The Analog Ensemble for Renewable Energy Applications: an Overview
Luca Delle Monache1 Stefano Alessandrini1, Guido Cervone2, Constantin Junk3, Daran Rife4, Jessica Ma4, Simone Sperati5, Sue Ellen Haupt1, Thomas Brummet1, Paul Prestopnik1, Gerry Wiener1, Jesper Nissen6, Sam Hawkins6
1National Center for Atmospheric Research − Boulder, CO, USA 2The Pennsylvania State University − State College, PA, USA 3ForWind, Car von Ossietzky University Oldenburg − Oldenburg, Germany 4DNV GL − USA 5RSE − Milan, Italy 6Vattenfall − Sweden
rd 3 International Conference Energy & Meteorology − Boulder, CO, USA, 25 June 2015 1 Acknowledgments
• COLLABORATORS o Roland Stull (University of British Columbia) o Lueder von Bremen, Detlev Heinemann (ForWind, Carl von Ossietzky University) o Iris Odak, Kristian Horvath (Meteorological and Hydrological Service of Croatia) o Badrinath Nagarajan (IBM) o Federica Davo’ (RSE) o Tony Eckel (Microsoft) o Thomas Nipen (Norwegian Meteorological Institute) o Laura Harding (The Pennsylvania State University) o Irina Djalalova, Jim Wilczak (NOAA) o Emilie Vanvyve (UK Met Office) o Jens Madsen (Vattenfall) o Christopher Rozoff, Will Lewis (University of Wisconsin, Madison) o Jan Keller (DWD, University of Bonn)
• SPONSORS o U.S. Department of Defense (DOD) Army Test and Evaluation Command (ATEC) o U.S. DOD Defense Threat Reduction Agency (DTRA) o U.S. Department of Energy (DOE) o U.S. National Aeronautics and Space Administration (NASA) o U.S. National Renewable Energy Laboratory (NREL) o U.S. National Oceanic and Atmospheric Administration (NOAA) Hurricane Forecast Improvement Program (HFIP) o Vattenfall o Vestas Wind Systems o Xcel Energy
2 Outline
• Analog Ensemble (AnEn) basic idea • AnEn successful applications AnEn for short-term (i.e., 0-72 h) wind power predictions AnEn for short-term (i.e., 0-72 h) solar GHI predictions AnEn for long-term (i.e., multi-year) wind resource assessment • Summary of recurring features of AnEn
3 AnEn has been successfully applied for: • Short-term predictions of: 10- and 80-m wind speed, 2-m temperature, etc. Wind power Solar GHI Load • Downscaling, resource assessment: Wind speed Computationally efficient dynamical downscaling
4 AnEn has been successfully applied for: • Short-term predictions of: 10- and 80-m wind speed, 2-m temperature, etc. Horvath (DHMZ) et al., earlier today Wind power Lueder von Bremen (ForWind) et al., earlier today Solar GHI
Load Alessandrini (NCAR) et al., next talk • Downscaling, resource assessment:
Wind speed Nissen (Vattenfall) et al., Tuesday Computationally efficient dynamical downscaling Daran Rife (DNV GL) et al., Tuesday
5 AnEn has been successfully applied for: • Short-term predictions of: 10- and 80-m wind speed, 2-m temperature, etc. Wind power Solar GHI Load • Downscaling, resource assessment: Wind speed Computationally efficient dynamical downscaling
6 Analog Ensemble (AnEn)
Mon Tue Wed Thu Fri Sat Sun t = 0 Time
7 Analog Ensemble (AnEn)
PRED
Mon Tue Wed Thu Fri Sat Sun t = 0 Time
8 Analog Ensemble (AnEn)
PRED OBS
Mon Tue Wed Thu Fri Sat Sun t = 0 Time
9 Analog Ensemble (AnEn)
PRED OBS
Mon Tue Wed Thu Fri Sat Sun t = 0 Time
10 Analog Ensemble (AnEn)
PRED OBS
Mon Tue Wed Thu Fri Sat Sun t = 0 Time
11 Analog Ensemble (AnEn)
PRED
Mon Tue Wed Thu Fri Sat Sun t = 0 Time
12 Analog Ensemble (AnEn)
PRED OBS
Mon Tue Wed Thu Fri Sat Sun t = 0 Time
Wed Fri Sat Tue Sun Mon Thu farthest closest analog analog “Analog” Space
Delle Monache et al. (MWR 2011, 2013) 13 Analog Ensemble (AnEn)
PRED OBS
Mon Tue Wed Thu Fri Sat Sun t = 0 Time
Wed Fri Sat Tue Sun Mon Thu farthest closest analog analog “Analog” Space
14 Analog Ensemble (AnEn)
PRED OBS
Mon Tue Wed Thu Fri Sat Sun t = 0 Time
2-member AnEn
Wed Fri Sat Tue Sun Mon Thu farthest closest analog analog “Analog” Space
15
Power predictions: Experiment design
• Test site: Wind farm in northern Sicily − 9 turbines, 850 kW Nominal Power (NP) • Training period: November 2010 - October 2012 • Verification period: November 2011 – October 2012 • Probabilistic prediction systems: ECMWF EPS, COSMO LEPS, AnEn
Alessandrini et al. (RE 2015) 16 RMSE of ensemble means
Analog Ensemble (AnEn) Mean European Center for Medium range Weather Forecasting (ECMWF) Ensemble Mean
17 SMUD observations data set
• 8 stations with available global horizontal irradiance (GHI) observations • GHI data for the 226-day period, from 04-25-2014 to 12-03-2014) • Average missing data: ~ 9% • So far only 3 stations are used (those with less QC issues) Alessandrini et al. (AE 2015) 18
Results: Time series example: AnEn, DICAST
Station SMUD 67, forecast initialized at 12 UTC, 20 September 2014
0
0
9
0 0
7 5- 95%
)
2 25-75%
− An-En Mean
m
0
0
W
(
5
I
H
G
0
0
3
0
0
1 0
0 3 6 9 13 17 21 25 29 33 37 41 45 49 53 57 61 65 69
Lead Time (Hour) 19 AnEn vs. DICAST (Brier Skill Score)
• The Brier Skill Score (BSS) compares the skill of two probabilistic predictions for a given event • Values > 0 indicate that AnEn has more skill than DICAST • Shown is the BSS index as a function of forecast lead time, with the event considered being GHI > mean(obs. GHI at the given lead time)
0
8 0
6 0
4 Conclusion: AnEn is nearly
0 )
2
AnEn better AnEn always better than the already
%
(
0
S optimized DICast forecast, even
1
S −
B with this short training and 0
4 validation period −
0
7
−
0
AnEn worse AnEn
0
1 − 1 5 9 14 20 26 32 38 44 50 56 62 68
Lead Time (Hours) 20 AnEn for wind resource assessment
• Recreate a long-term observation-based wind climatology at site • Downscale a long-term NASA Modern-Era Retrospective Analysis for Research and Applications (MERRA) time series using a short-term record of observations
Observed wind speed
MERRA wind speed
1992 19 years to downscale 2010 1 year where 2011 MERRA/obs. overlap
Vanvyve et al. (RE 2015), Zhang et al. (AE 2015) 21 AnEn for wind resource assessment
• Recreate a long-term observation-based wind climatology at site • Downscale a long-term NASA Modern-Era Retrospective Analysis for Research and Applications (MERRA) time series using a short-term record of observations
Observed wind speed
MERRA wind speed
1992 19 years to downscale 2010 1 year where 2011 MERRA/obs. overlap
22 AnEn for wind resource assessment
• Recreate a long-term observation-based wind climatology at site • Downscale a long-term NASA Modern-Era Retrospective Analysis for Research and Applications (MERRA) time series using a short-term record of observations
Observed wind speed
MERRA wind speed
1992 19 years to downscale 2010 1 year where 2011 MERRA/obs. overlap
23 AnEn for wind resource assessment
• Recreate a long-term observation-based wind climatology at site • Downscale a long-term NASA Modern-Era Retrospective Analysis for Research and Applications (MERRA) time series using a short-term record of observations
Observed wind speed
MERRA wind speed
1992 19 years to downscale 2010 1 year where 2011 MERRA/obs. overlap
24 AnEn for wind resource assessment
• Recreate a long-term observation-based wind climatology at site • Downscale a long-term NASA Modern-Era Retrospective Analysis for Research and Applications (MERRA) time series using a short-term record of observations
Observed wind speed
MERRA wind speed
1992 19 years to downscale 2010 1 year where 2011 MERRA/obs. overlap
25 AnEn for wind resource assessment
• Recreate a long-term observation-based wind climatology at site • Downscale a long-term NASA Modern-Era Retrospective Analysis for Research and Applications (MERRA) time series using a short-term record of observations
Analog ensemble Observed (deterministic + probabilistic) wind speed
MERRA wind speed
1992 19 years to downscale 2010 1 year where 2011 MERRA/obs. overlap
26 Downscaled wind tower locations
100m & 50m 27 Hourly wind speed Hourly wind speed Wind speed at Boulder, CO: Wind speed at Butler: WindHistogram distributions comparisonHistogram
0.20 Boulder, CO Observations Butler, WA Observations Training: 1 year MERRA Training: 1 year MERRA Downscaling: 3 years Downscaling: 5 years Analog ensemble 0.12 Analog ensemble 0.16
y 0.09 y
t Hourly wind speed
t
i i
s Outlier s Outlier
0.12 n Wind speed at Butler: n
e th e th 75 75 d
d Histogram
Inside Inside
y Mean y
Mean t t i inner i th
inner th l
l i i 50 50
b fence b fence th th a 25
a 25 b
b 0.06
o o
r Outlier
r 0.08 Outlier
P P
Observations MERRA Hourly wind speed Hourl0y .w0in3d speed 0.04 Wind speed at Goodnoe Hills, WA: 0.12 Wind speed at Utah: Analog ensemble Histogram Histogram
0.00 0.00
y 0.09
0 1 2 3 4 5 6 7 8 9 10 11121314151617181920 2122232425 30 0 1 2 3 4 5 6 t 7 8 9 10 11121314151617181920 2122232425 30 i
-1 s -1 Outlier 0.16 Wind speed (m s ) n Wind speed (m s ) e 75th d Observations Observations Inside
y Mean t
i inner th l MERRA Training period: last 365 days, period downscaled: last 3 entire yeMaErsR, RanAalogs: 25 0.16 Training period: i last 365 days, period downscaled: last 5 entire years, analogs: 25 50 b fence th Goodnoe Hills, WAAna log ensemble a Utah Prison, UT Analog ensemble 25 b 0.06
Training: 1 year o Training: 1 year r Outlier
Training: 2 Jan 2012 to 31 Dec 2012, period: 2 Jan 2009 to 1 Jan 2012, trend: 4 neDownscaling:ighbors, window: 0 nei g3hb yearsor, var: w s.wd Training: 4 Aug 2007 to 2 Aug 2008, peP riod: 5 Aug 2002 to 3 Aug 2007, trend: Downscaling:4 neighbors, window: 0 1 n eyearsighbor, v ar: ws.wd 0.12
0.12
y
y
t t
i 0.03 i
s Outlier
s Outlier
n n
e th e th 75
75 d
d
Inside Inside
y Mean y
Mean t t
i inner th i
inner th l l 0.08 i 50 i 50
b fence b fence th th a 25
a 25 0.08
b b
o 0.00 o
r Outlier
r Outlier P
P 0 1 2 3 4 5 6 7 8 9 10 11121314151617181920 2122232425 30 Wind speed (m s-1 )
0.04 0.04 Training period: last 365 days, period downscaled: last 5 entire years, analogs: 25
Training: 4 Aug 2007 to 2 Aug 2008, period: 5 Aug 2002 to 3 Aug 2007, trend: 4 neighbors, window: 0 neighbor, var: ws.wd
0.00 0.00 0 1 2 3 4 5 6 7 8 9 10 11121314151617181920 2122232425 30 0 1 2 3 4 5 6 7 8 9 10 11121314151617181920 2122232425 30 -1 Wind speed (m s-1 ) Wind speed (m s ) 28
Training period: last 365 days, period downscaled: last 3 entire years, analogs: 25 Training period: last 365 days, period downscaled: last 1 entire year, analogs: 25
Training: 1 Jan 2009 to 31 Dec 2009, period: 2 Jan 2006 to 31 Dec 2008, trend: 4 neighbors, window: 0 neighbor, var: ws.wd Training: 6 Jun 2005 to 5 Jun 2006, period: 6 Jun 2004 to 5 Jun 2005, trend: 4 neighbors, window: 0 neighbor, var: ws.wd AnEn recurring features • Reduction of systematic and random errors and increase of the correlation between the predictions and the observations with respect to the deterministic prediction used to generate AnEn
• The possibility of using of a higher resolution model (since only one deterministic forecast is needed to generate an ensemble), or to drastically reduce the computational burden
• No need for initial condition and model perturbation strategies to generate an ensemble
• Intrinsically reliable estimates (i.e., no postprocessing required)
• Ability to capture the atmospheric flow-dependent error characteristics
• A superior skill in predicting rare events when compared to state-of- the-science post-processing methods and to traditional ensemble
methods based on a set of numerical weather predictions 29 Thanks! ([email protected]) References 1. Delle Monache, L., T. Nipen, Y. Liu, G. Roux, and R. Stull, 2011: Kalman filter and analog schemes to postprocess numerical weather predictions. Mon. Wea. Rev., 139, 3554–3570 2. Delle Monache, L., T. Eckel, D. Rife, and B. Nagarajan, 2013: Probabilistic weather prediction with an analog ensemble. Mon. Wea. Rev., 141, 3498–3516 3. Mahoney, W.P., K. Parks, G. Wiener, Y. Liu, W.L. Myers, J. Sun, L. Delle Monache, T. Hopson, D. Johnson, S.E. Haupt, 2012: A wind power forecasting system to optimize grid integration. IEEE Trans. Sustainable Energy, 3, 670–682 4. Alessandrini, S., Delle Monache, L., Sperati, S., and Nissen, J, 2015. A novel application of an analog ensemble for short-term wind power forecasting. Renewable Energy, 76, 768-781 5. Vanvyve, E., Delle Monache, L., Rife, D., Monaghan, A., Pinto, J., 2015. Wind resource estimates with an analog ensemble approach. Renewable Energy, 74, 761-773 6. Nagarajan, B., Delle Monache, L., Hacker, J., Rife, D., Searight, K., Knievel, J., and Nipen, T., 2015. An evaluation of analog-based post-processing methods across several variables and forecast models. Conditionally accepted, Weather and Forecasting 7. Djalalova, I., Delle Monache, L., and Wilczak, J., 2015. PM2.5 analog forecast and Kalman filtering post- processing for the Community Multiscale Air Quality (CMAQ) model. Accepted to appear on Atmospheric Environment 8. Junk, C., Delle Monache, L., Alessandrini, S., von Bremen, L., and Cervone, G., 2015. Predictor-weighting strategies for probabilistic wind power forecasting with an analog ensemble. Accepted to appear on Meteorologische Zeitschrift 9. Alessandrini, S., Delle Monache, L., Sperati, S., and Cervone, G., 2015. Solar forecasting with an analog ensemble. Conditionally accepted, Applied Energy 10.Eckel, T., and Delle Monache, L., 2015. A hybrid, analog-NWP ensemble. Conditionally accepted, Monthly Weather Review 11.Zhang, J., Draxl, C., Hopson, T., Delle Monache, L., and Hodge, B.-M., 2015. Comparison of deterministic and probabilistic wind resource assessment methods on numerical weather prediction. Conditionally
accepted, Applied Energy 30 The metric (1)
Analog strength for a particular forecast lead time t is measured by the distance between current and past forecast, over a short window, 2 t wide +t 1 2 f : Forecasts’ standard deviation over ft - gt¢ = ( ft+k - gt¢+k ) å entire analog training period s f k=-t
Expanded to multiple predictor variables, but still focused on predictand f: (for wind speed, predictors are speed, direction, sfc. temp., sfp pressure, and PBL depth)
Nv +t w v v 2 v Nv : Number of predictor variables ft - gt¢ = å å ( ft+k - gt¢+k ) wv : Weight given to each predictor v=1 sf v k=-t
Current Forecast, f Past Forecast, g t+1
‘
WindSpeed t ‘ t‘1 t +1 t t1
0 1 2 3h 0 1 2 3h 31 The metric (2)
After finding the n strongest analogs, each of the n AnEn members is taken as the verifying observation from each analog.
Current Forecast, f Past Forecast, g t+1
WindSpeed t t+1 t1 t observation t1 AnEn member #7
0 1 2 3h 0 1 2 3h
32 How skillful is AnEn? • AnEn generated with Environment Canada GEM (15 km), 0-48 hours • Comparison with:
o Environment Canada Regional Ensemble Prediction System (REPS, next slide)
o Logistic Regression (LR) out of 15-km GEM
o LR our of REPS, i.e., Ensemble Model Output Statistics (EMOS) • Period of 15 months (verification over the last 3 months) • 10-m wind speed • 550 surface stations over CONUS (in two slides) • Probabilistic prediction attributes: statistical consistency, reliability, sharpness, resolution, spread-error consistency
Delle Monache et al. (MWR 2013) 33 Regional Ensemble Prediction System (REPS)
• Model: GEM 4.2.0 (vertical staggering) • 20 members + 1 control run • 72 hours forecast lead time • Resolution: ~33 km with 28 levels • Initial conditions (i.e., cold start) and 3-hourly boundary condition updates from GEPS (EnKF + multi-physics) • Physics:
o Kain et Fritsch (1993) for deep convection
o Li et Barker (2005) for the radiation
o ISBA scheme (Noilhan et Planton, 1989) for surface • Stochastic Physics: Markov Chains on physical tendencies
34 Ground truth dataset
• 550 hourly METAR Surface Observations • 1 May 2010 – 31 July 2011, for a total of 457 days • 10-m wind speed
35 Probabilistic forecast attributes: Resolution
Consider different classes of forecast events.
If all observed classes corresponds to different forecast classes, then the probabilistic forecast has perfect RESOLUTION.
36 Analysis of Resolution (1)
Relative Operating Characteristics skill score, 10-m wind speed ≥ 5, 10 m s-1
WSPD > 5 m s-1 WSPD > 10 m s-1
0.8 1 AnEn EMOS REPS LR (a) (b)
better 0.7 0.9
0.6
S S
C 0.8 O
R 0.5
0.7 0.4
worse 0.6 0 6 12 18 24 30 36 42 48 0 6 12 18 24 30 36 42 48 Forecast Lead Time (hours) Forecast Lead Time (hours)
37 Probabilistic forecast attributes: Reliability
Example:
① An event (e.g., wind speed > 5 m/s) is predicted to happen with a 30% probability
② We collect the observations that verified every time we made the prediction in 1
③ If the frequency of the event in the observation collected is 30%, then the forecast is perfectly RELIABLE
38 Analysis of reliability & sharpness Reliability and sharpness diagram: 10-m wind speed > 5 m s-1, 9-h fcst 1 1 (a) (b) REPS EMOS 0.8 0.8
0.6 0.6
y
c
n e
u 0.4 0.4
q
e
r
F
t
s 0.2 0.2
a
c
e
r
o F
0 0 ,
y 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
c
n
e
u
q
e
r
F
e 1 1
v i
t (c) (d)
a l
e LR AnEn R
0.8 0.8
d
e
v
r e
s 0.6 0.6
b
O
0.4 0.4
0.2 0.2
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Forecast Probability 39
AnEn in 2/3-D RUC (Ground-truth) (a) Hypothetical atmospheric release in Rapid Update Cycle, ground-truth at t = 0 Washington, D.C., at 9:00 UTC, 1 June 2011
Surface wind vectors (arrows with barbs) 6-h dosage (shading) Hazard contour (red line)
NAM (b) AnEn (c) North America Model Analog Ensemble mean 45-h forecast 45-h forecast
40 Probabilistic forecast attributes: Sharpness
Sharpness refers to the degree of concentration of a forecast PDF’s probability density, and is a property of the forecasts only.
Ideally, we want the forecast system, while mainly reliable, with as many forecasts as possible close to 0% and 100%, corresponding to a perfect deterministic forecast system. However, an improvement in sharpness does not necessarily mean that the forecast system has improved.
0 5 10 15 20 T 0 5 10 15 20 T Sharper Forecast Less Sharp Forecast
41 Analysis of reliability & sharpness Reliability and sharpness diagram: 10-m wind speed > 5 m s-1, 9-h fcst 1 1 (a) (b) REPS EMOS 0.8 0.8
0.6 0.6
y
c
n e
u 0.4 0.4
q
e
r
F
t
s 0.2 0.2
a
c
e
r
o F
0 0 ,
y 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
c
n
e
u
q
e
r
F
e 1 1
v i
t (c) (d)
a l
e LR AnEn R
0.8 0.8
d
e
v
r e
s 0.6 0.6
b
O
0.4 0.4
0.2 0.2
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Forecast Probability 42
AnEn sensitivity Relative Operating Characteristics skill score, 10-m wind speed ≥ 5 m s-1
AnEn with a shorter AnEn built with a coarser training data set (15 9 months) dynamical model (15 33 km)
0.8 better
0.75
S S
C 0.7
O R
0.65 AnEn LR Full Training Short Training worse 0.6 0 6 12 18 24 30 36 42 48 Forecast Lead Time (hours)
43 Analysis of spread-error consistency (2) Binned spread-skill diagram, 10-m wind speed, 42-h fcst
44 Probabilistic forecast attributes: Statistical and spread-error consistency
① The ensemble spread tell us how uncertain a forecast is. Ideally, large spread should be associate with larger uncertainties, low spread should indicate higher accuracy
① If an ensemble is perfect, than the observations are indistinguishable from the ensemble members
45 Spread-skill relathionship
ECMWF EPS
46 Analog predictors weighting strategies: Data locations
Junk et al. (WE 2015) 47 Analog predictors weighting strategies: Impact on deterministic prediction
48 Analog predictors weighting strategies: Impact on probabilistic prediction
49 Results: exampleTime ser ieofs o f time6-hourly wseriesind speed (Lamont,at Lamont, OK OK) Training period is last 365 days, analogs searched for period of observations
Analogs: 25, selected with ws.wd.p, 2 neighbors in trend, 0 neighbor in search window Lamont, OK: simple topography 16 12 8 2001 4 16 12 8 2002 4 16 12 8 2003 4 16 ) 12
1 2004
- 8 s
4
m ( 16 d 12 e 2005
e 8
p 4
s
d
n 16 i 12
W 8 2006 4 16 12 8 2007 4 16 12 8 2008 4 16 12 8 2009 4 Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec 60 m AGL observations 60 m AGL AN-KF 2 Training period: 2009-10-18 to 2010-10-17, downscale period: 2001-01-01 to 2009-10-17 Pearson r = 0.80, Spearman r = 0.89 (across period shown; daily) Vanvyve et al. (RE 2014) 50 PDF comparisons
Hourly wind speed Hourly wind speed Hourly wind speed Hourly wind speed Hourly wind speed Wind speed at Butler: Wind speed at Boulder, CO: Wind speed at Bovina: Wind speed at CapeMay: Histogram Wind speed at Bovina: Histogram Histogram Histogram Histogram
0.20 0.16 Observations Observations Observations Observations Observations MERRA MERRA MERRA MERRA MERRA Analog ensemble 0.12 Analog ensemble 0.12 Analog ensemble 0.12 Analog ensemble Analog ensemble 0.16 0.12
y 0.09
y y
t 0.09
y
i
t t
y
t i
i 0.09
i t
s Outlier
i s
s Outlier Outlier s
0.12 Outlier n
s n
n Outlier n
e th
n e e th th
e th 75 d
75 e th 75
d d
75
d
75 Inside d
Inside Inside y
Inside Mean
y y
Mean Inside t Mean
y
t t Mean i
y inner
t th
i i inner Mean l inner
th t th
i
l l
inner th i
i
l i
i inner th 0.08 50 l
50 i 50 b
i 50 fence b b fence 50 th fence
th b fence th a
b fence th 25 a a 25 th 25
a 25 a
25 b b
b 0.06 0.06
b b
0.06 o
o o r
o Outlier
r r 0.08 Outlier o Outlier r Outlier
r Outlier
P
P P
P P
0.04 0.04 0.03 0.03 0.03
0.00 0.00 0.00 0.00 0.00 0 1 2 3 4 5 6 7 8 9 10 11121314151617181920 2122232425 30 0 1 2 3 4 5 6 7 8 9 10 11121314151617181920 2122232425 30 0 1 2 3 4 5 6 7 8 9 10 11121314151617181920 2122232425 30 0 1 2 3 4 5 6 7 8 9 10 11121314151617181920 2122232425 30 0 1 2 3 4 5 6 7 8 9 10 11121314151617181920 2122232425 30 -1 -1 -1 Wind speed (m s ) Wind speed (m s-1 ) Wind speed (m s-1 ) Wind speed (m s ) Wind speed (m s )
Training period: last 365 days, period downscaled: last 3 entire years, analogs: 25 Training period: last 365 days, period downscaled: last 1 entire year, analogs: 25 Training period: last 365 days, period downscaled: last 1 entire year, analogs: 25 Training period: last 365 days, period downscaled: last 5 entire years, analogs: 25 Training period: last 365 days, period downscaled: last 1 entire year, analogs: 25
Training: 2 Jan 2012 to 31 Dec 2012, period: 2 Jan 2009 to 1 Jan 2012, trend: 4 neighbors, window: 0 neighbor, var: ws.wd Training: 3 Mar 2011 to 1 Mar 2012, period: 3 Mar 2010 to 2 Mar 2011, trend: 4 neighbors, window: 0 neighbor, var: ws.wd Training: 10 Oct 2011 to 8 Oct 2012, period: 10 Oct 2010 to 9 Oct 2011, trend: 4 neighbors, window: 0 neighbor, var: ws.wd Training: 4 Aug 2007 to 2 Aug 2008, period: 5 Aug 2002 to 3 Aug 2007, trend: 4 neighbors, window: 0 neighbor, var: ws.wd Training: 25 Sep 2008 to 24 Sep 2009, period: 26 Sep 2007 to 24 Sep 2008, trend: 4 neighbors, window: 0 neighbor, var: ws.wd
Hourly wind speed Hourly wind speed Hourly wind speed Hourly wind speed Hourly wind speed Wind speed at CedarCreek: Wind speed at CedarCreek: Wind speed at Cochran: Wind speed at Goodnoe Hills, WA: Wind speed at Hatfield: Histogram Histogram Histogram Histogram Histogram
0.16 Observations Observations Observations Observations Observations MERRA MERRA MERRA MERRA 0.16 MERRA Analog ensemble 0.12 Analog ensemble Analog ensemble Analog ensemble 0.12 Analog ensemble 0.100
0.12
0.12
y y
0.09 y
y y
t t
t 0.09
i
i
t t
i i
i 0.075
s s
Outlier s Outlier s
s Outlier Outlier Outlier
n
n
n
n n e e th th
e th e
e th 75 75 th d d 75
75 d 75
d d
Inside Inside Inside Inside Inside
y y
Mean y Mean y
y Mean t
Mean t Mean
t
t t
i i
inner th i inner th
i i l l inner th 0.08
inner th l inner th
l l
i
i
i i
i 50 50 50 50 50 b b fence fence
b fence b
b fence th th th th fence th
a a
25 a 25 a
a 25 0.08 25 25
b b
0.06 b b
b 0.06 o
0.050 o
o
o o r
r Outlier Outlier
r r
r Outlier Outlier Outlier
P
P
P
P P
0.04 0.025 0.03 0.04 0.03
0.000 0.00 0.00 0.00 0.00 0 1 2 3 4 5 6 7 8 9 10 11121314151617181920 2122232425 30 0 1 2 3 4 5 6 7 8 9 10 11121314151617181920 2122232425 30 0 1 2 3 4 5 6 7 8 9 10 11121314151617181920 2122232425 30 0 1 2 3 4 5 6 7 8 9 10 11121314151617181920 2122232425 30 0 1 2 3 4 5 6 7 8 9 10 11121314151617181920 2122232425 30 -1 -1 -1 -1 Wind speed (m s ) Wind speed (m s ) Wind speed (m s-1 ) Wind speed (m s ) Wind speed (m s )
Training period: last 365 days, period downscaled: last 3 entire years, analogs: 25 Training period: last 730 days, period downscaled: last 2 entire years, analogs: 25 Training period: last 365 days, period downscaled: last 3 entire years, analogs: 25 Training period: last 365 days, period downscaled: last 3 entire years, analogs: 25 Training period: last 365 days, period downscaled: last 3 entire years, analogs: 25
Training: 2 Jan 2012 to 31 Dec 2012, period: 2 Jan 2009 to 1 Jan 2012, trend: 4 neighbors, window: 0 neighbor, var: ws.wd Training: 2 Jan 2011 to 31 Dec 2012, period: 2 Jan 2009 to 1 Jan 2011, trend: 4 neighbors, window: 0 neighbor, var: ws.wd Training: 30 Jun 2011 to 28 Jun 2012, period: 30 Jun 2008 to 29 Jun 2011, trend: 4 neighbors, window: 0 neighbor, var: ws.wd Training: 1 Jan 2009 to 31 Dec 2009, period: 2 Jan 2006 to 31 Dec 2008, trend: 4 neighbors, window: 0 neighbor, var: ws.wd Training: 1 Jan 2005 to 31 Dec 2005, period: 2 Jan 2002 to 31 Dec 2004, trend: 4 neighbors, window: 0 neighbor, var: ws.wd
Hourly wind speed Hourly wind speed
Hourly wind speed Hourly wind speed Wind speed at StKillian: Wind speed at Utah: Wind speed at Isabella: Hourly wind speed Wind speed at Megler: Histogram Histogram Histogram Wind speed at Jewell: Histogram Histogram
0.20 Observations Observations Observations Observations MERRA MERRA Observations 0.16 MERRA MERRA 0.12 Analog ensemble Analog ensemble MERRA 0.16 Analog ensemble Analog ensemble Analog ensemble 0.16 0.100
0.12 y
y 0.09
t t i
0.12 i
y y s
s Outlier Outlier
t t
i i
n n
y
t
s s
i e
0.12 Outlier 0.075 Outlier e th th
n n
s 75 75 d
Outlier d
e e th th
n Inside Inside y
75 75 y d d Mean Mean
e th
t t
i Inside 75 Inside i inner inner
d th th
l l
y y
i
Mean Mean i t
t Inside 50 50
i i
y b
inner th Mean inner th b l l fence fence
t th th
i i
50 i inner 50 a th a
l 25 25
i b
b fence th 50 fence th 0.08 b
b 0.06
b a
a 25 fence th 0.08 25
o o
a b
b 25 r
r Outlier Outlier
b
o o
P P r
r 0.08 Outlier 0.050 Outlier o
r Outlier
P P P
0.04 0.04 0.03 0.04 0.025
0.00 0.00 0.00 0.000 0.00 5 7 5 7 0 1 2 3 4 5 6 7 8 9 10 11121314151617181920 2122232425 30 0 1 2 3 4 5 6 7 8 9 10 11121314151617181920 2122232425 30 0 1 2 3 4 5 6 7 8 9 10 11121314151617181920 2122232425 30 0 1 2 3 4 6 8 9 10 11121314151617181920 2122232425 30 0 1 2 3 4 6 8 9 10 11121314151617181920 2122232425 30 -1 -1 -1 -1 Wind speed (m s ) Wind speed (m s-1 ) Wind speed (m s ) Wind speed (m s ) Wind speed (m s )
Training period: last 365 days, period downscaled: last 5 entire years, analogs: 25 Training period: last 365 days, period downscaled: last 1 entire year, analogs: 25 Training period: last 365 days, period downscaled: last 1 entire year, analogs: 25 Training period: last 365 days, period downscaled: last 5 entire years, analogs: 25 Training period: last 365 days, period downscaled: last 1 entire year, analogs51: 25
Training: 20 Dec 2006 to 19 Dec 2007, period: 21 Dec 2001 to 19 Dec 2006, trend: 4 neighbors, window: 0 neighbor, var: ws.wd Training: 1 Jul 2004 to 30 Jun 2005, period: 2 Jul 2003 to 30 Jun 2004, trend: 4 neighbors, window: 0 neighbor, var: ws.wd Training: 3 Nov 2011 to 1 Nov 2012, period: 3 Nov 2010 to 2 Nov 2011, trend: 4 neighbors, window: 0 neighbor, var: ws.wd Training: 2 Jan 2004 to 31 Dec 2004, period: 3 Jan 1999 to 1 Jan 2004, trend: 4 neighbors, window: 0 neighbor, var: ws.wd Training: 6 Jun 2005 to 5 Jun 2006, period: 6 Jun 2004 to 5 Jun 2005, trend: 4 neighbors, window: 0 neighbor, var: ws.wd COSMO 6 km Reanalysis
– Regional reanalysis generated in the framework of the Hans-Ertel- Centre for Weather Research – COSMO model • 6 km horizontal resolution for domain covering whole Europe • 40 vertical levels – Output • 3D atmospheric state - 60 minutes interval • 2D parameters - 15 minutes interval – Production • 16 years finished by end of the year, final 35 year period intended
1979 1998 2002 2007 2013
planned in production available 52
COSMO 6 km Reanalysis
53 Precipitation downscaling over Germany with AnEn Brier Skill Score Against DWD 6-km Reanalysis (AnEn = Rean. BSS = 0, AnEn perfect BSS = 1)
0.1 mm 1.0 mm 5.0 mm
NOTE: Training:2010-2012, Verification: 2007-2009
Lead: Jan Keller (DWD, University of Bonn) 54 AnEn for wind resource assessment in areas with no observations
• Brute force WRF simulations at 17 mast sites for 2002-2012 • Nested 10 km – 2 km grid configuration (one way nesting) • Analog ensemble used to downscale 10 km WRF
Analog ensemble WRF 2 km (deterministic + probabilistic) wind speed
WRF 10 km wind speed
2002 9 years to downscale 2011 1 year where 2 and 2012 10 km WRF overlap
55 U.S. West Coast Site 1: Time series for 2 representative months
December 2011 Hourly wind speed 10km brute force
2km brute force
2km analog ensemble July 2011
Rife et al. (JAMC 2015) 56 U.S. West Coast Site 1: Fit to measurements
Daily wind speed
R² = 0.80 R² = 0.80
Measurements Measurements
2km Analog Ensemble WRF 2km brute force
RMSD [m/s] 1.28 RMSD [m/s] 1.29
BIAS [m/s] 0.23 BIAS [m/s] 0.24
57 Sensitivity tests: Specific year(s) used for training
How well does analog ensemble replicate brute force WRF 1.00
N=17 mast sites
0.00 BIAS [m/s] Hourly wind speed -1.00 1.00 AnEn
downscaled
2 year (2002)
R 0.80
0.60 AnEn
2.00
1.00
[m/s] RMSD 0.00
2
0
Index -2
Oceanic Nino Nino Oceanic 2002 2003 2004 2005 2006 2007 2008 2009 2010 Training year 10 km WRF downscaled to 2 km with AnEn. Training data: 2 km WRF for a single year. 58 Analog ensemble: Wind mapping
R2 = 0.996 Bias = -0.192 m/s CRMSD = 0.20 m/s
Courtesy of Daran Rife (DNV GL) 59 Ozone: AirNow sites distribution
NOTE: used 687 (blue) stations out of 1432 (blue + red) (80% data availability cut-off)
60 Ozone: RMSE
Ranking
1. KF-AN 2. KF 3. AN 4. PERS 5. 7-day 6. RAW
Analog predictors: Ozone and 2-m T, 20 analogs 61 Ozone: Correlation
Ranking
1. KF-AN 2. KF 3. AN 4. PERS 5. 7-day 6. RAW
Analog predictors: Ozone and 2-m T, 20 analogs 62 Analysis of spread-error consistency Binned spread-skill diagram Boulder, CO and Bovina at 50m
Hourly wind speed Hourly wNin =d 2s5p9e4e9d N = 8644 Wind speed at Boulder, CO: Wind speed at Bovina: Binned spread-skill diagram Binned spread-skill diagram
6.0 )
) 2.4
1 1
- -
s s
m m
( (
n n
a a e
e 4.5
m m
e e l
l 1.8
b b
m m
e e
s s
n n
e e
n n
i i
r
r 3.0
o o
r r r
r 1.2
e e
f f
o o
n n
o o
i i
t t
a a
i i
v v
e e
d d
1.5 d
d 0.6
r r
a a
d d
n n
a a
t t
S S Perfect spread-skill relationship Perfect spread-skill relationship 0.0 0.0 0.0 0.8 1.6 2.4 3.2 4.0 4.8 5.6 0.0 0.4 0.8 1.2 1.6 2.0 2.4 2.8 Binned ensemble spread (m s-1 ) Binned ensemble spread (m s-1 )
Training period: last 365 days, period downscaled: last 3 entire years, analogs: 25 Training period: last 365 days, period downscaled: last 1 entire year, analogs: 25
Training: 2 Jan 2012 to 31 Dec 2012, period: 2 Jan 2009 to 1 Jan 2012, trend: 4 neighbors, window: 0 neighbor, var: ws.wd Training: 10 Oct 2011 to 8 Oct 2012, period: 10 Oct 2010 to 9 Oct 2011, trend: 4 neighbors, window: 0 neighbor, var: ws.wd
63 NASA’s MERRA • Introducti on
. NASA Modern-Era Retrospective Analysis for Research and Applications (MERRA)
. Based on NASA’s Global Atmospheric Model
and Data Assimilation System
. 3-D worldwide record of weather from 1979
. 1/2 degrees latitude × 2/3 degrees longitude
. Hourly surface 2D and 6 hourly 3D fields
. Assimilation of all NASA historical satellite data
. Conventional data