Wind in complex terrain. A comparison of WAsP and two ... - WindSim

Wind in complex terrain.A comparison of WAsP and two CFD-models.Erik Berge Arne R. Gravdahl Jan Schelling Lars Tallhaug Ove UndheimKjeller Vindteknikk AS Vector AS Hydro Oil & Energy Kjeller Vindteknikk AS Institutt for energiteknikk*erik.berge@vindteknikk.no gravdahl@vector.no jan.schelling@hydro.com lars.tallhaug@vindteknikk.no ove.undheim@vindteknikk.no*Presently at Kjeller Vindteknikk ASAbstract:WAsP and the two CFD-models WindSim and3DWind have been compared at a complex terrainsite in western Norway for the two dominating winddirections south and north. One year of measurementsfrom two 50 m masts and one 10m mast wereavailable. The paper analyses the mean absoluteerror, the vertical variations of wind speed andturbulence with height, and meso-scale windvariations across the wind farm. Despite thesimplicity and known weaknesses of WAsP incomplex terrain, no improvements of the averagewind speed calculations were obtained by use of theCFD-models. Meso-scale wind variations of the order2-3 m/s during a particular weather situation wereencountered for a case study by use of a meso-scalemodel.Keywords: Complex terrain, WAsP, CFD-models.1 IntroductionLarge wind energy developments are at the planningstage in Norway. Many of the planned wind farmswill be located in rather complex coastal sites wherewind resource assessment is difficult.Wind models are potentially important at complexterrain sites in the wind resource assessment sincemeasurements can only be afforded at selectedpositions, and the wind variations may be large overshort distances.For micro-scale flow (spatial scales of 1 m to 2km), the WAsP model [1] is most commonly used inthe wind resource analysis, but in areas with flowseparation the model is not very well suited forresource assessment. However, simple terraincorrections (so called RIX-analysis see [2]) have beenapplied with success to the WAsP simulation for quitecomplicated terrain [2,3].CFD-models are also utilized in micro-scale windresource assessment. Most commonly, the CFDmodelsdevelop a steady-state time-independentsolution for the wind and turbulence fields. Thus, aphysically more realistic picture of the flow field canbe obtained compared to WAsP. Examples of suchmodels applied to Norwegian wind resourceassessment are WindSim [4] and 3DWind [5]. Still,these CFD-models avoid the inclusion of the energyequation, thus stratified flow cannot be realisticallyhandled. Also, the assumption of a steady stateturbulent field may be doubtful since the turbulenceactually often is very intermittent and transient.Nested models that handle real-timemeteorological data from the synoptic scale down tothe micro-scale, including buoyancy effects on themicro-scale flow, have also been developed and tested[6]. Such models still require weeks of computationalefforts on today’s most powerful computers [6], inorder to generate a wind resource map for a windfarm. Thus the temporal and economical costs of suchmodels may still be too high for practicalapplications.In complex terrain, meso-scale wind variationswithin a wind farm area may as well be anticipated(spatial scales of 2 km to 200 km). Meso-scalemodels, that solve the coupled equations of dynamicaland thermo-dynamical processes in the atmosphere,may be needed together with the micro-scale modelsto achieve a complete resource analysis. Exampleson such models are the WRF-model [7], the MC2-model [8] and the RAMS-model [9].The aim of the present study has been to validatethe two CFD-models WindSim and 3DWind and tocompare with the WAsP model at a complex terrainsite in Norway. In addition, examples on meso-scalewind variations based on the WRF-model are given.One year of measurements from one 10 m and two 50m measuring masts have been made available from aHydro Oil & Energy site in western Norway. Bothaverage mean wind and turbulence conditions havebeen analyzed. An important issue has been todetermine whether or not the CFD-models couldimprove the confidence in the wind map at thecomplex terrain site.

2 Measuring program2.1 The Gurskøy regionHydro Oil & Energy initiated in 2000 a fieldmeasuring campaign at Gurskøy in western Norway.The area is located about 25 km east of Stadt, one ofthe windiest regions in Norway. The Gurskøy islandhas a complicated local terrain structure withmountaintops of about 600 m with steep edges downto local valleys and fjords (see Figure 1).Considerable local flow separation is expected in thistype of terrain. From the sectors W to N open oceanflow modified by the local terrain may be expected.Toward S and E high mountains of 800-1500 m areencountered at distances of approximately 15 km ormore, with associated meso-scale wind effects.2.2 Description of the measuring sitesThe measurements were collected from april 2002 tomay 2003 for the sites 1 (z=424 m), 2 (z=390 m) and3 (z=352 m) (see Figure 1). At site 1, themeasurements were carried out with a 10 m steel tubemast, while 50 m steel tube masts where employed atsites 2 and 3.The wind speed measurements were collected withthe NRG Maximum #40 anemometer. Comparisonwith RISØ wind speed sensors have shown that theNRG sensor underestimates the wind speed for windspeeds lower than 2 m/s. For an annual average windspeed in the range 7-9 m/s the underestimation willbe about 0.04 m/s, i.e. much less than the overalluncertainty of the measurements, including theuncertainties of long-term variability of the wind, ofabout 5% (see [3]).The wind rose at site 1 is presented in Figure 2.We note the dominance of winds from S and SW andN and NE.The sites 1 and 2 are situated on an east-westridge with considerable speed-up for both southerlyand northerly flow. The hillside to the south stretchesall the way down to the fjord, while a mountain risesto about 400 m further to the south. Turbulence couldbe expected at all three sites north of this mountainfor the dominating southerly flow. At site 1 there issome sheltering toward W and NW, but this is a wind

Figure 1: Location of the three measuring sites 1, 2 and 3 at the Gurskøy island. Height contours are for every 20m.direction with low frequency. Site 2 is partly shelteredfor the direction NE. Site 3 is located in a SW to NEoriented hillside. Speed-up for SW flow can beexpected at this site. The wind rose is consistent withobservations from other stations in this region.270300240Expected Long−Term Wind Rose at 3043302100180Figure 2: Wind rose at site 1.30600.05 0.1 0.15 0.2 0.253 Micro-scale model set-up3.1 WAsP1501209025 −>20 − 2515 − 2010 − 155 − 100 − 5WAsP version 8.1 was employed in this study. Heightcontours for each 5 m were available for an area ofabout 3km*3km covering the met mast. Outside thisarea 20 m height contours were used. A backgroundroughness of 0.03 m describing the bare mountainareas, was applied. Higher roughness values wereapplied to forests and small towns.3.2 The CFD-modelsWindSim 3DWindSize of model domain 20*20 17*14(km 2 )Total number of cells 144*111*20 142*160*48Size of model domain 3*2 1.44*2.22with 30m resolution(km 2 )Number of cells with 100*67 48*7430m resolutionHeight (m) of the lowestlayers6, 20, 36, 57 7, 15, 24,33, 44Number of layers in the 20 48verticalHeight (m) of the model 2500 5000Boundary conditions (seetext for explanation)Coriolis force No YesTurbulent scheme k-ε k-εTable 1: Set-up of WindSim and 3Dwind for thesimulation.The details of the model set-up for the two CFDmodelsare compared in Table 1.Although an as similar set-up as possible has beenthe intention, inevitably some differences are found.Both WindSim (for a documentation of WindSimsee http://windsim.com) and 3DWind [5], [10] arewind flow solvers based on the 3D ReynoldsAveraged Navier-Stokes equations. The models solvethe atmospheric flow for a “steady-state” case for achosen wind direction. By “steady-state” it is meantthat the model is run until the solution converges toone wind and turbulence distribution for the entiredomain that do not vary outside a pre-set convergencelimit. By simulating for several different winddirections (sectors) an annual average wind speed canbe generated. The models are run for a given set ofconstructed boundary and initial conditions. The k-εturbulence closure scheme is applied in both models.An important advantage of a CFD-modelcompared to WAsP is that the CFD-model calculatesthe turbulent quantities of the flow, and thus givesinformation about high turbulence areas that shouldbe omitted for wind power exploitation. Also, thedirect effects of turbulence on the mean flow arecalculated. Especially in complex terrain, where acomplicated turbulence structure is encountered,benefits of a CFD-model are expected. However, anyturbulence modeling will inevitable introduce newparameterization assumptions and constants, whichalso introduces uncertainty in the modeling (see forexample [11]). Considerable uncertainty is thereforelinked to the quantification of the turbulence and itseffects on the mean flow.Most CFD-models assume a neutral stratificationof the atmosphere. For strong winds this is often agood approximation. But even for wind speeds up to10-15 m/s stable stratification has been observed, atleast in the coastal areas in central Norway [12].Vertical wind shears, turbulent fields etc. maydevelop quite differently during stable (or unstable)conditions [12]. Thus, a limitation in the CFDapproachis encountered here. Another limiting factoris the finite number of directions applied in thederivation of the annual average wind speed. Sincethe long-term wind speed is composed of wind flowsfrom all different directions (though with somepredominant directions), a sample of for example 12sectors may be too small to represent an annualaverage properly.

4 ResultsThe CFD-models are initialized based on assumptionsabout higher level (~500m) and lateral boundary levelwind speed and profiles. A logarithmic wind profile(corresponding to neutral stability conditions) isemployed at the boundaries in both models. Thus, inorder to compare the output of the CFD-models withany measurements, the modeled wind speed fieldshave to be scaled by use of the measurements. Forexample, for a specific annual wind speed from sectorS at 50 m at one of the sites, the output of the CFDmodelwill be assumed equal to the measurement atthis site. It is then further assumed that the wind fieldcan be scaled linearly based on the scaling at thisparticular site. By employing this method modelsimulations of the wind speed at all the measuringsites can be obtained. Thus it is assumed that the ratioof the modeled and measured values at one site willapply to the whole modeling domain. Thisassumption may not be valid everywhere in themodeling domain, but it is quite commonly applied tomodeling. In the following section we assess themodel simulations by using this method. The CFDmodelsare scaled to yield correct wind speed at the50 m level at Site 2 and Site 3 respectively, then theresults at the other measuring sites are extracted fromthe CFD-models.The WAsP model work differently. In order tomake the results from WAsP comparable to the CFDmodels,the measurements at 50 m from Site 2 andSite 3 are inserted in the model and the wind speed ofthe other measuring mast are obtained for the samesectors as utilized by the CFD-models.Note that the CFD-models were run only for thetwo sectors north (N) and south (S) to savecomputational efforts. Each sector covers 30°centered around 360° and 180° respectively.According to the wind rose (see Figure 2), thedirections N and S represent ca. 13% and 18% of thewind data respectively.4.1 Annual average wind speedIn Figure 3 the vertical profiles at site 2 have beenobtained based on a “true” solution at 50 m height atsite 3. For the southerly sector the wind direction isnearly perpendicular to the steep slope to the south ofsite 2 and a strong speed-up is encountered. BothWindSim and WAsP yield only small differencesfrom the observations, while 3DWind yields a largeoverestimation of the wind speed. For the northerlysector, the wind speed actually decreases from 10 mto 50 m at site 2. All three models underestimate thewind speed for this sector, still WAsP yields the bestresults in this particular case.A further comparison of the output of the modelsimulations with the observations has been carried out. Thethree models are all scaled to fit the observations at 50 mheight at either site 2 or site 3.Based on the scaled values of each of the two sites,the annual average wind speed is estimated at theother levels of the same mast (10m and 30m) andHeight (m)Height (m)605040302010Site 2 - S (scaled at site 3 50m)00.00 0.50 1.00 1.50 2.00605040302010Scaled windspeedSite 2 - N (scaled at site 3 50m)00.00 0.50 1.00 1.50 2.00Scaled windspeedOBSWASPWINDSIM3DWindOBSWASPWINDSIM3DWindFigure 3. Scaled vertical profiles at Site 2 for sector S(upper panel) and sector N (lower panel).Modelled2.001.801.601.401.201.000.800.600.400.200.00Obs vs model Sector S0.00 0.20 0.40 0.60 0.80 1.00 1.20 1.40 1.60 1.80 2.00ObservedWasp WINDSIM 3DWind

2.001.801.601.40Obs vs model Sector Nresults underline that a RIX-correction should alwaysbe supported by empirical data, and a generalizationof the method from one area to another is notrecommended.Modelled1.201.000.800.600.400.200.000.00 0.20 0.40 0.60 0.80 1.00 1.20 1.40 1.60 1.80 2.00ObservedWasp WINDSIM 3DWindFigure 4. Observed vs. modelled annual normalizedwind speed for sector S (upper panel) and sector N(lower panel).at the three measuring levels of the other 50m mastand finally at 10 m at site 1. The results aresummarized in Figure 4 and Table 2. The totalnumber of data pointsWAsPSWind-Sim S3D-Wind SWAsPNWind-Sim N3D-WindNMAE 0.11 0.10 0.28 0.14 0.24 0.24COR 0.88 0.84 0.74 0.89 0.58 0.79Table 2: Mean Absolute Error (MAE) and correlationcoefficient (COR) of ratios between models andmeasurement. Each statistical parameter is based on12 data values.to be verified are 12, thus any statistical analysis hasto be interpreted with care since the data points arefew. For the sector S, WAsP and WindSim showsimilar behavior, while the errors are somewhatlarger for 3DWind. For sector N, WAsP givessomewhat surprisingly the best correspondence, whileWindSim has the lowest correlation. Based on theresults of Figure 3 and 4 and Table 2 we concludethat the errors are as large (and actually a littlelarger) for the CFD-models compared to WAsP forthe two sectors analyzed. This is an importantfinding, although we lack insight into the reasonswhy the CDF-models are unable to improve thesimulations at Gurskøy. In the following sections wepresent a discussion aiming to give a betterunderstanding of the modeling results.A RIX-analysis was carried out for the three sitesto check if the WAsP results could be improved. TheRIX-values using a radius of 2 km were 24.2%, 20%and 17.2% for the site 1, site 2 and site 3 respectively.In contrast to [3] no relationship between thedifferences in RIX-values and the prediction errorwas found. For example, employing the 10 mmeasurements at site 1 gave an over prediction of thewind speed at site 2 and 3, while an under predictioncould be expected from the RIX-analysis. These4.2 Turbulence and the directionaldependenceIn Figure 5 the turbulence intensities at site 3 ispresented. While 3DWind overestimates theturbulence for sector S WindSim gives too low values.For sector N both models are closer to theobservations, but still 3DWind over predicts whileWindSim under predicts. The same pattern isencountered for site 2 (not shown), and in othersimilar model experiments using 3DWind. Ourconclusion from these investigations is that the overprediction of the turbulence in 3DWind at site 3 mostprobable is linked to an underestimation of theaverage wind speed at the same site. This also leadsto the overestimation of the predicted average windspeed level of 3DWind at site 2 based on the windfield scale to site 3 (upper panel of Figure 3). InFigure 6 the turbulence pattern derived from 3DWindis presented for the wind directions 180° and 190°.Separated flow and strong turbulence is developed onthe lee side of the mountain to the south of the ridgewhere the three meteorological masts are located. Atboth sites 2 and 3 it is likely that the turbulenceintensities could be rather sensitive to the exact winddirection.From Figure 7 it is also noted that the annualaverage wind speed varies largely for different windspeed directions. The variability at site 2 for thedirection 170°, 180° and 190° is quite well capturedby 3DWind. It is also observed that the outcome ofHeight (m)605040302010Site 3 sector S00 0.05 0.1 0.15 0.2 0.25 0.3Turbulence intensityObs WindSIM 3DWind

Height(m)Site 3 sector N60504030201000 0.05 0.1 0.15 0.2 0.25 0.3Turbulence intensity1.801.601.401.201.000.800.600.400.20Site 1Site 2Site 33DWind at Site 2Wind speed during the period April 2002 to May 2003Obs WindSIM 3DWindFigure 5: Observed vs. modeled turbulence intensityat Site 3.the CFD-model is largely dependent on the selectionof sectors. In the present case with complex terrain 12sectors yield inevitable to course resolution. Ideally atleast 36 sectors should be utilized, which on the otherhand increase the computational demands.5. Comparison of the CFD modeldataIn Figure 8 the differences between WindSim and3DWind for southerly flow are presented. The wind6.9096.908x 10 6Turbulence Intensity []10.000 3 6 9 12 15 18 21 24 27 30 33 36DirectionFigure 7: Wind speed as a function direction scaledby the annual average wind speed at 50 m at site 2.fields of the two models have been scaled in order tominimize the differences between the two models.Note that during the scaling linearity is assumed,although this may not strictly be the case. From thefigure we see quite much higher wind speeds inWindSim than in 3DWind especially in the turbulentlee wakes. This is consistent with the findings insection 4.2 where a higher turbulent level isencountered in 3DWind. This will reduce thehorizontal wind speed, since a larger part of theavailable kinetic energy of the flow is consumed byturbulent processes. Although the two models employthe same turbulent scheme we encounter a largedifference in the resulting wind field and turbulence0.96.9070.8Northing [UTM EUREF89 sone 32]6.9066.9056.904HaugshornetLitlabøhornetHornelva0.70.60.50.46.9030.36.9020.26.9010.13.2 3.21 3.22 3.23 3.24 3.25 3.26 3.27 3.28 3.29Easting [UTM EUREF89 sone 32]x 10 5x 10 6Turbulence Intensity []6.90916.9080.96.9070.8Northing [UTM EUREF89 sone 32]6.9066.9056.9046.9036.9026.901HaugshornetLitlabøhornetHornelva3.2 3.21 3.22 3.23 3.24 3.25 3.26 3.27 3.28 3.29Easting [UTM EUREF89 sone 32]x 10 5Figure 6: Turbulence intensities for wind from 180°(upper panel) and 190° (lower panel) based on3DWind model runs.0.70.60.50.40.30.20.1Figure 8: WindSim – 3DWind at 50 m for sector S.field. This emphasizes the importance of the turbulentprocesses in rough terrain and the complexity of aproper modeling of the turbulence.The frequency distribution of the wind speedscovering the domain of Figure 8 is presented inFigure 9. The wind speed of each single grid-point at50 m in the modeling domain has been included inthis distribution. The data are not scaled and therebynot strictly comparable, still a good picture of thebehavior of the two models is given. We immediatelyobserve the much higher frequency of low wind

Figure 10: Outer domain (red) and inner domain(blue) of the meso-scale simulation.The present analysis shows that, despite thecomplex terrain, WAsP compares better than theCFD-models to the observations when results for thesectors south and north are considered. Thisconclusion is valid both for the vertical wind profilesand the annual average wind speed levels. The meanabsolute error for sector south is 11%, 10% and 28%for WAsP, WindSim and 3DWind respectively. Forsector north the corresponding numbers are 14%,24% and 24%. The RMS-differences in the windspeeds between the two CFD-models have beencalculated for an area where wind energy exploitationis realistic. For both sector south and north anaverage RMS-difference of about 1.5 m/s is found.Meso-scale simulations with a completemeteorological model indicate large wind variationswithin the island due to the mountains to the south.For one particular case with southerly flow, windspeed variations of 2-3 m/s at 100 m height areencountered moving from the western to the easternside of the island. A dependence of the vertical windprofile on vertical stability is also found in the mesoscalemodel simulations. One disadvantage of WAsPand the two CFD-models is that meso-scale windvariations are not taken into account in the modeling.In particular, the effects of the vertical stability onboth the meso-scale and the micro-scale flow may beanticipated to be important (see for example [8]), butit is not accounted for in the micro-scale models. Thisbecomes a limitation of the three models whenapplied to a complex terrain site such as Gurskøy.An advantage of the CFD-models compared toWAsP, is the explicit calculation of the turbulencefield. The comparison of the measured and modeledturbulence intensities shows that 3DWindoverestimates the turbulence while WindSim tends togive too low values. The directional classificationuses 12 sectors of 30°. The model runs show thatlarge variation in the turbulence level may occurwithin a 30° sector when the terrain is complex. Thussmaller sectors should be employed, although thisincreases the computational efforts of the CFDmodels.Through the analysis of the modeledturbulence fields a better understanding of themeasurements at Gurskøy is obtained. For example,the 3DWind runs clearly indicate an influence of thefirst mountain to the south on the turbulence level(and consequently also the average wind speed level)at the measuring sites. This mountain is of the sameheight as sites 2 and 3 and located about 3 km to thesouth.The results of this study emphasize theimportance of high quality measurements in complexterrain for a reliable wind resource mapping. Still,both micro- and meso-scale models are important forthe interpretation of the measurements and for aninvestigation of the turbulence and speed-up effects incomplex terrain. The data material of this report islimited, and therefore further validations of themicro- and meso-scale models in complex terrain arerecommended. This is anticipated to give increasedinsight into the potential and limitations of themodels for wind farm development.

Figure 11: Wind speed at 50 m height at 03 UTC 01.01.2005. WRF-run with 1 km horizontal resolution. Heightcontours (black lines) are drawn every 30 m. The vertical profiles are from the positions 1 (SW) and 2 (SE).Figure 12:Vertical profiles of wind speed and potential temperature at the SW corner (red lines) and SE corner(blue lines of the island Gurskøy. Height represent the heights above sea level.References[1]. WAsP Manual: Wind Analysis and ApplicationProgram (WAsP). Vol 2: Users Guide. RisøNational Laboratory, Roskilde, Denmark, ISBN87-550-178, 1993.[2]. Bowen, A. J. and Mortensen, Niels G. Exploringthe limits of WAsP the wind atlas analysis andapplication program. In Proceedings of EWEC-1996, Gøteborg, Sweden, 1996.[3]. Berge, E., Nyhammer, F.K., Tallhaug, L. andJakobsen, Ø. An evaluation of the WAsP modellat a coastal mountainous site in Norway, 2006.

Wind Energy, in press.[4]. WindSim documentation. http://windsim.com.[5]. Undheim, O. The non-linear microscale flowsolver 3DWind. Developments and validation.Ph.d. thesis at the Norwegian TechnicalUniversity, Trondheim, Norway. 2005.[6]. Eidsvik, K. J. A system for wind powerestimation in mountainous terrain. Prediction ofAskervein hill data. Wind Energy, 2005, 8, 237-249.[7]. Michalakes, J., S. Chen, J. Dudhia, L. Hart, J.Klemp, J. Middlecoff, and W. Skamarock:Development of a Next Generation RegionalWeather Research and Forecast Model.Developments in Teracomputing: Proceedings ofthe Ninth ECMWF Workshop on the Use of HighPerformance Computing in Meteorology. Eds.Walter Zwieflhofer and Norbert Kreitz. WorldScientific, Singapore, 2001, pp. 269-27.[8]. Benoit, R., Desgagné, M., Pellerin, P., Pellerin,S., Charter, Y. The Canadian MC2: a semi-Lagrangian, semi-implicit wide-bandatmospheric model suited for finescale processstudies and simulation. Monthly Weather Review1997, 125, 2382-2415.[9]. Pielke, R.A., Cotton, W.R., Walko, R.L.,Tremback, C.J., Lyons, W.A., Grasso, L.D.,Nicholls, M.E., Moran, M.D., Wesley, D.A., Lee,T.J. and Copeland, J.H. A comprehensivemeteorological modeling system–RAMS. Meteor.Atmos. Phys., 49, 1992, 69-91.[10]. Undheim, O. 2003. Comparison ofturbulence models for wind evaluation incomplex terrain. In preparation for Proceedingsof EWEC-2003 Madrid, Spain.[11]. Stull, R. B. An introduction to boundarylayer meteorology. Kluwer Academic Publisher,London, 1988.[12]. Aasen, Svein Erik. The Skipheia WindMeasurements Station. Instrumentation, WindSpeed Profiles and Turbulence Spectra. Ph.d.thesis, University of Trondheim, Department ofPhysics, AVH, Trondheim, Norway. 1995.[13]. Klemp, J.B., Skamarock, W.C. and DudhiaJ. Conservative split-explicit time integrationmethods for the compressible non-hydrostaticequations (see http://www.wrf-model.org).