6.陳瀅依 CWP2023_Turbine interaction models and validation_DNV（換pdf格式公開）.pdf

《6.陳瀅依 CWP2023_Turbine interaction models and validation_DNV（換pdf格式公開）.pdf》由會員分享，可在線閱讀，更多相關《6.陳瀅依 CWP2023_Turbine interaction models and validation_DNV（換pdf格式公開）.pdf（29頁珍藏版）》請在三個皮匠報告上搜索。

1、Turbine interaction models of different fidelity and their validations CWP 2023Yingyi Chen Renewable Advisory,Energy System17th Oct 2023DNV 17 OCTOBER 2023Turbine Interaction Models DNV 17 OCTOBER 2023Turbine interaction modelling history198019902000201020203Jensen/Katic(Park)Ainslie(Eddy viscosity)

2、DNV CFD turbine interaction modelWind tunnels+first wind farm validations10%power overpredictions at Horns Rev!(80 turbines)Large wind farm correction modelsDTU FugaBlockage effect!DNV BEET Blockage correction modelconcern about long range wind farm wakes between clusters growsJan 2022 Orsted releas

3、e TurbOParkAnsys CFD Wind-Modellermodel fidelityLESNon-linear RANS EngineeringLinearised RANSCommercial scale wind farms WakeBlasterWind farm design software released:WindFarmerWindProDNV 17 OCTOBER 2023Cost-fidelity relationship for turbine interaction models4Source:DNVDNV 17 OCTOBER 2023WRF-CFD va

4、lidation results DNV 17 OCTOBER 2023Cost-fidelity relationship for turbine interaction models6Source:DNVDNV 17 OCTOBER 2023Engagement with EnBW7Two successful measurement campaigns with EnBW dual-scanning lidar SCADA (met mast data)Combination of Wind farm blockage Cluster wakesCombination of WRF an

5、d CFD modelling Validation of Blockage(WESC 2021,WindEurope 2021)Blockage and cluster wakes(WESC 2023)Results publishedDNV 17 OCTOBER 2023 Focus on directions where the measurements along the lidar lines are only blockage affected Example from direction 210 Wind speed on the plateau of the thrust cu

6、rve Example on the right:%change in wind speed at hub height(for unstable conditions)Wind speed reduction(through blockage)seen upstream of the wind farm Measurements along lines A and BExample 1:BlockageSouth-Westerly wind directions 8Direction 210ABDNV 17 OCTOBER 2023Line West B and West A,All sta

7、bility,plateau of thrust curve9line West B30 sector:193-223,all stabilityline West A30 sector:193-223,all stabilityOver all stability(plateau of thrust curve),measurements capture 3%wind speed reduction between 4.8 km and 0.5 km,combination of WRF and CFD simulations get the magnitude rightDNV 17 OC

8、TOBER 2023Pattern of production(193-223),all stability10CFDSCADA Whole array normalised power well captured by CFD model.(Note:turbines with availability 85%not plotted in SCADA PoP plot).Leading turbines PoP:not a large variation across the line of leading turbinesCFD vs SCADA,arrayLeading turbines

9、DNV 17 OCTOBER 2023 Directions where the measurements along the lidar lines are affected by both blockage AND wakes from neighbouring clusters Example from direction 271 Wind speed on the plateau of the thrust curve Example on the right:%change in wind speed at hub height(for unstable conditions)Win

10、d speed reduction(through blockage)seen upstream of the wind farm Wake effects from upstream cluster dominate measurements along lidar lineExample 2:Cluster wakesWesterly wind directions11Direction 271DNV 17 OCTOBER 2023Measurements along line West B,All stability,plateau of thrust curve1230 sector(

11、253-283)Direction 271Measurements and CFD capture cluster wakes and blockage%change in wind speed at hub heightDNV 17 OCTOBER 2023Pattern of production(253-283),stable13CFDSCADACFD vs SCADA,array Whole array normalised power well captured by CFD Large variation across the line of leading turbineslow

12、est producing produces 33%less than highest producing turbine*Note:only applicable on plateau of thrust curve,will be less at higher wind speeds!Leading turbinesDNV 17 OCTOBER 2023DNVs fast CFD surrogate model,CFD.ML14DNV 17 OCTOBER 2023Cost-fidelity relationship for turbine interaction models15Sour

13、ce:DNVDNV 17 OCTOBER 2023What is CFD.ML?16CFDDNVs CFD RANS modelling of wind farm flows.The highest fidelity modeling applied at scale in wind farm energy production assessments.MLMachine learning model based on graph neural networks.CFD.MLA surrogate model for RANS CFD applied to turbine interactio

14、n modeling.Fast enough to use in wind farm optimization context.Captures flow physics better than engineering wake modelsno up-front tuning to SCADADNV 17 OCTOBER 202317windHypothetical Wind Farm The Bowl”110 turbines,7D spacingCFD.MLRANS CFDCan CFD.ML replicate CFD?Front row power variations Array

15、efficiencydifference driven partly by CFD.MLs ability to include blockage.1-2 hours on a HPC clusterFew seconds via the CFD.ML cloud APICFD.ML-performs well at approximating RANS CFD.-Is fast,may be used in an optimization context.DNV 17 OCTOBER 202318internal wakesexternal or cluster wakeswind farm

16、 blockage=Upstream wind farmYour wind farmInflow wind 270 15Validation against operational data-the three key aspects of a turbine interaction modelDNV 17 OCTOBER 202319Wind FarmS1M1L1L2XL1XL2ModelSEVM0.00.0-1.9-0.50.50.8CFD.ML-0.7-0.9-2.0-2.0-0.73.0wfEV 120D(newa)-0.8-1.9-0.90.3wfEV 120D(wti)-0.9-1

17、.8-1.7-0.9-2.00.3wfPARK-2.5-5.2-7.3-2.9-8.01.2Bias in power output for each time stampMean bias over time stamps NRelative mean bias over time stamps N=modelled measuredMBEP=1rMBEP=MBEP-6 offshore wind farms,validation focussed on internal wakes-CFD.MLs validation points to a slight overprediction o

18、f(internal)wakes,although validation framework uncertainty remains-CFD.MLs error spread is smaller than in the case of engineering models in 5/6 wind farms better predictions of production patternsValidation against operational data-internal wakesDNV 17 OCTOBER 202320NESWNESWValidation against opera

19、tional data-blockage-induced front-row power variationsFront row turbines subject to blockage induced speed-ups and slow-downsFront row turbinesWind farm blockage zoneDNV 17 OCTOBER 2023Validation against operational data-blockage-induced front-row power variations21Turbine positions along the front

20、-rowPower output relative to the mean of the row-A large offshore windfarm-Relative power output variations along the front row-Flowcases shown have a long fetch and no neighboring wakes.-20deg bins,no filter on stability in data-Boxes entail 50%of data,outer whiskers entail 90%of data,centerline is

21、 the median-The more data,the better the agreement.DNV 17 OCTOBER 202322NESWWake ZoneTurbine ATurbine BNESW110 RDWake ZoneTurbine ATurbine BValidation against operational data-cluster wakesDNV 17 OCTOBER 2023Validation against operational data-cluster wakes23Wind directionPower ratio of two corner t

22、urbines-CFD.ML stable does a better job.It is a 50/50 blend of predictions from a neutral-only gnn and a gnn trained on neutral&stable CFD sims.Its an experimental approach.-These observations agree with the(earlier discussed)ongoing refinements in the underlying CFD model.-CFD.ML neutral fails to c

23、apture the amplitude of the signal(underprediction of cluster wakes).-5 deg directional bins,1 m/s wind speed bin,no data filtering on stability-Boxes entail 50%of data,outer whiskers entail 90%of data,centerline is the medianDNV 17 OCTOBER 2023Validation studies for Classic Turbine Interaction Mode

24、ls DNV 17 OCTOBER 2023Cost-fidelity relationship for turbine interaction models25Source:DNVClassic turbine interaction modelsDNV 17 OCTOBER 2023Relative Mean Bias Error in modelled power26Bias in power output for each time stampMean bias over time stamps NRelative mean bias over time stamps N=modell

25、ed measuredMBEP=1rMBEP=MBEPWind Farm#Turbines:125 125S1M1L1L2XL1XL2ModelSEVM0.00.0-1.9-0.50.50.8CFD.ML-0.7-0.9-2.0-2.0-0.73.0wfEV 120D(newa)-0.8-1.9-0.90.3wfEV 120D(wti)-0.9-1.8-1.7-0.9-2.00.3wfPARK-2.5-5.2-7.3-2.9-8.01.227Results for multiple wake modelsAggregating all 48 casesMinMeanMax *No blocka

26、ge model/correction has been included.Errors are relative to the observed wake loss.Credit:Ewa Johansson,rsted,June 2023,“Benchmarking results from multiple wake models on operational data from offshore wind farms”,Wind Europe technical workshop18DNV 17 OCTOBER 2023Classic wake modelsNext generation

27、 turbine interaction models(CFD.ML)Summary Trade-off between computation cost and fidelity for turbine interaction models.Classic turbine interaction models provides a low-cost solution but with higher uncertainties.Use of higher fidelity models for cases outside of validation envelope to reduce unc

28、ertainties CFD.ML has the potential to become the next generation,fast-turnaround turbine interaction model when applied stand-alone in energy production assessments of wind farms.28Blockage treated separately with:flat,farm-level AEP corrections or dedicated blockage-only modelsWakes&blockage treated togetherCourtesy:Vattenfall,Christian SteinessThank you