possibly also the term prediction models. The scheduling of a quick-response accuracy of the forecasts. grid can be done with time-series analysis models alone, based on past production. For the trading horizon, it is necessary to Besides the forecasted value for the power, the utilities also would like to have an estimate on the accuracy of the current prediction. It has been established [3, 4] that the errors of the NWP models are not highly dependent on the level of the wind speed. That means that the shape of the power curve has a large influence on the uncertainty of the forecasts: where the power curve is steep, the error is amplified, while in the flat areas of the power curve, the error becomes less relevant. Combining this with the historical performance of the model and a term depending on the horizon is the state-of-the-art in uncertainty prediction these days. A completely different class of uncertainty could be introduced if we could assess the predictability of the current weather situation. This is where ensemble forecasting comes in. The idea in ensemble forecasting is to cover a larger part of the possible futures through introduction of variation in the initial conditions. This can be used as a sensitivity analysis on the Figure 1: A spaghetti plot of tracks over the influence of variations in different factors. There are four main continental US. Image Peter Houtekamer, Environment Canada possibilities for the calculation of ensembles of forecasts: • Every model run is started with a little variation in the initial conditions. In this way it can be investigated how 2 PREVIOUS WORKS sensitive the result is against small changes in the initial A number of groups in the field are currently investigating the conditions (a.k.a. the butterfly effect). These small benefits of ensemble forecasts. variations are still compatible with the likely error in the Giebel et al [5] and Waldl and Giebel [6,7] investigated the analysis at time zero. Keep in mind that the meteorological relative merits of the Danish HIRLAM model, the observational network has a density over land in the order Deutschlandmodell of the DWD and a combination of both for of 20-50 km. The ECMWF (European Centre for Medium a wind farm in Germany. There, the RMSE of the Range Weather Forecasts in Reading) runs this kind of Deutschlandmodell was slightly better than the one of the ensemble twice a day with 50 members. NCEP in the US Danish model, while a simple arithmetic mean of both models run another one with 11 members. yields an even lower RMSE. • A multi-scheme ensemble starts with identical data input, Moehrlen et al [8] use a multi-model ensemble of different but uses different variants of the same model. This can be parameterisation schemes within HIRLAM. They make the different data assimilation techniques (optimal interpolation, point that, with the spacing of the observational network being 3D-Var or 4D-Var), different numerical integration schemes 30-40 km, it might be a better use of resources to run the NWP (Eulerian or Lagrangian) or different physical model not in the highest possible resolution (in the study 1.4 parameterisations. Dependent on the choices made, the km), but use the computer cycles instead for calculating model behaviour changes. Note that this does not ensembles. A doubling of resolution means a factor 8 in necessarily mean that the different combinations do better running time (since one has to double the number of points in or worse in the traditional verifications scores. all four dimensions). The same effort could therefore be used • A variation on the scheme above is a proper multi-model to generate 8 ensemble members. The effects of lower ensemble, using results from the operation models of resolution might not be so bad, since effects well below the different institutes, eg the Deutscher Wetterdienst or the US spacing of the observational grid are mainly invented by the National Weather Service. model anyway, and could be taken care of by using direction • The easiest possibility is to use a poor man’s ensemble. dependent roughness instead. This is only valid if the Since for instance DMI delivers a new model run every 6 resolution is already good enough to properly represent fronts hours for 48 hours in advance, this means that at every point and meso-scale developments. in time there are up to 7 overlapping forecasts, done at Their group is also the leader of an EU-funded project called different start times in the past. Honeymoon. One part of the project is to reduce the large- A typical way to plot ensemble members is the so-called scale phase errors using ensemble prediction. "spaghetti plot". Where all the lines are together, the Landberg et al [4] used a poor man’s ensemble to estimate the probability for this low-pressure trajectory happening is large, error of the forecast for one wind farm. The assumption is that while in the centre of the picture it is uncertain which one is when the forecasts change from one NWP run to the next, then going to happen in real life. the weather is hard to forecast and the error is large. However, The idea here is that for a hard to predict weather situation, this uncertainty forecast fared no better than an uncertainty there is large spread between the ensemble members, and the derived from the wind speed level. skill of the prediction is low. To which extent there really is a Roulston et al [9] evaluated the value of ECMWF forecasts correlation between spread and skill, will be investigated in for the power markets. Using a rather simple market model, this project. they found that the best way to use the ensemble was what they called climatology conditioned on EPS (the ECMWF Ensemble Prediction System). The algorithm was to find 10 In our project we use both the raw ECMWF ensemble days in a reference set of historical forecasts for which the forecasts every six hours (horizontal resolution ~ 75 km), and wind speed forecast at the site was closest to the current we use the DMI-HIRLAM model to downscale the ECMWF forecast. This set was then used to sample the probability ensemble forecasts to a horizontal resolution of approximately distribution of the forecast. This was done for the 10 th , 50 th 20 km for the first 72 hours with output every hour. and 90 th percentile of the ensemble forecasts. 3.3 Multi-Model Ensembles 3 ENSEMBLE PROVIDERS The greatest possible variation when using the same input data is found when using models from different institutes, running 3.1 NCEP different numerical models and maybe even different data NCEP (National Centre of Environmental Prediction, acquisition schemes. This is the case eg for DMI-HIRLAM http://www.emc.ncep.noaa.gov) in the US operates a twice and the DWD Lokalmodell (LM). The latter even uses its own daily ensemble of global weather forecasts. At 00Z, 12 global model to drive it. While HIRLAM has a largely forecasts are made comprising of 1 control (unperturbed rectangular grid, the LM uses a triangular grid. The DWD forecast) at high resolution (T170 truncation from 0 to 168 does its own data assimilation, both for the LM and the hours and then T126 to 384 hours), 1 control at the ensemble driving Globalmodell (GME). The GME has a icosahedral- resolution (T126 from 0 to 84 hours ahead and then T62 out to hexagonal grid with a mesh size of ~60 km horizontal, and 384 hours), and 10 perturbed forecasts at the ensemble uses 31 layers in the vertical. The 200-s internal time step is resolution. At 12Z, 1 control and 10 perturbed forecasts are run twice a day out to 174 hours. The GME data assimilation made all at ensemble resolution. for the wind field is done by 3D multivariate optimal The perturbations are constructed by a so-called “breeding interpolation of deviations of observations from 6-h forecasts. cycle”. The breeding of one such perturbation is given as an This model provides the hourly boundary values to the LM. example. It is started by creating an alternative initial state of The LM is a non-hydrostatic meso- scale regional NWP the atmosphere by “seeding” the best guess state of model for central and western Europe. It is run on a rotated atmosphere with a random perturbation. The two slightly spherical grid with ~7 km mesh size horizontally, with 35 different states of the atmosphere are integrated forward in levels vertically, resolving the lower 1500 m above the model time using the forecast model. One day later the difference orography with 10 layers. It is run on 00, 12 and 18 UTC, with between the two forecasts is used, with a rescaling, as the a horizon of 48 hours. Data assimilation is by nudging towards initial perturbation in the next forecast. The rescaling is done observations. to limit the amplitude of the perturbation to a scale similar to Earlier comparisons with the older Deutschlandmodell have that of errors in the initial best guess state of the atmosphere. shown an improvement in accuracy from combining the two The structures that emerge from this “breeding cycle” are the models (see section 2). In this project, two advantages are fastest growing non-linear perturbations and are thought to considered: having two models should improve the resilience resemble the important errors in the initial best guess state of of the forecasting model against missing NWP data, and from the atmosphere. the combination of both models it is expected to gain in The data is collected automatically twice daily from accuracy and to gather a measure of uncertainty. ftp://ftpprd.ncep.noaa.gov/ in the form of GRIB files. These are then locally unpacked and for a region covering Europe and part of North Africa (12°W-25°E, 25°N-65°N) data from 4 THE PSO ENSEMBLE PROJECT a selection of fields is extracted and stored as netCDF files. The Zephyr [10] collaboration has developed a new generation The meteorological fields stored are: the u-component and v- in the Danish tradition of short-term forecasting models. The component of the wind at 10m and 850hPa, the temperature at next step was now, after Zephyr mostly concentrated on 2m and 850hPa, and mean sea level pressure. This forecast software architecture, to go back to provide the best possible data has a look-ahead time increment of 6 hours out to 84 models and integrate the latest methods in . hours and the grid resolution is 1° x 1°. In particular, this involves ensemble forecasting, in all the forms described above. 3.2 ECMWF The project mainly tries to investigate the possibilities for producing good uncertainty estimates of the predictions. One The ECWMF ensemble comprises one global 10-day control problem of ensemble forecasting is that there is very much forecast and 50 similar forecasts that are based on small, initial information in the whole ensemble. The operators can have perturbations to the control initial condition. The perturbations trouble perceiving all members at the same time. Therefore it that are based on so-called singular vectors, are calculated so is necessary to reduce the amount of information to a as to maximise the linear growth of the perturbation kinetic manageable size. One possibility is to show minimum, energy after 48 hours. Perturbations are calculated separately maximum and mean values, another is to show chosen for the Northern and Southern Hemispheres and the Tropics. confidence intervals. The whole plume of forecasts can also be In addition, the effect of model errors is addressed by the use shown. Within the project it should be investigated how to of ‘stochastic physics’ in the NWP model. For each perturbed present the information contained in the ensembles in the most ensemble member the combined effect of the physical meaningful way. parameterisations is randomly perturbed by up to 50% every six hours. 5 CONCLUSIONS After the development of Zephyr, the Danish wind power prediction grouping is now working with an analysis of different options on the side of the NWP input. In particular, the use of ensemble models is investigated to increase the accuracy of the forecasts, for yielding longer forecasts and to get a better measure of the uncertainty of the forecasts. The project runs from 2002-2005.

ACKNOWLEDGEMENTS This project is funded under the Danish PSO rules (ORDRE- 101295 / FU 2101).

REFERENCES

1 Landberg, L., G. Giebel, H.Aa. Nielsen, T.S. Nielsen, H. Madsen: Short-term Prediction – An Overview . Wind Energy 6(3), pp. 273- 280, June 2003. DOI 10.1002/we.96 2 Giebel, G., R. Brownsword, G. Kariniotakis: The State-of-the-Art in Short-Term Prediction of Wind Power – A Literature Overview . Project report under preparation for the ANEMOS project, to be published on anemos.cma.fr. See also these proceedings. 3 Lange, M., H.-P. Waldl: Assessing the Uncertainty of Wind Power Predictions with Regard to Specific Weather Situations . Proceedings of the European Wind Energy Conference, Copenhagen, Denmark, 2- 6 June 2001, pp. 695-698, ISBN 3-936338-09-4. (Note: the paper is misprinted in the proceedings, better follow the link provided to their university homepage.) 4 Landberg, L., G. Giebel, L. Myllerup, J. Badger, T.S. Nielsen, H. Figure 2: Two different visualisations of the same Madsen: Poor man’s ensemble forecasting for error estimation. ensemble members. The upper one is showing the plume of AWEA, Portland/Oregon (US), 2-5 June 2002 formed by all ensemble members; the lower one uses the 5 Giebel, G., L. Landberg, K. Mönnich, H.-P. Waldl: Relative derived confidence intervals. The forecasts are from the NCEP Performance of different Numerical Weather Prediction Models for ensemble system for a grid point near Risø. Short Term Prediction of Wind Energy . Proceedings of the European Wind Energy Conference, Nice, France, 1-5 March 1999, pp. 1078- With the purpose of producing information regarding 1081, ISBN 1 902916 00 X uncertainty of the wind power forecast based on ensemble 6 Waldl, H.-P., and G. Giebel: The Quality of a 48-Hours Wind Power forecasts two main aspects are important. Firstly, the Forecast Using the German and Danish Weather Prediction Model. information on uncertainty contained in the ensemble forecast Wind Power for the 21st Century, EUWEC Special Topic Conference, must reflect the actual uncertainty observed at one or more Kassel (DE), 25-27 Sept 2000 specific sites or it must be possible to establish a link between 7 Waldl, H.-P., and G. Giebel: Einfluss des dänischen und des deutschen Wettervorhersagemodells auf die Qualität einer 48- the ensemble and the actual uncertainty. Secondly this Stunden-Windleistungsprognose . 5. Deutsche Windenergiekonferenz information will be translated into an uncertainty of the wind DEWEK 2000, Wilhelmshaven (DE), 7-8 Jun 2000, pp. 145-148 power forecast. 8 Moehrlen, C., J. Jørgensen, K. Sattler, E. McKeogh: Power Another part of the project is to investigate whether ensembles Predictions in Complex Terrain With an Operational Numerical can yield better forecasts. While we are at the introduction of Weather Prediction Model in Ireland Including Ensemble new NWP models, we will try to estimate how much longer Forecasting . Poster on the World Wind Energy Conference in Berlin, forecasts help with the maintenance planning at a utility. One Germany, June 2002 wish of the utility partners is actually to have forecasts 9 Roulston, M.S., D.T. Kaplan, J. Hardenberg, L.A. Smith: Value of the ECMWF Ensemble Prediction System for Forecasting Wind spanning a whole weekend ( ie , 72 hours or more) to be able to Energy Production . Proceedings of the European Wind Energy trade on the electricity exchanges in Germany – they close Conference, Copenhagen, Denmark, 2-6 June 2001, pp. 699-702, during the weekends. ISBN 3-936338-09-4 Currently, data acquisition is underway. Some initial runs of 10 Giebel, G., L. Landberg, T.S. Nielsen, H. Madsen: The Zephyr DMI’s HIRLAM ensembles based on the ECMWF EPS have Project – The Next Generation Prediction System . Poster P_GWP145 been performed. The choice of test cases is pretty complete. on the Proceedings CDROM of the Global Windpower Conference The analysis is getting underway now, but has not yielded and Exhibition, Paris, France, 2-5 April 2002 results yet.