validation of new model for short-term forecasting of ......dnv gl © 2015 contents background...
TRANSCRIPT
DNV GL © 2015 SAFER, SMARTER, GREENERDNV GL © 20151
Validation of New Model for Short-term Forecasting of Turbine Icing
Beatrice Brailey
EWEA Wind Power Forecasting 2015
DNV GL © 2015
Contents
Background
Methods
Validation data
Results
Value to forecast users
Conclusions
1 October, 2015
2
DNV GL © 2015
Background – Icing losses
Icing losses for wind farms at high latitudes are variable and can be highly
significant
– Annual energy production losses from ~0% up to >10%
– Monthly energy production losses from 1% up to >50%
(Staffan Lindahl: Quantification of energy losses cause by blade icing using SCADA data, Winterwind 2014)
Individual icing events can lead to full loss of power
1 October, 2015
3
-3
0
3
6
9
12
15
-0.2
0
0.2
0.4
0.6
0.8
1
11-Dec 13-Dec 15-Dec 17-Dec 19-Dec 21-Dec 23-Dec 25-Dec 27-Dec 29-Dec 31-Dec
Win
d S
peed
(m
/s)
Po
wer (
% c
ap
acit
y)
Actual Power
Expected Power
Wind Speed
DNV GL © 2015
0%
20%
40%
60%
80%
100%
12-Mar 15-Mar 18-Mar 21-Mar 24-Mar 27-Mar
Po
we
r (%
Cap
acit
y)
ForecastActual
Background – Short-term forecasting
The value of short-term forecasting is well understood
State of the art forecasts are typically high accuracy
Beneficial to model blade icing when forecasting for wind farms in cold climates
Icing prediction is woefully unvalidated
– Reliable observation data is scarce
1 October, 2015
4
DNV GL © 2015
Methods – base forecast
1 October, 2015
5
Wind SpeedWind DirectionTemperaturePressureRelative Humidity
Ice accretion
DNV GL © 2015
Methods – icing model
1 October, 2015
6
Wind SpeedWind DirectionTemperaturePressureRelative Humidity
EnsemblePredictions
Adapted from Frohboese and Anders (2007)
Ensemble NWP predictions
Predicted freezing/thawing time
Ice accretion parameterization
Power adjustment
Z
Base
Icing
DNV GL © 2015
Methods – icing model power conversion
1 October, 2015
7
Borrowed from http://www.tuuliatlas.fi/icingatlas/icingatlas_6.html
Finnish Wind Energy Atlas
Ljungberg & Niemelä Model
Converts blade ice, wind speed to power
degradation
DNV GL © 2015
Validation Data
3 wind farms, ~40 wind turbines
Projects in Region 2 and Region 3, where there is
sufficient icing to test model
For each site:
– ~1 year of data for model training
– ~1 year of data for validation
For all projects the turbines remain operational during
blade icing periods
1 October 2015
8
Region 3
Region 1
Region 2
DNV GL © 2015
Results
1 October 2015
9
DNV GL © 2015
Results
1 October, 2015
10
0%
2%
4%
6%
8%
10%
12%
14%
16%
18%
20%
0 6 12 18 24
Mean
Ab
so
lute
Erro
r
(%
cap
acit
y)
Forecast Horizon (hours)
Wind Farm 1 – Avg. 13% icing loss over 4 years
Base Forecast Power
Icing Forecast Power
Day ahead accuracy improvement = 0.95%
DNV GL © 2015
Results
1 October, 2015
11
0%
2%
4%
6%
8%
10%
12%
14%
16%
18%
20%
0 6 12 18 24
Mean
Ab
so
lute
Erro
r
(%
cap
acit
y)
Forecast Horizon (hours)
Wind Farm 2 - Avg. 6% icing loss over 3.5 years
Base Forecast Power
Icing Forecast Power
Day ahead accuracy improvement = 0.65%
DNV GL © 2015
Results
1 October, 2015
12
0%
2%
4%
6%
8%
10%
12%
14%
16%
18%
20%
0 6 12 18 24
Mean
Ab
so
lute
Erro
r
(%
cap
acit
y)
Forecast Horizon (hours)
Wind Farm 3 - Avg. 4% icing loss over 3.5 years
Base Forecast Power
Icing Forecast Power
Day ahead accuracy improvement = 0.21%
DNV GL © 2015
Results
1 October, 2015
13
0%
2%
4%
6%
8%
10%
12%
14%
16%
18%
20%
0 6 12 18 24
Mean
Ab
so
lute
Erro
r
(%
cap
acit
y)
Forecast Horizon (hours)
Overall
Base Forecast Power
Icing Forecast Power
Day ahead accuracy improvement = 0.6%
DNV GL © 2015
Value to forecast users
Icing modelling improved forecast accuracy
Increase energy revenue (based on day ahead trading in liberalized market)
Operational planning
Grid management
1 October, 2015
14
Forecast scenario Average tradingrevenue (€/MWh)
MAE
No forecast 56.3 -
Basic forecast 57.7 22%
State of the art 61.6 12%
Perfect 64.6 0%
Based on day ahead energy trading in the UK
Parkes et al. Wind Energy Trading Benefits Through Short Term Forecasting, EWEC 2006
DNV GL © 2015
Value to forecast users – an advanced warning system
1 October, 2015
15
0%
20%
40%
60%
80%
100%
16-Oct 18-Oct 20-Oct 22-Oct 24-Oct 26-Oct 28-Oct
Po
wer (
% c
ap
acit
y)
Icing Forecast Power Icing Probability Actual Power
DNV GL © 2015
Conclusions
Validation shows icing model adds value to forecasts
– Successful in varying levels of icing
– Reduces MAE by up to 0.95% capacity (average improvement = 0.6%)
Ice accretion ice load power is well modelled
Scope for model improvement
– Meteorological conditions icing freezing time
– Thawing/ice throw
– Upper limit for ice load
Forecast accuracy improvement = increased revenue, informed operations,
improved grid management
1 October, 2015
16
DNV GL © 2015
SAFER, SMARTER, GREENER
www.dnvgl.com
Questions?
1 October 2015
17
Beatrice Brailey
+44 (0) 117 972 9900