1. Medium-range ensemble prediction of hydrological droughts
Felix Fundel, Stefanie Jörg-Hess, Massimiliano Zappa
NRP-61 Drought-CH
NFP61 sustainable water management
Are we prepared for droughts?

2. Motivation Who could be interested in drought forecasts?
Hydro power generation, water quality, tourism, agriculture, power plant cooling etc.
Every interest group is affected differently, i.e. numerous drought indicators exist.
Here, only hydrological droughts are considered.

3. Hydrological droughts
runoff
time
threshold
severity
duration
Hydrological droughts are defined by looking on runoff only
From the runoff you can derive a number of drought indices
Therefore you define a threshold you consider as representative for a drought
You have a drought event whenever the runoff falls below the threshold
For each event you can calculate duration, severity (deficit) or magnitude (not independent of course)
Common drought characteristics (e.g. Tallaksen et al., 1997; Fleig et al., 2006)
Duration = length of longest forecast period below threshold
Severity = cumulative runoff deficit w.r.t. threshold
Magnitude = severity/duration

4. Study Site Area 1750 km2 Altitude 370m – 2350m runoff climatology
Gurtz et al., 1999 Hydrol. Proc.
Area km2
Altitude 370m – 2350m
runoff climatology
( )

5. Forecast system Validation ECMWF VarEPS PREVAH Meteorological forcing
Reforecast
4+1 members, weekly, lead-time 32 days
downscaling
bilinear interpolation (t, td, rr, rh, rad, v, ssd, alb)
lapse rate correction (t, td)
PREVAH
Hydrological model
Hydrotope units 500m x 500m
Initial states from observation based reference run
Gauge observations
Runoff forecast
Validation
Verification of 5 member daily mean forecast

6. Forecast quality Is this good (skillful)? At 32 d?
Only every other week
Those forecasts are quite common in operational hydrology today. Maybe not so much on time scales of 32 days, and there are good reasons for that:
Is this good (skillful)? At 32 d?

7. Forecast quality Limited predictability of peak-flow
This probabilistic verification of the forecast shown before as function of lead-time and runoff quantile form 0 to 1.
The score used is the 2AFC (two alternatives forced choice), a convenient measure, reflecting the potential forecast skill. Convenient because of its intuitive interpretation (.5 better than climatological guess), 1 perfect
And it can be used for probabilistic, continuous, ensemble , deterministic … forecasts.
You see that forecasts of peak-flow lose their skill rapidly, after day 10 there is hardly skill left.
The situation is better for low flow, where the predictability is extended to longer lead-times.
Now this is a verification on a daily basis, by using drought indices as described the forecast complexity is further reduced
VarEPS/PREVAH ensemble runoff forecast skill drops with lead-time
low-flow has higher predictability at longer lead-times
2AFC: Mason and Weigel 2009

8. Drought Indices Example
32 days runoff prediction  low-flow index
lead-time is no longer involved
For this study only the longest event within the 32d period was considered
Low-flow threshold 15th quantile (seasonal)

9. Low-Flow Indices
Gray areas give the ensemble range for each forecast, the red line shows the observed index in the 32d period of each forecast.
The resemblance is visible to the naked eye, especially the 2003 drought is standing out here.
Maybe underestimation in case of severity and magnitude
Why did I choose the 15th quantile as threshold?
Of course every other threshold might have been picked and be interesting for a specific user group.
Howerer a analysis of different the forecast quality of different thresholds revealed…

10. Dependence on choice of threshold
Forecast quality
Dependence on choice of threshold
2AFC for low-flow index forecasts
Mean forecast vs. observed index
Q15th (varying seasonal)
Good compromise of forecast quality and event frequency
The quality of low-flow index forecasts depends on the chosen threshold and seems to peak at around the 15th quantile.
Also picking lower thresholds had impact on the robustness of the scores.

11. Forecast quality
For the drought threshold of 15th quantile the properties of the drought index forecasts are as follows
Read points
The post-processing should at least correct for the underestimation
Not perfect but significant correlation of observation and forecast mean
Severity and magnitude forecasts underestimate
Duration forecasts seem to be reliable
Potential for statistical post-processing

12. Economic Value
Relative improvement in economic value when making forecast based decisions
Considering the cost/loss ratio of the forecast user
Considering different event thresholds (50th to 98th quantile of the low-flow indices)
1: Forecasts are most valuable for users whose C/L ratio is close to the climatological event frequency
One remaining question concerns the timing of the event within the forecast. E.g. a 10 day event could befrom day 1-10 in the forecast and day in the observations.
Obviously forecasts are most valuable for users with C/L around freqclim
Valuable forecasts of all low-flow indices are possible for all users

13. Timing of forecast events
observed
Yes No
42.5%
24.5%
7.5%
25.5%
Contingency table for drought forecast & obs.
forecast
Only 42% of hits do overlap
Forecast users should allow for some temporal fuzziness

14. 2006 Drought extremely hot period but shorter than 2003 Temp. anomaly
mean
min
max
Rebetez et al., 2009, Theor. Appl. Climatol.
Temp. anomaly
July 2006
July temp. distribution
( , France < 500m)
The 2006 heat wave was described and classified by Rebetez. In parts of Europe the monthly temperatures (min and max) were amongst the highest ever recorded. As at this time the 51 member monthly forecast system of ECMWF already was operational we took the opportunity to compare the runoff forecast during this event with those of the reforecasts. The questions were: was there an event? Was it forecast? What is the added value of having 51 members compared to only 5 in this case?
extremely hot period but
shorter than 2003

15. Forecasts using the operational 51 member VarEPS, the runoff is inverted to emphasis low-flow, also the runoff was normed so that the threshold (15th) equals 1.
Forecasts start at end of June and are then issued each successive week, whereas the start of the reforecast covering the event is a couple of days later.
Of course, forecasts and reforecasts are not based on the same model version. Which is why the predictions are not strictly comparable.

16. Summary Predictability of surface peak runoff is limited
Low-flow events are predictable longer in advance
Forecasting low-flow indices (averaging over several days) additionally enhanced predictability
Monthly low-flow index ensemble forecasts provide a valuable tool for decision making
Predictability of surface runoff is limited: depending on catchment an forecast system but generally not much longer than 10 days in advance
Low-flow events are predictable longer in advance: initial conditions get more important than the role of the meteorological forcing
Forecasting low-flow indices (averaging over several days) additionally enhanced predictability: by averaging the complexity is reduced and also skill from shorter lead-times is assigned to longer lead-times
Monthly low-flow index ensemble forecasts can be useful for decision making: monthly forecast of course are mainly interesting or sectors vulnerable on such short scales like power production or water quality management.