際際滷

際際滷Share a Scribd company logo

"Streamflow prediction in River Rhine: Exploring combinations of bias-correcting forcing and bias-correcting flow

Download as PPT, PDF
Jan Verkade
Jan Verkade

This document explores bias correction of temperature, precipitation, and streamflow predictions for the Rhine River. It finds that bias correcting temperature and precipitation predictions improves their accuracy by 20-60% and 20-30% respectively, but does not consistently translate to improved streamflow prediction accuracy. Preserving the spatial and cross-variable dependencies between predictions is important for streamflow skill. Future work will focus on determining whether the benefits of bias correcting forcing inputs are maintained after also bias correcting streamflow predictions.

1 of 16
Downloaded 10 times
Streamflow prediction in River Rhine: Exploring combinations of bias-correcting forcing and bias-correcting flow Jan Verkade (Deltares and Delft University of Technology) James Brown (NOAA-NWS-OHD and UCAR)
Motivation and research questions Biases/uncertainty in predicted forcing used for streamflow prediction: NWP models are skillful, but biased (mean, spread,..) This bias/uncertainty propagates from forcing to flow Bias-correction of precipitation is complex Ultimately, flow bias-correction is always needed Key research questions: What is the signal from bias-correction of forcing in streamflow? Is this signal maintained after bias-correction of flow , i.e. is forcing correction needed?
Research design Raw forcing (T,P) Hydrologic model Raw flow Ensemble verification B-C forcing (T,P) Hydrologic model Raw forcing (T,P) Hydrologic model Scenario 1 Scenario 2 Baseline B-C streamflow B-C streamflow
油
Data kindly provided by Florian Pappenberger @ ECMWF
Observed forcing data: E-OBS dataset Downloadable from KNMI @ http://eca.knmi.nl/
Bias-correction of temperature, precipitation and flow The random variables (one time/location): Predictand Y = observed temp/precip/flow. Assumed unbiased! Potential predictors X = {X1,,X5,, Xm}; biased. The bias-corrected forecast: How to parameterize for T and P? Parsimonious model (subject to skill!) Model the statistical dependence (traces)
Bias-correction of temperature, precipitation and flow (2/2) Temperature normal regression: linear regression in normal space Precipitation logistic regression: linear regression in logistic space Streamflow: Krzysztofowicz approach: Hydrologic Uncertainty Processor Prior: unconditional climatology Posterior: distribution of flow conditional on ensemble mean
B-C: preservation of space-time dependencies How to parameterize dependence? Space-time patterns of T and P Cross-variable dependence in T and P Critical for streamflow prediction Empirical approach Based on Schaake shuffle (Clark et al.) Shuffle the bias-corrected ensemble members to preserve rank-ordering of the raw ensemble members
Skill of T correction CRPSS = % gain over raw forecast ~20-60% gain Gradual decline with lead time
Skill of P correction ~20-30% gain Faster drop after 24 hour lead time
Skill of P correction for > 1-in-10 day observed P amount ~small gain or loss Failure of logistic regression to remove conditional bias (under-prediction of large P)
Skill of S for T and P correction. ~-10% to +10%
Skill of S for > 1-in-10 day observed S, with T and P correction. ~-40% to +20% Loss of skill at long lead times. Caution when correcting high P at long lead times!
Next steps Q1: What is the signal from bias-correction of forcing in streamflow?: Some way towards answering that question Need to establish why skilful forcing correction is not consistently translating into flow skill. Could it be due to the space-time and cross-variable dependence (Schaake Shuffle)? Try Brown and Seo (2011) approach to conditional bias (bias-penalized kriging) Next, well focus on Q2: Is the signal from forcing bias-correction lost following flow bias-correction?
Questions? (slides available from slideshare.net/janverkade) Contact: jan.verkade@deltares.nl, twitter.com/janverkade [email_address]

More Related Content

"Streamflow prediction in River Rhine: Exploring combinations of bias-correcting forcing and bias-correcting flow

  • 1. Streamflow prediction in River Rhine: Exploring combinations of bias-correcting forcing and bias-correcting flow Jan Verkade (Deltares and Delft University of Technology) James Brown (NOAA-NWS-OHD and UCAR)
  • 2. Motivation and research questions Biases/uncertainty in predicted forcing used for streamflow prediction: NWP models are skillful, but biased (mean, spread,..) This bias/uncertainty propagates from forcing to flow Bias-correction of precipitation is complex Ultimately, flow bias-correction is always needed Key research questions: What is the signal from bias-correction of forcing in streamflow? Is this signal maintained after bias-correction of flow , i.e. is forcing correction needed?
  • 3. Research design Raw forcing (T,P) Hydrologic model Raw flow Ensemble verification B-C forcing (T,P) Hydrologic model Raw forcing (T,P) Hydrologic model Scenario 1 Scenario 2 Baseline B-C streamflow B-C streamflow
  • 4. 油
  • 5. Data kindly provided by Florian Pappenberger @ ECMWF
  • 6. Observed forcing data: E-OBS dataset Downloadable from KNMI @ http://eca.knmi.nl/
  • 7. Bias-correction of temperature, precipitation and flow The random variables (one time/location): Predictand Y = observed temp/precip/flow. Assumed unbiased! Potential predictors X = {X1,,X5,, Xm}; biased. The bias-corrected forecast: How to parameterize for T and P? Parsimonious model (subject to skill!) Model the statistical dependence (traces)
  • 8. Bias-correction of temperature, precipitation and flow (2/2) Temperature normal regression: linear regression in normal space Precipitation logistic regression: linear regression in logistic space Streamflow: Krzysztofowicz approach: Hydrologic Uncertainty Processor Prior: unconditional climatology Posterior: distribution of flow conditional on ensemble mean
  • 9. B-C: preservation of space-time dependencies How to parameterize dependence? Space-time patterns of T and P Cross-variable dependence in T and P Critical for streamflow prediction Empirical approach Based on Schaake shuffle (Clark et al.) Shuffle the bias-corrected ensemble members to preserve rank-ordering of the raw ensemble members
  • 10. Skill of T correction CRPSS = % gain over raw forecast ~20-60% gain Gradual decline with lead time
  • 11. Skill of P correction ~20-30% gain Faster drop after 24 hour lead time
  • 12. Skill of P correction for > 1-in-10 day observed P amount ~small gain or loss Failure of logistic regression to remove conditional bias (under-prediction of large P)
  • 13. Skill of S for T and P correction. ~-10% to +10%
  • 14. Skill of S for > 1-in-10 day observed S, with T and P correction. ~-40% to +20% Loss of skill at long lead times. Caution when correcting high P at long lead times!
  • 15. Next steps Q1: What is the signal from bias-correction of forcing in streamflow?: Some way towards answering that question Need to establish why skilful forcing correction is not consistently translating into flow skill. Could it be due to the space-time and cross-variable dependence (Schaake Shuffle)? Try Brown and Seo (2011) approach to conditional bias (bias-penalized kriging) Next, well focus on Q2: Is the signal from forcing bias-correction lost following flow bias-correction?
  • 16. Questions? (slides available from slideshare.net/janverkade) Contact: jan.verkade@deltares.nl, twitter.com/janverkade [email_address]

Editor's Notes

  • #5: Basin: Rhine basin (approx. 160e3 km2), focus on Moselle sub-basin Available modeling system: HBV rainfall runoff model at daily time step Within a Delft-FEWS forecast production system (CHPS)
  • #6: 3164 issued reforecasts between 1991 and 2010 1/10th-by-1/10th degree 30 days leadtime we use 10 days only
  • #7: E-OBS is a daily gridded observational dataset for precipitation, temperature and sea level pressure in Europe. The full dataset covers the period 1950-01-01 until 2011-06-30. It has originally been developed as part of the ENSEMBLES project (EU-FP6) and is now maintained and elaborated as part of the EURO4M project (EU-FP7).