# Sloboda Prague

Published on September 25, 2007

Author: Dabby

Source: authorstream.com

Using Bayesian Belief Network to Calibrate Environmental Process Models:  Using Bayesian Belief Network to Calibrate Environmental Process Models Markiyan Sloboda, David Swayne, Vimal Sharma Prague, ISESS 2007 Presentation Structure:  Presentation Structure Introduction Environmental Models and Bayesian Networks Calibration Using the Bayesian Network A Simple Example Results and Analysis Conclusions and Future Work Introduction:  Introduction Environmental models are important tools for environmental assessment and management Almost all environmental models require computationally expensive calibration Manual and auto calibration methodology Environmental Models:  Environmental Models Non-point source (NPS) models overview Classical calibration of environmental models Analysis of the calibration The Bayesian Network:  The Bayesian Network A structure called a Bayesian network is used to represent dependence between variables and to give a concise specification of the joint probability distribution The Bayesian network captures believed relations between set of variables, which are relevant to some problem. Monte Carlo simulations Store and Forward Approach in the NPS Modelling:  Store and Forward Approach in the NPS Modelling Calibration Using the Bayesian Network:  Calibration Using the Bayesian Network Building a 'reservoir' of search attempts Populate the Bayesian Network running many simulations Environmental models have parameters, that have a certain probability distribution Calibration Using the Bayesian Network:  Calibration Using the Bayesian Network Search within the Bayesian network using targeting search is much more efficient than a regular search used to calibrate a model Parallelization and high-performance computing is used to construct the Bayesian Network Calibration Using the Bayesian Network:  Calibration Using the Bayesian Network The network will predict the value of the calibration parameter by calculating their expected value from the belief propagation For multi-parameter calculation, the correct values of the calibration parameters is a system of K equations with K unknowns (possibly non-linear) from which the 'correct' c parameters must be derived The GAMES Model – a Simple Example:  The GAMES Model – a Simple Example The Guelph model for evaluating the effects of Agricultural Management systems on Erosion and Sedimentation (GAMES) Watershed is discretized into field sized elements Can be used for annual or seasonal assessments The GAMES Model – a Simple Example:  The GAMES Model – a Simple Example Erosion Component is Based on the Universal Soil Loss Equation (USLE) Sediment yield delivered from a field The GAMES Model – a Simple Example:  The GAMES Model – a Simple Example Delivery ratio is the percentage of soil loss delivered downstream Seasonal version of the USLE uses seasonal values for the factors Stratford Avon Watershed:  Stratford Avon Watershed Stratford Avon Watershed Drainage Network DistributedFramework:  Distributed Framework Testing Example Using the SHARCNET:  Testing Example Using the SHARCNET The SHARCNET stands for the Shared Hierarchical Academic Computing Network. Established in 2000 The SHARCNET network can be physically described as a 10G WAN between 8 sites Parallel algorithm for populating the Bayesian network was compared to the earlier proposed algorithm Testing Example Using the SHARCNET:  Testing Example Using the SHARCNET Time necessary to populate the Bayesian network Testing Example Using the SHARCNET:  Testing Example Using the SHARCNET Results and Analysis:  Results and Analysis Calibration of the GAMES model is a minimization of the difference between the predicted total sediment load and the observed load of the watershed constructed Instead of estimating each time the α parameter, the α distribution generated is considered, in this case a logarithmic distribution Results and Analysis:  Results and Analysis The populated network, by the definition will be able to generate the calibrated value for the parameter, by just simply specifying the predicted total sediment load The advantage of this approach is that once the Bayesian network was created it will not be reconstructed to get the calibrated value of α for a specific scenario, which is used over the watershed, unlike, in case of standard calibration procedure Conclusion:  Conclusion In this study an alternative approach was proposed to model calibration Results indicate that use of the Bayesian network for calibration is possible and is reasonable for the cases when the data is changing, whenever the model parameters have to be calibrated The parallel approach to populate the Bayesian network was shown using the SHARCNET Future Work :  Future Work Construct and populate the Bayesian network using more complicated hydrological models, such as SWAT Try calibrating parameters that are dependent on each other to better understand their dependence and behaivour Thank you!:  Thank you!

16. 11. 2007
0 views

15. 06. 2007
0 views

09. 10. 2007
0 views

25. 09. 2007
0 views

25. 09. 2007
0 views

11. 10. 2007
0 views

16. 10. 2007
0 views

17. 10. 2007
0 views

19. 10. 2007
0 views

22. 10. 2007
0 views

22. 10. 2007
0 views

25. 09. 2007
0 views

25. 09. 2007
0 views

05. 10. 2007
0 views

23. 10. 2007
0 views

26. 11. 2007
0 views

28. 11. 2007
0 views

25. 09. 2007
0 views

16. 10. 2007
0 views

26. 10. 2007
0 views

01. 11. 2007
0 views

02. 11. 2007
0 views

29. 10. 2007
0 views

21. 11. 2007
0 views

28. 12. 2007
0 views

02. 01. 2008
0 views

03. 01. 2008
0 views

02. 08. 2007
0 views

02. 08. 2007
0 views

25. 09. 2007
0 views

25. 09. 2007
0 views

04. 10. 2007
0 views

26. 11. 2007
0 views

02. 08. 2007
0 views

22. 10. 2007
0 views

04. 03. 2008
0 views

29. 09. 2007
0 views

12. 10. 2007
0 views

07. 11. 2007
0 views

03. 01. 2008
0 views

12. 03. 2008
0 views

25. 09. 2007
0 views

25. 03. 2008
0 views

26. 03. 2008
0 views

25. 09. 2007
0 views

07. 04. 2008
0 views

28. 03. 2008
0 views

30. 03. 2008
0 views

31. 12. 2007
0 views

09. 04. 2008
0 views

10. 04. 2008
0 views

13. 04. 2008
0 views

14. 04. 2008
0 views

17. 04. 2008
0 views

17. 10. 2007
0 views

22. 04. 2008
0 views

25. 09. 2007
0 views

08. 10. 2007
0 views

02. 11. 2007
0 views

15. 10. 2007
0 views

05. 01. 2008
0 views

18. 06. 2007
0 views

18. 06. 2007
0 views

18. 06. 2007
0 views

18. 06. 2007
0 views

07. 01. 2008
0 views

24. 11. 2007
0 views

01. 01. 2008
0 views

12. 10. 2007
0 views

19. 11. 2007
0 views

16. 10. 2007
0 views

25. 09. 2007
0 views

26. 10. 2007
0 views

23. 10. 2007
0 views

15. 06. 2007
0 views

15. 06. 2007
0 views

15. 06. 2007
0 views

15. 06. 2007
0 views

15. 06. 2007
0 views

15. 06. 2007
0 views

15. 06. 2007
0 views

15. 06. 2007
0 views

09. 10. 2007
0 views

15. 06. 2007
0 views

15. 06. 2007
0 views

25. 09. 2007
0 views

28. 12. 2007
0 views

07. 01. 2008
0 views

23. 11. 2007
0 views

15. 06. 2007
0 views

18. 06. 2007
0 views

18. 06. 2007
0 views

25. 09. 2007
0 views

07. 11. 2007
0 views

18. 06. 2007
0 views

29. 11. 2007
0 views

25. 09. 2007
0 views

07. 10. 2007
0 views