• U.S. Department of Energy (DOE) goal of having 20% of the nation’s electrical energy from wind by 2030

• DOE workshop report states  “…1% error in wind speed estimates for a 100-MW wind generation facility can lead to losses approaching $12,000,000 over the lifetime of that plant ..."  

• To optimize wind for power generation, accurate weather forecasts are needed

• Better weather forecasts lead to greater confidence and more reliance on wind energy as a reliable energy source

• Typically, weather forecasting done for precipitation and temperature, not wind – until recently

• Meteorologists traditionally have focused wind forecasts at the 10 m level, a height strongly influenced by surface friction

• Prior wind forecasting research in the western United States has focused on flow in complex terrain (e.g. Wood 2000, Ayotte et al. 2001)

• Not applicable in Iowa where low-level jets and changing surface conditions are likely to be the dominant factors

• Statistical approach to predict wind speed at different levels (Huang et al. 1996, Kamal et al. 1997) – time series analysis

• With the increased growth in the wind energy sector, wind speed forecasts at turbine hub height (80 m) are now needed

• Due to the lack of observations, validating forecasts at this height has been difficult and little attention has been paid to wind forecasts at 80 m in the meteorological community

• In this study, an ensemble was created based on six different ways to represent drag as well as other forecasting techniques to improve wind speed forecasting at 80 m

Fd(z,t) V“PBL Scheme”

• PBL schemes were developed to help resolve the turbulent fluxes in the boundary layer – aka the drag force

• Due to the complex nature of boundary layer, very difficult to model – changes diurnally and seasonally

• Smaller BL in winter – snow pack

• To parameterize the planetary boundary layer - both local and non-local parameterizations are used

• Local closure - estimates unknown fluxes using known values and/or gradients at the same point (Stull 1988)

• Non-local closure - estimates unknown fluxes using known values and/or gradients at many points in space (Stull 1988)

Pomeroy Iowa Wind Farm

Weather forecast model tested• Weather Research and Forecasting (WRF) model

Model Forecast Period• 54 hour forecasts – first 6 hours not used - model spin-up

Six different ways to Represent Drag Force• Yonsei University Scheme (YSU) - WRF• Mellor-Yamada-Janjic (MYJ) - WRF• Quasi-Normal Scale Elimination PBL (QNSE) - WRF• Mellor-Yamada Nakanishi and Niino Level 2.5 PBL (MYNN2.5) - WRF• Mellor-Yamada Nakanishi and Niino Level 3.0 PBL (MYNN3.0) - WRF• Pleim PBL scheme (also called Asymmetric Convective Model (ACM2)) – WRF

• Driving model (Initial and lateral boundary conditions) - Regional and Global models that provide data to limited domain used in the WRF

• Global Forecast System (GFS) • North American Model (NAM)

• Observation data • 80 m meteorological tower on the southwest side of the Pomeroy, Iowa wind

farm• Data was taken at 10 minute increments and averaged over one hour periods

centered on each hour; to match model output

• Evaluation period• From June 2008 through September 2010, excluding periods where missing

data was observed

Perturbation Number

GFS MYNN 3.0 MAE ( ms-1)

GFS YSU MAE ( ms-1)

Ensemble MAE ( ms-1)

2 2.34 2.06 2.05

4 2.18 2.04 1.98

15 2.27 2.18 2.08

Time Initialization

GFS MYNN 3.0 MAE ( ms-1)

GFS YSU MAE ( ms-1)

Ensemble MAE ( ms-1)

18 UTC 1.88 1.78 1.69

00 UTC 1.82 1.74 1.63

06 UTC 1.83 2.07 1.73

Ensemble - best model skill

Time Initialization - Higher model skill (lower MAE) than

Perturbations

MAE of three different GFS perturbations using the YSU and MYNN3.0 PBL schemes from 10 cases in January 2010

MAE associated with the wind speed at 80 m from three different initialization times from 10 cases in January 2010

Day Ahead Market

Noon NoonMidnight

Day Ahead Market

Noon NoonMidnight

Day Ahead Market

Noon NoonMidnight

Larger the model spread – less confidence in forecast

Day Ahead Market

Noon NoonMidnight

Day Ahead Market

Noon NoonMidnight

Day Ahead Market

Noon NoonMidnight

Grid SpacingGFS MYNN 3.0 MAE

( ms-1)GFS YSU MAE

( ms-1)Ensemble MAE

( ms-1)

10 km 1.82 1.74 1.63

4 km 2.16 1.79 1.73

MAE of wind speed at 80 m from two different grid spacings (4 km and 10 km) from 10 cases in January 2010

Highest Model skill associated with 10

km grid spacing

Computing power limited in most

private companies, running 10 km model runs are

much more feasible than 4 km runs

Day Ahead Market

Noon NoonMidnight

Day Ahead Market

Noon NoonMidnight

Similar model skill – with terrain effects, like mountains – results would be much different

Day 2 Picked ensemble best MAEDay 2 Non-Picked ensemble

best MAEDay 2 All Member Ensemble

best MAE

5/15 4/15 6/15

Training of the model based on day 1 results15 cases from June 2008 to May 2009

Training approach was not a reliable method to predict wind speed as conditions change from day to day

Picked Ensemble – showed best model skill

only 33% of time

Non-Picked Ensemble – showed best model skill

27% of time

Model Number

00 UTC MYJ GFS - 10 km grid spacing

00 UTC MYJ NAM - 10 km grid spacing

00 UTC Pleim NAM - 10 km grid spacing

00 UTC Pleim GFS - 10 km grid spacing

00 UTC YSU NAM - 10 km grid spacing

00 UTC YSU GFS - 10 km grid spacing

• Bias corrections developed from 30 cases (all seasons) from June 2008 to Jan. 2010• Applied to Case study from Oct. 11, 2009 to Nov. 11, 2009• Wind Speed bias correction seen as best way to improve forecast (green box)• Non bias correction showed worst results (red box)

Day Ahead Market

MidnightNoon Noon

Model over-predictionNighttime

Model under-predictionDaytime

MAE MYJ (m/s)

MYNN 2.5 (m/s)

MYNN 3.0 (m/s)

Pleim (m/s)

QNSE (m/s)

YSU (m/s)

Ensemble (m/s)

GFS 00Z 1.59 1.66 1.66 1.52 1.65 1.57 1.48

GFS 18Z 1.68 1.81 1.72 1.61 1.77 1.63 1.58

NAM 00Z 1.67 1.71 1.69 1.63 1.71 1.57 1.56

NAM 18Z 1.66 1.75 1.74 1.60 1.70 1.63 1.57

Green boxes show highest model skill

Member Number

PBL Scheme

Time Initialization

Land Surface Scheme Land Layer Scheme Initial Boundary

Conditions

1 ACM2 18 UTC Pleim-Xiu Pleim-Xiu GFS

2 ACM2 18 UTC Pleim-Xiu Pleim-Xiu NAM

3 ACM2 00 UTC Pleim-Xiu Pleim-Xiu GFS

4 YSU 00 UTC Noah Monin-Obukhov NAM

5 YSU 00 UTC Noah Monin-Obukhov GFS

6 MYJ 00 UTC Noah Janjic Eta Monin-Obukhov GFS

• Bias corrections for 00Z and 18Z time initializations and NAM and GFS initial boundary conditions over a period from Aug. 14-28, 2009

• Six schemes that showed best model skill (lowest MAE) formed operational model

• Five out of six scheme either YSU or Pleim – both non-local turbulent closure schemes• Four out of six scheme use the 00Z time initialization• Four out of six scheme use the GFS initial boundary conditions

EnsembleMAE after Bias

Correction (m/s)MAE Prior to Bias Correction (m/s)

Standard Deviation after Correction (m/s)

GFS 00Z 1.67 1.99 0.74

GFS 18Z 1.66 2.05 0.80

NAM 00Z 1.68 1.91 0.67

NAM 18Z 1.70 1.93 0.73

Deterministic Forecast

1.70 1.77 - - -

Operational Model

1.52 1.67 0.98

• Tested over 25 cases during the summer and fall of 2010• Best model skill seen in Operational Model after wind speed bias correction (Green Box)• Largest standard deviation (measure of model spread) in operation model ensemble• Deterministic forecast is the best individual model found from the period studied

• Ramp event - changes in wind power of 50% or more of total capacity in four hours or less (Greaves et al. 2009)

• Approximated using a typical wind turbine power curve such that any wind speed increase or decrease of more than 3 ms-1 within the 6-12 ms-1 window in four hours or less was considered a ramp

• Fifty eight cases spanning 116 days from June 2008 through June 2009 were validated – Models all used GFS initial boundary conditions

Sensitive Area – Ramp events not important above or below this area

PBL Scheme MYJ MYNN 2.5 MYNN 3.0 Pleim QNSE YSU Obs

Ramp-up 23 29 27 19 26 16 35

Ramp-down 23 28 21 14 28 13 31

Total Ramp Events

46 57 48 33 54 29 66

Number of ramp events during Day 1 for GFS initial boundary conditions (06-30 hours after model start up) PBL Scheme MYJ MYNN 2.5 MYNN 3.0 Pleim QNSE YSU Obs

Ramp-up 17 25 24 17 26 11 37

Ramp-down

19 22 16 20 23 11 35

Total Ramp Events

36 47 40 37 49 22 72

Number of ramp events during Day 2 for GFS initial boundary conditions (30-54 hours after model start up)

All PBL schemes under-predict

number of ramp events

Fewer Ramp events forecasted on Day 2

Event was considered a ramp event if change in wind power was 50% or more of total capacity in four hours or less - wind speed increase or decrease of more than 3 m/s within the 6-12 m/s

PBL Scheme MYJ (ms-1)MYNN 2.5

(ms-1)MYNN 3.0

(ms-1)Pleim (ms-1)

QNSE (ms-1) YSU (ms-1) Obs (ms-1)

Ramp-up(Day 1)

4.50 4.62 4.75 4.85 4.60 4.67 4.53

Ramp-up(Day 2)

4.54 5.16 5.2 4.56 4.69 4.73 4.01

Ramp-down(Day 1)

3.74 4.62 4.20 4.60 4.31 4.17 4.34

Ramp-down(Day 2)

3.83 4.28 4.46 4.27 4.59 4.43 4.21

• Amplitude was over-predicted by all six PBL schemes for ramp-up events• Obs. show on average over 4m/s ramp event• If ramp-up event occurred at 6m/s and went to 10m/s within 4hrs• – Power increase from 216 kW to 1000 kW

Average amplitude of ramp up and ramp down events for GFS initial boundary conditions

Sensitive Area – Ramp events not important above or below this area

• Three hour averaged diurnal cycle of ramp-up events using the midpoint of the ramp event

• Peak at 01Z – LLJ related• Peak at 16Z – Growth of BL

• Three hour averaged diurnal cycle of ramp-down events using the midpoint of the ramp event

• Less noticeable trend

• To forecast winds, we have focused on drag (six different schemes)

• However, explored other methods to improve forecasts– Perturbations of GFS model

• Low skill

– Varied Time Initializations • High skill

– Grid Spacing• Little difference as terrain is flat – less computing power with 10km

– Training of the Model• Not a reliable method as conditions change from day to day

– Bias Correction• Noticed a diurnal bias in the model data• Investigated whether other biases existed – wind speed bias correction

• Combination of techniques yields a model that is significantly more skillful

• All six PBL schemes tested underestimated the number of ramp-up and ramp-down events

• Average ramp-up events around 4m/s increase– Increase in power produced from 216 kW to 1000 kW– For example, caused a blackout in Texas

• Modeled ramp-up events occurred most often between 22 UTC and 01 UTC - closely matched observed ramp-up events (occurred most frequently around 01 UTC)

• LLJs - first described in the late 1930’s by Goualt (1938) and Farquharson (1939)

• Areas of relatively fast-moving winds in the lower atmosphere, LLJs were first studied because of their roll in transporting warm, moist air from the Gulf of Mexico into the Great Plains, leading to convective events (Stensrud 1996)

• Most well-known LLJs occur over the Great Plains of the United States, although found around the world - Europe, Africa, and Australia (Stensrud 1996)

• Maximum winds during nocturnal LLJ events over the Great Plains are between 10 ms-1 and 30 ms-1 (Whiteman et al. 1997)

•Whiteman (1997) classified two years of LLJs in northern Oklahoma and discovered that LLJs occur:

– 47% of the time during the warm season – 45% of the time during the cold season

•Whiteman (1997) found approximately 50% of the maximum wind speeds during LLJ events occurred less than 500 m above the surface

•With the potential for wind turbine hub heights to increase from 80 m to 120 m or higher, LLJ interaction with wind turbines could largely affect the power performance of wind farms (Schwartz and Elliot, 2005).

• Few studies have examined the performance of forecasting models during LLJ events, or the sensitivity to how surface drag is represented in models

• In this study, the ability of the WRF model to accurately reproduce vertical wind structure during LLJ events was evaluated using six different drag schemes in the WRF model to observations from the Lamont, OK wind profiler site

• Observed data from the U. S. Department of Energy ARM project located at Lamont, OK.

• 915-MHz wind profiler - measure wind speeds below 500 m, unlike the NOAA 404-MHz profilers

• Observed data ranged from 96 m - 2462 m above the surface; vertical resolution of 60 m

• 30 LLJ cases and 30 non-LLJ cases were chosen between June 2008 and May 2010

• Same forecasts model was used as in Pomeroy – different dates, different heights evaluated, different location

MYJ (ms-1)

Pleim (ms-1)

YSU (ms-1)

QNSE (ms-1)

MYNN2.5 (ms-1)

MYNN3.0 (ms-1)

OBS (ms-1)

Maximum LLJ Wind

Speed19.0 18.2 16.3 19.1 18.2 17.9 22.7

Ho: ; Ha:

MYJ MYNN 2.5 MYNN 3.0 Pleim QNSE

P-value < 0.0001 < 0.0001 0.0002 < 0.0001 < 0.0001

All PBL schemes under-predict max

LLJ wind speed

Lowest predicted max LLJ wind speed

occurred in YSU scheme

Average Maximum LLJ Wind Speed over 30 LLJ events for GFS initial boundary conditions

Null hypothesis - difference between the maximum LLJ wind speed of the YSU scheme (u1) will be equal to the maximum LLJ wind speed of the other PBL schemes (u2)

P-values of the YSU PBL scheme vs. other PBL schemes for maximum LLJ wind speed Null hypothesis was

rejected in favor of the alternative hypothesis,

indicating that the under-prediction of the wind speed in the YSU PBL

scheme was highly significant

Ho: ; Ha:

MYJ MYNN 2.5 MYNN 3.0 Pleim QNSE

P-value < 0.0001 0.0005 < 0.0001 0.0007 0.0038

All PBL schemes under-predict height

of LLJ max

Highest predicted height of LLJ max occurred in YSU

scheme

Average Maximum LLJ Wind Speed over 30 LLJ events for GFS initial boundary conditions

Null hypothesis - difference between the height of LLJ maximum of the YSU scheme (u1) will be equal to the

height of the LLJ maximum of the other PBL schemes (u2)

P-values of the YSU PBL scheme vs. other PBL schemes for height of LLJ maximum Null hypothesis was

rejected in favor of the alternative hypothesis,

indicating that the higher height predicted by the

YSU PBL scheme was highly significant

MYJ (m)

Pleim (m)

YSU (m)

QNSE (m)

MYNN2.5 (m)

MYNN3.0 (m)

OBS (m)

Height of LLJ Maximum

371.2 427.0 538.3 344.5 365.3 340.3 553.0

Under-prediction of wind speed maximum

in YSU Scheme

LLJ structure not present in YSU

scheme

LLJ feature present in all other PBL schemes

YSU scheme - eddy viscosity value five

times larger than any other scheme

Larger eddy viscosity - more mixing and

turbulence

Higher speeds occurred above and below the jet core,

with higher momentum air being mixed closer to the

surface – resulting in substantially weaker

LLJ with a higher elevation of the

maximum

• All six PBL schemes showed the maximum LLJ wind speed occurring near or just after midnight

• Observed maximum LLJ wind speeds occurred a little later, with dual peaks at 08 UTC (2 am LST) and 10 UTC (4 am LST)

• Overall, the PBL schemes appeared to predict the timing of the peak LLJ occurrence reasonably well with perhaps a small early bias

LLJ (ms-1) Non-LLJ (ms-1)

96 m Wind Speed 6.3 5.1

157 m Wind Speed 9.9 7.8

Speed Shear 5.58 2.91

LLJ vs. Non-LLJ event comparison96 m: Wind speed 1.2 m/s stronger during LLJ events – power increase

of 117 kW

157 m: Wind speed 2.1 m/s stronger during LLJ events – power increase

of 496 kW

Speed shear – difference between 157 m and 96

m wind speed - is almost double during LLJ

events• Number of kW to power an average home per day = 50-70 kW• Speed shear is present at current hub height (80m)• Project this summer will focus on speed shear from 10m to 80m

and from 40m to 120m (entire reach of wind turbine blades)

96 mMYJ (ms-1)

MYNN 2.5 (ms-1)

MYNN 3.0 (ms-1)

Pleim (ms-1)

QNSE (ms-1)

YSU (ms-1)

Wind Speed Bias

3.26 4.38 4.14 2.45 3.15 1.80

Wind Speed MAE

4.52 5.48 5.28 4.04 4.38 3.34

All six PBL schemes over-predicted the wind speed

during LLJ events

YSU scheme showed the lowest MAE, while the

highest was observed with the MYNN 2.5 scheme

Bias and MAE associated with 96 m wind speed forecasts during LLJ events for GFS initial boundary conditions

157 mMYJ (ms-1)

MYNN 2.5 (ms-1)

MYNN 3.0 (ms-1)

Pleim (ms-1)

QNSE(ms-1)

YSU (ms-1)

Wind Speed Bias

3.20 3.66 3.40 1.87 3.15 0.15

Wind Speed MAE

3.80 3.94 3.69 3.14 3.72 2.09

All six PBL schemes over-predicted the wind speed

during LLJ events – although YSU small

positive bias

YSU scheme showed the lowest MAE, while the

highest was observed with the MYNN 2.5 scheme

Bias and MAE associated with 157 m wind speed forecasts during LLJ events for GFS initial boundary conditions

• LLJ maximum wind speeds were under-predicted by all PBL schemes - largest under-prediction occurred with the YSU scheme – larger drag present

• All the PBL schemes except the YSU scheme under-predicted the height of the LLJ maximum by more than 125 m

• YSU scheme - likely cause of the under-predicted wind speed and higher jet elevation - result of the strong eddy viscosity occurring during stable conditions

• Increased mixing - LLJs in the YSU scheme were substantially under-predicted and momentum was spread out over a deeper layer of the atmosphere

• Timing or temporal trends of the LLJ maximum - models had wind speed maxima occurring near or just after midnight (06-08 UTC), typically a few hours before observed LLJs (08-10 UTC)

• LLJ impacts at 96 m and 157 m - increased wind speeds and speed shear during LLJ events compared to non-LLJ events

• Implies that wind production would increase during LLJ events however, wind turbine durability would need to be improved to accommodate the increased shear

• At 96 and 157 m, the YSU PBL scheme showed significantly better skill (lower MAE) than the other schemes

• Non-local PBL schemes appear to show better model skill overall, however no one scheme is the answer for predicting low level winds

• Example - YSU– High model skill at predicting 80 m wind speeds – High model skill during LLJ events at 96 m and 157 m– Ramp events are poorly predicted– LLJ max wind speeds are significantly under-predicted

• As a result, no one scheme performed considerably better than any other and all showed room for improvement.

Questions?

References• Arribas A., K. B. Robertson, and K. R. Mylne, 2005: Test of a poor man’s ensemble prediction system for short-range probability forecasting. Mon. Wea. Rev.,

133, 1825–1839.• Ayotte, K.W., R.J. Davy and P.A. Coppin, 2001: A simple temporal and spatial analysis of flow in complex terrain in the context of wind energy modeling. Bound.-

Layer Meteorol., 98, 275–295.• Bélair, Stéphane, Jocelyn Mailhot, J. Walter Strapp, J. Ian MacPherson, 1999: An examination of Local versus Nonlocal Aspects of a TKE-Based Boundary Layer

Scheme in Clear Convective Conditions. J. Appl. Meteor., 38, 1499–1518.• Bowler, N. E., A. Arribas, and K. R. Mylne, 2008: The benefits of multianalysis and poor man’s ensembles. Mon. Wea. Rev., 136, 4113–4129.• Bradford, K. T., R. L. Carpenter, and B. Shaw, 2010: Forecasting Southern Plains wind ramp events using the WRF model at 3km. Ninth Annual Student

Conference, Atlanta, GA, Amer. Meteor. Soc., [Available online at http://ams.confex.com/ams/90annual/techprogram/paper_166661.htm].• Dalcher, A., E. Kalnay, and R. N. Hoffman, 1988: Medium range lagged average forecasts. Mon. Wea. Rev., 116, 402–416.• Department of Energy, 2008: 20% wind energy by 2030. Energy Efficiency and Renewable Energy Rep. DOE/GO-102008-2567, 1-2.• Schreck, S, J. Lundquist, and W. Shaw, 2008: U.S. Department of Energy Workshop Report—Research needs for wind resource characterization. NREL Rep. TP-

500-43521, 81-82.• Shin, H.H., S. Hong, 2011: Intercomparison of Planetary Boundary-Layer Parametrizations in the WRF Model for a Single Day from CASES-99. Boundary-Layer

Meteor., 139, 261-281.• Stensrud, David J., Jian-Wen Bao, Thomas T. Warner, 2000: Using Initial Condition and Model Physics Perturbations in Short-Range Ensemble Simulations of

Mesoscale Convective Systems. Mon. Wea. Rev., 128, 2077–2107.• Stull, R.B., 1998: An Introduction to Boundary Layer Meteorology. Boston, Massachusetts: Kluwer Academic Publishers. 197-242 pp.• Sukoriansky S, Galperin B, Perov V (2005) Application of a new spectral theory of stable stratified turbulence to the atmospheric boundary layer over sea ice.

Boundary-Layer Meteorol. 117, 231–257.• Theis, S. E., A. Hense and U. Damrath, 2005: Probabilistic precipitation forecasts from a deterministic model: a pragmatic approach. Meteorological

Applications, 12, 257-268.• Walser, A., D. Lüthi, and C. Schär, 2004: Predictability of precipitation in a cloud-resolving model. Mon. Wea. Rev., 132, 560–577.• Wei, M., Z. Toth, R.Wobus, Y. Zhu, C. H. Bishop and X. Wang 2006: Ensemble Transform Kalman Filter-based ensemble perturbations in an operational global

prediction system at NCEP, Tellus 58A, 28-44.• Whitaker, Jeffrey S., Andrew F. Loughe, 1998: The Relationship between ensemble Spread and Ensemble Mean Skill. Mon. Wea. Rev., 126, 3292–3302.• Wilks, D. S., 1995: Statistical Methods in the Atmospheric Sciences. Academic Press, 467.• Wood, N., 2000: Wind flow over complex terrain: A historical perspective and the prospect for large-eddy modeling. Bound.-Layer Meteor., 96, 11–32.• Xiu, A., and J. E. Pleim, 2001: Development of a land surface Model. Part I: Application in a mesoscale meteorological model. J. Appl. Meteor., 40, 192–209.• Zack, John W., 2007: Optimization of Wind Power Production Forecast Performance During Critical Periods for Grid Management. AWEA Windpower 2007.