PREDICTING SUSPENDED SEDIMENT CONCENTRATION AT NOORDWIJK

1
PREDICTING SUSPENDED SEDIMENT CONCENTRATION
AT NOORDWIJK

• Biswa Bhattacharya
• UNESCO-IHE Institute for Water Education

Delft Cluster II Morphodynamics of the North
Sea Workpackage 4 (A B) CT 05.20 North Sea and
Coast Sponsor Delft Cluster Rijkswaterstaat
2
Data-driven modelling - introduction

• "Data-driven" model is defined as a model
  connecting the systems input and output
  variables with some knowledge about the physical
  process
• example rainfall-runoff model
• The quality and quantity of measured data is
  important
• Important methods artificial neural networks,
  fuzzy logic, regression and model trees, etc.
• DDM is often called input-output models, or
  black box models
• black box - because the model does not
  reproduce how actually the modelled system works,
  but just the input-output relationship

3
Data-driven modelling introduction (2)

• Basic idea of DDM minimize the model error by
  tuning the model parameters
• optimization plays here central role

4
Artificial neural networks - introduction

There are (Ninp1)Nhid (Nhid1)Nout weights to
be identified

• Applications
• Rainfall runoff modelling
• Water level prediction, water systems control
• Sediment transport
• Prediction of harbour sedimentation

ANN demonstration
5
DDM in the context of this workpackage

• Depends upon available data, data analysis and
  literature review
• Stand alone data-driven models is not the
  objective focus is on hybrid modelling
• Three types of data are considered
• wind, waves, river discharge, water level, etc.
• sediment concentration (SPM) data along the Dutch
  coast
• Focus of DDM to model the SPM supply from the
  meteo forcing

SPM Suspended particulate matter
6
Data at Noordwijk

• Collected by CEFAS (Centre for Environment,
  Fisheries and Aquaculture Sciences, UK) and RIKZ
• CEFAS Minipod deployed at three sites, Noordwijk
  2, 5 10
• Hourly SPM and wave data during 2000 2002
• Data was collected using a SmartBuoy with Optical
  Backscatter technology

7
Data filtering
8
Low pass wave vs low pass SPM

• Low pass SPM is highly correlated with low pass
  significant wave height
• Time lag of about 10 hours
• Correlation coefficient is between 0.9 and 1.0

9
High pass wave vs high pass SPM

• High pass SPM is less correlated with high pass
  significant wave height
• Time lag of about 6 hours
• Correlation coefficient is between 0.5 and 0.8

10
Data analysis SPM bed shear stress
11
Storms and calm periods

• SPM do have correlation with bed shear stress
• Relationship is complex, depends upon other
  factors

12
Missing data

• ANN models were built to estimate the missing
  wave height and wave periods
• The accuracy of the models is high
• In the future the possibility of reproducing the
  wave data using data-driven models may be
  explored

13
Data-driven model at Noordwijk

• Inputs
• Low pass bed shear stress (three time steps)
• SW Wind
• 7Day moving average significant wave height
• Output
• Low pass SPM
• Method ANN modelling

14
Result on training data
15
Result on testing data
16
Result on testing data (2)
17
Testing with another testing data at Noordwijk-10
18
Subsequent steps in using the data-driven model

• Using the data-driven model at other locations
  with local hydro-meteo data
• Correct the prediction for cross-shore gradients
• Using the data-driven models prediction at the
  Southern boundary of a local fine-grid model
  (hybrid model)

19
Using the built model at Noordwijk-2
20
Predicted SPM at the south boundary
21
Local model
22
Hybrid modelling

• Use the output (SPM) of data-driven model as the
  south boundary of the numerical model

23
Conclusions (2)

• Data-driven models to predict spm from local
  (point) hydro-meteo conditions perform
  moderately well
• Inclusion of other important factors influencing
  spm may further improve the accuracy
• DDM predicts reasonably well at other locations
  it needs more testing at other locations
• Further actions
• Use remote sensing to test the DDM
• Improve corrections for cross-shore gradient
• Recommendations
• Collect more data. Presently the spatial and
  temoral resolution of SPM measurements are in no
  match with the spatial and temporal resolution of
  meteo data.

24
Thank you

• Acknowledgement
• Qinghua Ye, Li Wang, Sowed Sewagudde, Johan de
  Kok, Thijs van Kessel and Dano Roelvink
• Delft Cluster and Rijkswaterstaat