1. Application of a global probabilistic hydrologic forecast system to the Ohio River Basin
Nathalie Voisin1, Florian Pappenberger2, Dennis Lettenmaier1, Roberto Buizza2,
and John Schaake3
1 University of Washington
2 ECMWF
3 National Weather Service – NOAA
European Geophysical Union General Assembly , May

2. Background Existing Flood Alert Systems in mostly-ungauged basins
Limpopo 2000
Early Flood Alert System for Southern Africa (Artan et al. 2001)
South Asia 2000
Mekong River Commission – basin wide approach for flood forecasting
Bangladesh 2004
(Hopson and Webster 2010)*
Horn of Africa 2004
(Thiemig et al. 2010, EU - AFAS)*
Zambezi 2001,2007,2008
(EU-AFAS, in process)*
Message:
1. existing flood alert system in developed countries : many use ensemble weather forecasts, most important they use in-situ data:
rain gauge station observations, radars, river flow gauge etc.
And they are for one basin in particular
2. in ungauged basins:
- usually financed by international organization following desastrous events - case by case basis
Some are ensemble some are not
- basin dependent
- maintenance issue : educated people leave AND expensive
3. Goal here:
- develop a system hydrological forecast that would use ensemble weather forecast
- for large domains : not one basin at a time but a continent ( except for flow )
-global approach for calibrating the weather forecasts
- using only globally available tools, and can be run remotely to avoid cost and maintenance to population
TOO LONG
Global application i.e. have one consistent system over entire Africa and South East Asia run at once remotely:
-> main interest: reduce the financial burden of maintenance and operation
 cover entire continent instead of being on a basin per basin basis ( AFAS like)
* Ensemble flow forecasting

3. Objective
Develop a medium range probabilistic quantitative hydrologic forecast system applicable globally:
Using only (quasi-) globally available tools:
Global Circulation Model ensemble weather forecasts
High spatial resolution satellite-based remote sensing
Using a semi distributed hydrology model
applicable for different basin sizes, not basin dependent
flow forecasts at several locations within large ungauged basins
Daily time steps, up to 2 weeks lead time
Reliable and accurate for potential real time decision in areas with no flood warning system, sparse in situ observations (radars, gauge stations, etc) or no regional atmospheric model.
Note that similar to EFAS

4. Forecast scheme Today Initial State Voisin et al. (2010, in review)
Explain the scheme and state how it differs from the existing systems in ungauged basins ( global application)
Today
Initial State

5. Science Questions What is the forecast skill of the system?
What are the resulting hydrologic forecast errors related to errors in the calibrated and downscaled weather forecasts?
Is the forecast skill different for basins of different size?

6. Ensemble precipitation forecast calibration and downscaling
Shows snapshots of two observed events:
First one is the TMPA largest basin average precip day in the period
-> interpolation-like for RMSD and RMSDmean because…..
-> unrealistic spottyness for BCSD
- Second is a day with significant precipitation (upper tercile ) but without interpolation substitution in RMSD and RMSDmean
All look more realistic than interpolation method : higher precipitation values in specific grid cells, precipitation patterns.
Analog method vs interpolation:
- maintained resolution & discrimination
slightly lower predictability
BUT largely improved reliability
smaller mean error
more realistic precipitation patterns

7. (substitute for observations, Climatology) Forecast –
Reference
(substitute for observations,
Climatology)
Forecast –
Clim & null Precip 4
15-member ensemble, 15-day daily forecast:
Day 1-10: ECMWF EPS fcst
Day 11-15: Zero precip. 1
Deterministic
15-day daily fcst
Day 1-15:
- Zero precip.
ECMWF analysis fields:
with TMPA precipitation
Daily, period, degree 2 3
ECMWF EPS fcst
Interpolated to .25o
ECMWF EPS fcst
Calibrated & downscaled
( analog method) …
Initial hydrologic state
VIC
period
VIC
15-day simulation …
Daily
simulated runoff,
soil moisture, SWE
Substitute for observed
runoff
15-member ensemble
15-day distributed
runoff forecast
deterministic
15-day distributed
runoff forecast …
Initial
flow conditions
Routing model
15-day simulation
Routing model
period …
15-member ensemble
15-day flow forecast
at 4 stations with
different drainage areas
15-day deterministic
flow forecast
at 4 stations with
different drainage areas
simulated daily flow
Substitute for observations

8. Calibration of the hydrology and routing models
→ Use “simulated observed flow” as reference
(ECMWF Analysis and TMPA precipitation)
→Focus on weather forecasts errors
No flow observation uncertainties
No hydrology model and routing model ( structure, parameter estimation) uncertainties
Re-explain briefly why the application is over the Ohio River Basin ( previous work)

9. Verification of ensemble runoff forecasts
Ohio River Basin day forecasts (10 day fcst, +5 days 0-precip) o grid cells
CRPS is a great measure for probabilistic fcst:
Handy because get a measure for one forecast
BUT very sensitive to the size of the ensemble spread ( range)

10. Ensemble flow forecasts verification

11. Ensemble flow forecasts verification
Ensemble reliability at Metropolis and Elizabeth

12. Conclusions
A preliminary probabilistic quantitative hydrologic forecast system for global application was developed and evaluated:
Skill for 10 days for spatially distributed runoff
Skill for day forecasts depending on concentration times at the flow forecast locations
For small basins : skills for 10 days, with good reliability for short lead times
For larger basins: for 10 days + concentration time
Ensemble weather forecasts need to be calibrated:
for better hydrologic probabilistic forecasts ( reliability )
For better forecast accuracy in sub basins locations
Will incorporate PUB and HEPEX results and ideas. ( PUB: Predictions in Ungauged Basins HEPEX: Hydrologic Ensemble Prediction Experiment)

13. Thank you – Questions?

14. Forecast Verification
Which forecasts?
Spatially distributed ensemble runoff forecasts
Ensemble flow forecasts at 4 locations
Verification:
Deterministic Forecast Skill Measures:
Bias ( accuracy, mean errors)
RMSE (accuracy)
Correlation (accuracy, predictability)
Probabilistic Forecasts Skill Measures:
Continuous Rank Probability Skill Score (accuracy, reliability, resolution, predictability)
Rank Histograms ( ensemble spread i.e. probabilistic forecast reliability)
For forecast categories: What can I expect when a forecast falls in a certain forecast category? ( oriented for real-time decision )

15. Calibration of the hydrology and routing models
-Differences between TMPA and observed precipitation
-Daily flow fluctuations due to navigation, flood control, hydropower generation
Uncertainties in VIC and routing models physical processes, structure and parameters
Re-explain briefly why the application is over the Ohio River Basin ( previous work)
→ Use “simulated observed flow” as reference
→Focus on weather forecasts errors

16. Ensemble forecast verification
Relative Operating characteristic (ROC) Plot Hit Rate vs. False Alarm Rate for a set of increasing probability threshold to make the yes/no decision. Diagonal = no skill Skill if above the 1:1 line Measure resolution A bias forecast may still have good discrimination.

17. Ensemble Forecast Verification
Ensemble reliability:
Reliability plot: PROBABILISTIC fcsts
Choose an event = event specific
Each time the event was forecasted with a specific probability ( 20%, 40%, etc), how many times did it happen ( observation >= chosen event). It requires a sharpness diagram to give the confidence in each point. It should be on a 1:1 line.
Talagrand diagram (rank): PROBABILISTIC QUANTITAVE fcsts
Give a rank to the observation with respect to the ensemble forecast ( 0 if obs below all ensemble members, Nmember + 1 if obs larger )
Is uniform if ensemble spread is reliable, (inverse) U-shaped if ensemble is too small (large), asymetric is systematic bias.

18. Continuous Rank Probability Score
Probabilistic quantitative forecast verification
measures the difference between the predicted and observed cumulative distribution functions: resolution, reliability, predictability
For one forecast(gridcell, lead time, t): d1 d2 d3
dNmember
magnitude
Prob Fcst
∆P12
∆PN2 1
forecast is best. 1 1 1