Download as pdf or txt
Download as pdf or txt
You are on page 1of 36

WindTwin

Brunel Innovation Centre Contribution


to WindTwin:

Predictive regression algorithms for


modelling health monitoring sensors and
predicting machinery health deterioration
WindTwin
Contents

1. Brunel Innovation Centre Introduction


2. Brunel Innovation Centre in WindTwin
3. Parameter selection
4. Modelling
5. Performance comparison
6. Transfer learning
7. Conclusion
WindTwin
Mechanical Test Rig Design | Electronic Systems

High Power Sonication


Power Electronics
WindTwin
NDT | Condition and Structural Monitoring | High
Power Sonication
WindTwin
Machine Learning for Diagnosis/Automated Defect
Recognition/Condition Monitoring
WindTwin
Machine Learning for Diagnosis/Automated Defect
Recognition/Condition Monitoring
WindTwin

Renewable Energy Research


Hybrid Anti-icing Fouling prevention for the
technology ladder and the transition
piece using high power
sonication

Digital Twin of Wind Automated Robotic


Turbine Machinery Inspection of Welds using
Eddy Current/TOFD/UT
WindTwin
WindTwin
Contents

1. Brunel Innovation Centre Introduction


2. Brunel Innovation Centre in WindTwin
3. Parameter selection
4. Modelling
5. Performance comparison
6. Transfer learning
7. Conclusion
WindTwin

BIC in WindTwin
WindTwin

Results presented on Westmill WT9


WindTwin
Contents

1. Brunel Innovation Centre Introduction


2. Brunel Innovation Centre in WindTwin
3. Parameter selection
4. Modelling
5. Performance comparison
6. Transfer learning
7. Conclusion
WindTwin
Model
WindTwin

WindTwin Deep Learning Model Prediction


5 ML models with 5 different combination of inputs

wtc_AmbieTmp wtc_A1ExtTmp_ wtc_A4IntTmp_ wtc_AmbieTmp_ wtc_AcWindSp wtc_AcWindSp_ wtc_ActPow wtc_ActPow


Model
_Mean mean mean stddev _mean stddev er_mean er_stddev

1 ✓ ✓ ✓ ✓ ✓ ✓
2 ✓ ✓ ✓ ✓ ✓
3 ✓ ✓ ✓
4 ✓ ✓ ✓
5 ✓ ✓ ✓
WindTwin

Model assessment

Minimum error Maximum error Standard deviation


Model offset value offset value error offset value

1 0.0 73.51 4.10

2 0.0 55.19 3.80

3 0.0 66.67 4.32

4 0.0 67.18 3.99

5 0.0 73.89 4.40


WindTwin
WindTwin training models predictions’ error histogram
Predicted Temperature Error Destribution for 200K testing points
30000
wtc_Am wtc_A1E wtc_A4I wtc_Am wtc_Ac wtc_Ac wtc_Act wtc_Act
Model bieTmp_ xtTmp_ ntTmp_ bieTmp_ WindSp_ WindSp_ Power_ Power_s
25000 Mean mean mean stddev mean stddev mean tddev

1 ✓ ✓ ✓ ✓ ✓ ✓
20000 2 ✓ ✓ ✓ ✓ ✓

3 ✓ ✓ ✓
15000 4 ✓ ✓ ✓

5 ✓ ✓ ✓
10000

5000

0
0 5 10 15 20 25 30
Model1 Model2 Model3 Model4 Model5
WindTwin
Contents

1. Brunel Innovation Centre Introduction


2. Brunel Innovation Centre in WindTwin
3. Parameter selection
4. Modelling
5. Performance comparison
6. Transfer learning
7. Conclusion
WindTwin
Updated Model

995<Shaft speed<1005
High speed
Active power model
Wind speed
Residual
Ambient temperature
temperature
Nacelle temperature
Current Low speed
1495<Shaft speed<1505 model
WindTwin
Selected ML architecture
conv layer 1

conv layer 0

conv layer 2

Input (n) n*2


Output
(n*2) +2 1
WindTwin
Inputs
Mean value and standard deviation

Model 1 Model 2 Model 3

Ambient temperature
Ambient temperature Ambient temperature
External temperature
External temperature External temperature
Internal temperature
Internal temperature Internal temperature
Wind speed
Input Wind speed Wind speed
Active power
Active power Active power
Current
Historical data of Generator
Current
temperature

Output Generator temperature Generator temperature Generator temperature


WindTwin
Contents

1. Brunel Innovation Centre Introduction


2. Brunel Innovation Centre in WindTwin
3. Parameter selection
4. Modelling
5. Performance comparison
6. Transfer learning
7. Conclusion
WindTwin
Performance comparison

High Low High Low


Mean Mean Std Std
Model 1 2.31 2.33 1.85 1.85

Model 2 2.29 2.14 2.34 1.79

Model 3 0.26 0.25 0.33 0.26


WindTwin
Performance comparison

High Low High Low


Mean Mean Std Std
Model 1 2.31 2.33 1.85 1.85

Model 2 2.29 2.14 2.34 1.79

Model 3 0.26 0.25 0.33 0.26


WindTwin
Performance comparison using training data
WindTwin
Model 2 Performance using training data
WindTwin
Model 2 testing data

Real vs predicted

Single point

6 hour sliding window (36


points averaged)
WindTwin
Model 2 testing data

Real vs predicted

Single point

6 hour sliding window (36


points averaged)
WindTwin
Model 2 testing data

Real vs predicted

Single point

6 hour sliding window (36


points averaged)
WindTwin
Model 2 Confusion Matrix
normal 6hour/36points sliding window
WindTwin
Contents

1. Brunel Innovation Centre Introduction


2. Brunel Innovation Centre in WindTwin
3. Parameter selection
4. Modelling
5. Performance comparison
6. Transfer learning
7. Conclusion
WindTwin
Transfer learning
WindTwin
Transfer learning
Traditional learning Transfer learning
WindTwin
Transfer learning: High Speed Gen

Single point

6 hour sliding window (36


points averaged)
WindTwin
Transfer learning: Low Speed Gen

Single point

6 hour sliding window (36


points averaged)
WindTwin
Contents

1. Brunel Innovation Centre Introduction


2. Brunel Innovation Centre in WindTwin
3. Parameter selection
4. Modelling
5. Performance comparison
6. Transfer learning
7. Conclusion
WindTwin
Conclusions

• The regression model allows to predict the monitoring sensor output based on
operational parameters of the wind turbine without knowing anything about the
specificities of the monitored asset. BUL has successfully developed virtual
monitoring sensors with an excellent 99.8% accuracy.
BUL has basically created a digital twin of a generator monitoring sensor and use the
deviation of the real sensor to infer the degradation of the asset

• The regression models developed based on temperature of the generator allow to


detect a defect developing 2 months before the wind turbine shutdown

• The transfer learning model needs further development to allow quick training of
new assets

You might also like