application of a multi-scheme ensemble prediction system for wind power forecasting in ireland
TRANSCRIPT
Application of a
Multi-Scheme
Ensemble Prediction System
for Wind Power Forecasting
in Ireland
WEPROG ApS, DenmarkWeather and Wind Energy Prognosis
Corinna Möhrlen, [email protected]
Jess Jørgensen, [email protected]
University College Cork, IrelandSustainable Energy Research Group,Department of Civil and Environmental Engineering
Steven Lang, [email protected]
Brian Ó Gallachóir, [email protected]
E. McKeogh, [email protected]
INTRODUCTION & RATIONALE
ENSEMBLE PREDICTION SYSTEMS (EPS)
WIND POWER PREDICTION & UNCERTAINTY
RESULTS & VALIDATION
CONCLUSIONS
INTRODUCTION
Reliable wind power forecasting is vitally important to:
* Enable high wind penetration
* Decrease costs of balancing power
* Maximise CO2 benefit of wind generation
* Ensure power system security and stability, particularlyon weakly interconnected grids
IMPORTANCE OF FORECASTING ON IRISH GRID
* Total installed generation on network is ~ 6300MW
* Maximum demand 4800MW & minimum demand 2000MW
* Installed wind generation was 500MW at end of 2005,and an additional 780MW with connection agreements
* Further 2700MW applications to connect to grid
* Weak interconnection of Republic of Ireland grid with Northern Ireland (NI) grid, and only weak interconnectionof NI with Scotland and the rest of UK.
‘TRADITIONAL’ WIND POWER FORECASTING
* Persistence
* Physical models
* Statistical models
* Hybrid models of the above
Most rely on input weather forecast datafrom national meteorological services…
These deterministic forecasts of wind speed and direction are not usually designed for wind power prediction, and introduce the greatest errors to predicted wind power
ENSEMBLE PREDICTION SYSTEMS (EPS)
A group, or ‘ensemble’, of weather forecasts produced in order to quantify the uncertainty of the forecast.
Different approaches:
* Ensemble Kalman Filter
* Singular vector
* Breeding vector
* Multi-model EPS
* Multi-scheme EPS
MULTI-SCHEMEENSEMBLE PREDICTION SYSTEM (MS-EPS)
* 75-member, limited area EPS
* 75 different Numerical Weather Prediction (NWP)model parameterisations, or ‘schemes’
* Each member’s scheme differs in formulation of fast meteorological processes
* Multi-scheme method reduces ensemble bias and quantifies forecast uncertainty
BACKGROUND TO DEVELOPMENT OF MS-EPS
* Research at UCC since 2000
* Operational system launched by WEPROG at Energinet.dk (then Eltra), 2003
* Testing in research projects, e.g. Honeymoon, 2003-05
* Currently forecasting ~ 20GW wind power
* Operating real-time, world-wide by WEPROG
* Ongoing research and development by UCC and WEPROG
WEATHER PREDICTION WITH MS-EPS
12 hour Forecast 10m wind speed, UK and Ireland, 23/1/06
WIND POWER PREDICTION MODULE
Converts weather forecast to wind power:
1 – Calibration Step
* ‘Training’ of each ensemble member using historicalpower production data
* Direction dependent, time independent power curvesproduced for each ensemble member
2 - Forecast Step
* Predict power using directional power curves
WIND POWER PREDICTIONEnerginet.dk - Operational System since 2003
72 hour Wind Power Forecast for Eltra area, Denmark, 12/1/06
IRISH RESULTS
Validation against data from Golagh wind farm, Co. Donegal, northwest Ireland (complex terrain, high load factor)
Photo courtesy B9 Energy
VALIDATION
Error Descriptors:
* MAE = mean absolute error
* Bias
* Standard deviation and RMSE
All normalised to the installed capacity of the wind farm or the aggregate operational area
Golagh Wind Farm Verification 2/1/05 – 1/5/05
---- Observed power data with 1 hr smoothing
IRISH RESULTS
Golagh observed power data is dominated by large fluctuations with amplitude comparable to the EPS spread
- similar effects have been observed at Horns Rev:
---- Observed power, raw 15 min data Horns Rev output (from Eltra System Plan 2004)
IRISH RESULTS – Daily Forecasts for Golagh
Example 00UTC 48hr forecasts, 2/1/05 – 13/1/05
MS-EPS IS ABLE TO QUANTIFY UNCERTAINTY
---- Observed power data with 1 hr smoothing
QUANTIFICATION OF UNCERTAINTYIS AN IMPORTANT FEATURE OF THE MS-EPS
* Physically realistic uncertainty estimate
* Grid operators have difficulty dealing with forecasting system which uses single, deterministic weather forecasts from national met services as input to forecasting tool – forecasts can be sometimes ‘way out’
* Minimise balancing generation and associated costs
* System security is enhanced with better forecasts and information on uncertainty – assists in operating the system during atypical weather events
IRISH RESULTS
Variation of forecast quality at Golagh Wind Farm
2005 Bias nMAE SD R2 Vavg (m/sec)
Jan -2.3% 13.8% 19.4% 0.76 12.9
Feb -2.6% 10.8% 14.4% 0.85 9.6
Mar -1.2% 9.8% 13.7% 0.87 9.2
Apr 0.6% 11.0% 14.3% 0.81 8.4
AVG -1.4% 11.4% 15.7% 0.84 10.1
•Error statistics generated from 24-48 hr forecasts
•30m agl model wind speed
•Normalised to wind farm capacity of 15MW
IRISH RESULTS
Variation of forecast error with forecast length - Golagh
Normalised mean absolute error out to 48 hour horizon
SOLID __ Statistical best guess
Dashed --- Mean
Dotted … Best member
COMPARISON WITH DANISH & GERMAN RESULTS
* To study any differences between forecasting for single sites and aggregate areas of wind power production
* To investigate the effect of geographical dispersion of turbines on forecasting error
RESULTS – Germany / Denmark / Ireland
Area/Site Germany Denmark West
Golagh
(Ireland)
Horns Rev
(Denmark)
Scaled to Represent:
17GW 2.5GW 15MW 160MW
Average Load Factor
24% 28% 35 – 55% 35 – 55%
nMAE 4.4% 8.2% 12.5% 14.5%
Standard Deviation
6% 12% 18% 21%
DISTRIBUTION OF ERRORS
Frequency distribution of errors for single sitesand Danish and German aggregate areas
0-1 1-2 2-4 4-9 9-16
16-25
25-36
36-49
49-64
64-81
0
5
10
15
20
25
30
35
40
45
50
Germany
error bins [%]
num
ber
of
case
s in
each
bin
0-1 1-2 2-4 4-9 9-16 16-25
25-36
36-49
49-64
64-81
0
5
10
15
20
25
30
35
40
45
50
Denmark
error bins [%]
num
ber o
f cas
esin
eac
h bi
n
0-1 1-2 2-4 4-9 9-16
16-25
25-36
36-49
49-64
64-81
0
5
10
15
20
25
30
35
40
45
50
Golagh
error bins [%]
num
ber
of
case
s in
each
bin
0-1 1-2 2-4 4-9 9-16 16-25
25-36
36-49
49-64
64-81
0
5
10
15
20
25
30
35
40
45
50
HornsRev
error bins [%]
num
ber of
cas
es in
eac
h bi
n
CONCLUSIONS
* Golagh and Horns Rev have significant power output fluctuations and higher forecast errorsthan aggregate wind power production areas
* Forecast errors appear to increase with increasing load factor, due to increasingatypical weather events and the greater number of hours at turbine cut-off
CONCLUSIONS
* Study suggests the prediction error in Ireland will be considerably lower with geographical dispersion of wind farms
* Forecasting for individual farms is more difficult and less accurate than aggregatedwind power forecasts
CONCLUSIONS
* The Multi-Scheme Ensemble Prediction System offers the possibility to estimate the uncertainty of the forecasts
* This provides operators more security when handling wind power and hence enables higher wind penetration
ACKNOWLEDGEMENTS
Sustainable Energy Ireland:Study funds under RE/W/03/006
ESB National Grid:Data provision and support