Page 1
Wind power prediction risk indices based on numerical weather prediction ensembles
Erik Holmgren, Nils Siebert, George [email protected] , [email protected]
Renewable Energies TeamMINES ParisTech / ARMINES
European Wind Energy Conference 2010, Warsaw, Poland
Page 2
Outline
• Introduction and objectives
• Definition and evaluation of risk indices
• Risk indices in decision making
• Conclusions
2
Page 3
Introduction: ensemble forecasting
• Typical Wind Power Forecasting (WPF) modelling scheme• NWP can also be provided as meteorological ensembles
– Alternative forecasts representing different scenarios– Computed by perturbing initial conditions– Control forecast + ensemble forecast members
3
NWP ensemblesMeasurements ModelMeasurements ModelModel
+1h +6h +12h +18h +24h +30h +36h … +48ht0
Present Future
0
0.5
1
1.5
2
2.5
3
3.5
0
0.5
1
1.5
2
2.5
3
3.5
+1h +6h +12h +18h +24h +30h +36h … +48ht0 +1h +6h +12h +18h +24h +30h +36h … +48ht0
Present Future
Production [MW]
ForecastMeasured production
Numerical Weather Predictions (NWP)
Control forecastEnsemble forecastsMeasured production
Page 4
Objectives
• Investigate relationship between ensemble spread and forecast error
• Quantify the ensemble spread through risk indices
4
ModelModel
+1h +6h +12h +18h +24h +30h +36h … +48ht0
Present Future
0
0.5
1
1.5
2
2.5
3
3.5
0
0.5
1
1.5
2
2.5
3
3.5
+1h +6h +12h +18h +24h +30h +36h … +48ht0 +1h +6h +12h +18h +24h +30h +36h … +48ht0
Present Future
Production [MW]
Control forecastEnsemble forecastsMeasured production
Period with smaller ensemble spread
Period with larger ensemble spread
Page 5
Outline
• Introduction and objectives
• Definition and evaluation of risk indices
• Risk indices in decision making
• Conclusions
5
Page 6
Notation: for
Production [MW]
Present Future
0
0.5
1
1.5
2
2.5
3
3.5
0
0.5
1
1.5
2
2.5
3
3.5
+1h +6h +12h +18h +24h +30h +36h … +48ht0 +1h +6h +12h +18h +24h +30h +36h … +48ht0
Present Future
+1h +6h +12h +18h +24h +30h +36h … +48ht0
Quantify ensemble spread through risk indices
• Definition: the Normalized Prediction Risk Index (NPRI)– Weighted standard deviation
of ensemble members
– Average value over a
look-ahead time window
6
ModelModel
Control forecast
Ensemble forecasts
Ensemble mean
Measured production
Page 7
Relationship between risk indices and forecast errors
• Definition: Energy Imbalance for a look-ahead time window– Sum of absolute errors
• Compute relative imbalances by dividing by long-term average :
7
ModelModel
+1h +6h +12h +18h +24h +30h +36h … +48ht0
Present Future
0
0.5
1
1.5
2
2.5
3
3.5
0
0.5
1
1.5
2
2.5
3
3.5
+1h +6h +12h +18h +24h +30h +36h … +48ht0 +1h +6h +12h +18h +24h +30h +36h … +48ht0
Present Future
Control forecast
Measured production
Production [MW]
Page 8
Evaluation of relationship risk index - forecast error
• Validation of approach on more test cases– Three French onshore wind farms with different characteristics
• Different prediction models
• Temporal scales– Size and position of look-ahead time window
• Spatial scales– Aggregate of wind farms
• Alternative definitions of risk indices– e.g. Range of ensemble members
8
Page 9
Evaluation
• NWPs from ECMWF*– 51 members
• 1 control member• 50 perturbed members
– Temporal resolution: 6 hours (0 - 240 hours)– Spatial resolution: 1o longitude, 1o latitude
– 2 prediction runs per day (midnight & noon)– 1.5 year of data
• WPF model
* European Centre for Medium-range Weather Forecasts
• Wind Farm– French onshore farm
• Flat terrain• Installed capacity: 10.1 MW
Model Type Explanatory variables Procedure Accuracy
Random Forest (RF)
Statistical • Predictions of wind speed• Predictions of wind direction• Last available power measurement
• Model learning on control member• Model testing on all ensemble members
• NMAE = 7 – 9 % for 6 – 48 hours ahead
9
Page 10
NPRI
rela
tive e
nerg
y im
bala
nce
[%
]
NPRI [%]
rela
tive e
nerg
y im
bala
nce
[%
]
1 2 3 4 5
NPRI Class
1 2 3 4 5
Evaluation: NPRI results
Steps1: Estimate Normalized Prediction Risk Indices (NPRI)
2: Compute energy imbalances
3: Derive NPRI – energy imbalance pairs
4: Group into classes based on NPRI
5: Calculate imbalance distributions
Results for day 2 ahead
10
Page 11
Outline
• Introduction and objectives
• Definition and evaluation of risk indices
• Risk indices in decision making
• Conclusions
11
Page 12
• Operational context: – Give alerts when high risk for large energy imbalance
• Alert function:
• Parameters:x = 1.5
y = 0.2
Exploration: risk indices for decision making - proposals
12
Energy imbalance
Mean energy imbalance
“Make an alert when the probability of an energy imbalance larger than x times the average is greater than y”
Page 13
Exploration: risk indices for decision making - results
• Evaluation: – Confusion matrix
• Results – Day 2 ahead
– Interesting approach• Alert function parameters x and y can be tuned depending on risk tolerance• More investigation needed to validate results
13
Alert needed Alert not needed Total
Alert made 4 1 5
Alert not made 15 70 85
Total 19 71 90
Alert needed Alert not needed
Alert made True alert False alert
Alert not made False non alert True non alert
“Make an alert when the probability of an energy imbalance larger than 1.5 times the average is greater than 0.2”
Page 14
Conclusions
• Evaluation• Larger ensemble spread indicates
– Larger average forecast error– Larger uncertainty in the predictions – Higher risk of large forecast error
• Risk indices useful to extract and display this information• Better understanding and extended validation of the concept of risk indices
• Value of risk indices for decision making• Useful in giving alerts for large energy imbalances
• Future work• More focus on the use of risk indices in operational contexts
14
Page 15
Thank you for your attention!
Acknowledgement: European R&D project SafeWind (FP7)