Diese Präsentation wurde erfolgreich gemeldet.
Die SlideShare-Präsentation wird heruntergeladen. ×

DSD-INT 2022 Prediction of Wind-Waves Using Long-Short Term Memory (LSTM) Models - Choo

Anzeige
Anzeige
Anzeige
Anzeige
Anzeige
Anzeige
Anzeige
Anzeige
Anzeige
Anzeige
Anzeige
Anzeige
Nächste SlideShare
HollingsPresentation
HollingsPresentation
Wird geladen in …3
×

Hier ansehen

1 von 25 Anzeige

DSD-INT 2022 Prediction of Wind-Waves Using Long-Short Term Memory (LSTM) Models - Choo

Presentation by Jian Feng Choo (Technology Centre for Offshore and Marine, Singapore (TCOMS), Singapore), at the Delft3D User Days, during Delft Software Days - Edition 2022. Monday, 14 November 2022.

Presentation by Jian Feng Choo (Technology Centre for Offshore and Marine, Singapore (TCOMS), Singapore), at the Delft3D User Days, during Delft Software Days - Edition 2022. Monday, 14 November 2022.

Anzeige
Anzeige

Weitere Verwandte Inhalte

Ähnlich wie DSD-INT 2022 Prediction of Wind-Waves Using Long-Short Term Memory (LSTM) Models - Choo (20)

Weitere von Deltares (20)

Anzeige

Aktuellste (20)

DSD-INT 2022 Prediction of Wind-Waves Using Long-Short Term Memory (LSTM) Models - Choo

  1. 1. Prediction of Wind-Waves Using Long-Short Term Memory (LSTM) Models Authors: Jian Feng CHOO (Presenter) Jeng Hei CHOW Pavel TKALICH Technology Centre for Offshore and Marine, Singapore (TCOMS)
  2. 2. 2 Contents ➢Introduction ➢Methodology ➢Results ➢Conclusion / Future Works
  3. 3. 3 Introduction ➢Introduction • Objective • Singapore’s Location • Singapore’s Monsoon Season • Long-Short Term Memory (LSTM)
  4. 4. 4 Objective To develop a machine learning model that provides cheap, fast and accurate prediction of waves for self- driving vessels. Singapore is the second largest port in the world, and a large portion of its revenue is from import and export. These accidents are usually caused by human error or unpredictable/constant changing weather.
  5. 5. 5 Pacific Ocean Where is Singapore? Indian Ocean South China Sea
  6. 6. 6 Singapore’s Monsoon Season Period: November to February The winter air from China blows across east Asia towards the tropical ocean and Australia (Summer). Period: June to August Cold air from Australia blows towards the warmer tropical ocean and the Asia continent.
  7. 7. 7 Long Short Term Memory (LSTM) Neural Network What is LSTM Neural Network? • A deep neural network that is capable of learning long term sequential data. • LSTM are widely used in sentiment analysis, language modelling, speech recognition and time series prediction. • Examples of applications: o Netflix recommendations algorithms o Financial market predictions (Stocks, Bonds, Housing Pricings, etc) o Chatbots
  8. 8. 8 Methodology ➢Methodology • Source of Datasets • Pre-processing • Models schematics • Model training parameters
  9. 9. 9 Source of Datasets Parameters Significant Height of Wind Wave Wind 10m Above Sea Level Source ECMWF ECMWF Type ERA5 Reanalysis ERA5 Reanalysis Period 2010 - 2021 2010 - 2021 Temporal Resolution Hourly Hourly Spatial Resolution 0.5 Deg x 0.5 Deg 0.25 Deg x 0.25 Deg
  10. 10. 10 Data Pre-processing Find strongest correlation between two points and its respective lag. Time Lag Time Lag Cross-Correlation
  11. 11. 11 Data Pre-processing U-velocity Wind V-velocity Wind Significant Height of Wind Wave Wind Speed Cross Correlation between Wind Speed and Significant Height of Wind Wave Cross Correlation between Wind Direction and Significant Height of Wind Wave Wind Direction Split data time series into respective monsoon seasons
  12. 12. 12 Northeast Monsoon (Nov – Mar)
  13. 13. 13 Inter Monsoon 1 (Apr – May)
  14. 14. 14 Southwest Monsoon (Jun – Aug)
  15. 15. 15 Inter Monsoon 2 (Sept – Oct)
  16. 16. 16 Input (Highest Correlation Point) / Output Wind Wave (Output) UV Wind (Input)
  17. 17. 17 LSTM Model Schematics LSTM t + hours prediction ahead U wind (Input) V wind (Input) Wind Wave (Output) t - observations t t - observations t
  18. 18. 18 LSTM Model Training Parameters Training Test 10 Years 1 Year 2010 2021 2020
  19. 19. 19 Results ➢Results • 10 Observation to Predict 1 Hour Ahead • Multiple Observation, Multiple Prediction Ahead • N – Highest Correlated Points
  20. 20. 20 10 Observation to Predict 1 Hour Ahead (Time series) Predicted VS Actual Significant Height Wind Wave (m) Time
  21. 21. 21 10 Observation to Predict 1 Hour Ahead (Scatter Plot) Predicted Actual Actual vs Predicted Values R^2 = 0.958 RMSE = 0.0732
  22. 22. 22 Multiple Observation, Multiple Prediction Ahead Observe (Hours) Prediction ahead (Hours) RMSE R^2 10 1 0.0732 0.958 10 2 0.0987 0.9235 10 5 0.1491 0.8254 20 1 0.0733 0.9581 20 2 0.1064 0.9111 20 5 0.1437 0.8147 30 1 0.0734 0.9578 30 2 0.1191 0.9007 30 5 0.1472 0.8299 • Increasing observations doesn’t increase accuracy. • Increasing prediction ahead decreases accuracy.
  23. 23. 23 N – Highest Correlated Points Wind Wave (Output) UV Wind (Input)
  24. 24. 24 N – Highest Correlated Points Number of Highest Correlated Points (Input) RMSE R^2 1 0.07585 0.9548 3 0.07067 0.9613 5 0.06939 0.9618 • Increasing the number of input of highly correlated points increases the accuracy, but not significant.
  25. 25. 25 Conclusion and Future Works Conclusion 1. LSTM shows promising results to predict wind waves from wind. 2. Increase in observations have no effect in prediction accuracy. 3. Accuracy decreases for increasing ahead prediction. 4. Increasing the number input points with “n” highest correlation points increases the accuracy, however the computational speed to train the model increases. Future Works 1. Embed physics about wind wave interaction into LSTM model. 2. Using ConvLSTM model for 2D time series prediction.

×