1. Long-lead streamflow forecasts: 2
Long-lead streamflow forecasts: An approach based on ensemble climate forecasts
Andrew W. Wood, Dennis P. Lettenmaier, Alan .F. Hamlet
University of Washington
NWS/OGP Climate Prediction Assessments Workshop
Alexandria, VA Oct, 2002
important points:
Arun Kumar works for the NCEP Environmental Modeling Center, Climate Modeling Branch and runs the global spectral model to produce forecasts.
this study period and region are important because there is an evolving drought that we would like to use as a test situation for the method’s skill.
this approach was previously applied to the East Coast during Summer 2000, when there was an evolving drought in that region – a paper evaluating the project is currently submitted to JGR-Atmospheres.

2. Overview Research Objective
To produce monthly to seasonal snowpack, streamflow, runoff & soil moisture forecasts for continental scale river basins
Underlying rationale/motivation
Global numerical weather prediction / climate models exploit SST – atmosphere teleconnections
Hydrologic models add soil-moisture / snowpack influence on future hydrologic conditions and streamflow (persistence)
important point(s):
the approach attempts to make use of forecast skill from 2 sources:
better understanding of synoptic scale teleconnections and the effects of persistence in SSTs on regional climate, as reproduced in coupled ocean-atmosphere models;
the macroscale hydrologic model yields an improved ability to model the persistence in hydrologic states at the regional scale (more compatible with climate model scales than prior hydrologic modeling).
Climate forecasts with monthly and seasonal horizons are now operationally available, and if they can be translated to streamflow, then they may be useful for water management.

3. Previous Experimental Applications
Columbia River Basin
Summer 2001 drought
East Coast,
Summer 2000 drought

4. UW Experimental West-wide hydrologic prediction system
climate model T & P output
NCEP, NSIPP, CCM, MPI
*ESP as baseline
(note: not using official tercile forecasts, yet)
Real-time
Ensemble Forecasts
Downscaling
VIC hydrologic simulations
Ensemble Hindcasts
(for bias-correction
and preliminary
skill assessment)
UW (Dennis Lettenmaier, Andy Wood and Alan Hamlet) has been exploring an approach for downscaling climate model ensemble forecast output for use in driving real-time hydrologic and streamflow forecast simulations (with the VIC hydrologic model). Currently they’re working with the NCEP GSM, but plan to add other model’s outputs now that they are available (thanks in part to IRI). Down the road, they also plan to use the official tercile format forecasts of some of the forecasting centers a basis for a parallel ensemble set.
As a baseline for comparison with the climate model ensembles, they have been using ESP, extended streamflow prediction, in which current hydrologic states are combined with an ensemble of historically observed met. traces in simulating hydrologic forecast ensembles.
They use both forecast and hindcast datasets. The hindcasts are useful for bias- correcting the forecast output before downscaling, and also can be run separately as a retrospective forecast assessment.
The current UW effort, which focused previously on the Columbia River Basin, is being expanded to the western US domain shown at upper right, and an effort is being made to forge connections with user groups so that forecast products can be tailored to their needs.
West-wide forecast products
streamflow
soil moisture, snowpack
tailored to application sectors
fire, power, recreation
current work
* ESP
extended streamflow prediction
(unconditional climate forecasts
run from current hydrologic state)

5. Simulations VIC model spin-up VIC forecast ensemble
start of month 0
end of month 6
Forecast Products
streamflow
soil moisture
runoff, snowpack
VIC model spin-up
VIC forecast ensemble
climate forecast information (from climate model output, or terciles)
VIC climatology ensemble
1-2 years back
NCDC meteorol. station obs. up to 2-4 months from current
LDAS/other meteorol. forcings for remaining spin-up
data sources
important point(s):
after bias correcting and downscaling the climate model forecasts, the procedure for producing hydrologic forecasts is as follows:
we spin up the hydrologic model to the start of the forecast using observed met. data (from 2 sources: NCDC cooperator stations through 3-4 months before the start of the forecasts, then LDAS 1/8 degree gridded forcings thereafter).
The GSM forecasts comprise 2 sets of ensembles, one for climatology and one for the forecast. The climatology ensemble yields a distribution of the conditions we’ve seen over the period , while the forecast ensemble yields the distribution of the conditions we might see for the next 6 months. Although the climatology ensemble is nominally unbiased against a simulated climatology based on observed met. data (rather than bias-corrected, downscaled GSM met. forcings), we compare the forecast and GSM climatology so that any unforeseen biases (resulting, perhaps, from the downscaling method) occur in both climatology and forecast. Eventually this cautionary step may be eliminated, and we’ll compare directly to the simulated observed climatology.
at the end of the spin-up period and one month before month 1 (out of 6) of the forecasts, we save the hydrologic model state. The state is then used for initializing the forecast runs. Through the first month, the model runs on observed data to the last date possible, then switches to the forecast data. Usually, we process the observed forcings up through the 15th to 25th of this initialization month, then the forecast forcing data carries the run forward for the remaining days in the month, and throughout the following 6 month forecast period. Note, the state files used for the climatology runs correspond to the spin-up associated with the particular year (out of ) from which the climatology ensemble member is drawn.
the spin-up period captures the antecedent land surface hydrologic conditions for the forecast period: in the Columbia basin, the primary field of interest is snow water equivalent.
forecast products are spatial (distributed soil moisture, runoff, snowpack (swe), etc.), and spatial runoff + baseflow is routed to produce streamflow at specific points, the inflow nodes for a management model, perhaps.
NOTE: In using climate model output, BIAS is a major obstacle.

6. Bias Example: NCEP Global Spectral Model (GSM)
important point(s):
here we see the biases from a spatial view -- note the large temperature bias in raw GSM forcing (for July), and in precip over the eastern edge of the Ohio Basin (as an example). In the third column of images, the biases have been removed by our method, so they should look like the first column (observed) – and they do.
Bias is removed at the monthly GSM-scale from the meteorological forecasts

7. GSM Bias Example (cont): for one cell over Ohio River basin
biases in monthly Precip & Temperature can be too large for use as hydrologic simulation inputs
important point(s):
the first hurdle in making forecasts is to overcome the regional biases. this and the next slide show the bias -- first for one GSM cell, where you can see plotted the observed forcings vs. the GSM raw climatology forcings (for the climatology period ‘79-’99). biases are so large in temperature that a hydrologic model would be blown out of the water.
this is from one set (May) of climatology & forecast ensembles back on the East Coast

8. UW climate model downscaling and bias adjustment approach
bias correction - percentile-based mapping of model output to climate model-scale observations (i.e., spatially averaged, temporally aggregated)
downscaling - interpolation of monthly anomalies to 1/8 degree, application to long term 1/8 degree observed means
disaggregation – by resampling observed daily sequences
Don’t need to go into much detail here – suffice it to say that the climate-model based hydrologic forecasting approach involves 3 steps.
The first, bias-correction, is done at the climate model scale, adjusting monthly outputs based on a comparison of the climate model climatology to the observed climatology for a concurrent period.
The second, downscaling, involves spatial interpolation of the bias-corrected forecast anomalies to the 1/8 degree hydrologic model grid, and application of these (multiplicative, for precip, and additive, for temp) to the observed 1/8 degree mean fields.
The third, temporal disaggregation, requires random resampling of daily observed sequences at 1/8 degree with subsequent shifting/scaling to reproduce the monthly forecast signal.
We’ve tried this in several test applications and it appears to work pretty well.

9. GSM Bias Example (cont): after procedure, most monthly biases removed
important point(s):
the first hurdle in making forecasts is to overcome the regional biases. this and the next slide show the bias -- first for one GSM cell, where you can see plotted the observed forcings vs. the GSM raw climatology forcings (for the climatology period ‘79-’99). biases are so large in temperature that a hydrologic model would be blown out of the water.
this is from one set (May) of climatology & forecast ensembles back on the East Coast

10. Comparison with Dynamical Downscaling
PCM: DOE Parallel Climate Model
(2.8 degree resolution)
RCM: PNNL Regional Climate Model
(1/2 degree resolution)
important point(s):
The streamflow results are shown for The Columbia River at the Dalles as two ensembles (climatology and forecast) plotted alongside each other, for each forecast month.
The clear message from this plot is that the forecast (red plusses) distributions are clustered significantly below the climatology distributions, to a diminishing extent at the end of the forecast period (when runoff is dominated more by precipitation than snowmelt).
Note, however, that the climatology period doesn’t include the recent end point, 1977, so if it did, the distributions might extend further down with respect to the forecast ensembles – not enough to change the significance of the results, however.
Forecasting approach after dynamical downscaling
used HISTORICAL climate scenario from PCM,
for 20 year period
Forecasting approach

11. Downscaling Method Comparisons
Domain and Model resolutions
important point(s):
The streamflow results are shown for The Columbia River at the Dalles as two ensembles (climatology and forecast) plotted alongside each other, for each forecast month.
The clear message from this plot is that the forecast (red plusses) distributions are clustered significantly below the climatology distributions, to a diminishing extent at the end of the forecast period (when runoff is dominated more by precipitation than snowmelt).
Note, however, that the climatology period doesn’t include the recent end point, 1977, so if it did, the distributions might extend further down with respect to the forecast ensembles – not enough to change the significance of the results, however.

12. pcm rcm
OBS
Downscaling Method Comparisons Precipitation downscaled vs. observed ( averages)
Methods – PCM vs RCM
I Interpolation
SD Statistical (Spatial) Downscaling alone
BCSD Bias-correction and SD
important point(s):
The streamflow results are shown for The Columbia River at the Dalles as two ensembles (climatology and forecast) plotted alongside each other, for each forecast month.
The clear message from this plot is that the forecast (red plusses) distributions are clustered significantly below the climatology distributions, to a diminishing extent at the end of the forecast period (when runoff is dominated more by precipitation than snowmelt).
Note, however, that the climatology period doesn’t include the recent end point, 1977, so if it did, the distributions might extend further down with respect to the forecast ensembles – not enough to change the significance of the results, however.

13. pcm rcm
Downscaling Method Comparisons Temperature downscaled vs. observed ( averages)
OBS
important point(s):
The streamflow results are shown for The Columbia River at the Dalles as two ensembles (climatology and forecast) plotted alongside each other, for each forecast month.
The clear message from this plot is that the forecast (red plusses) distributions are clustered significantly below the climatology distributions, to a diminishing extent at the end of the forecast period (when runoff is dominated more by precipitation than snowmelt).
Note, however, that the climatology period doesn’t include the recent end point, 1977, so if it did, the distributions might extend further down with respect to the forecast ensembles – not enough to change the significance of the results, however.

14. Downscaling Method Comparisons Columbia River Basin Averages (1975-95)
hydrology based on
downscaled vs. observed
( averages)
important point(s):
The streamflow results are shown for The Columbia River at the Dalles as two ensembles (climatology and forecast) plotted alongside each other, for each forecast month.
The clear message from this plot is that the forecast (red plusses) distributions are clustered significantly below the climatology distributions, to a diminishing extent at the end of the forecast period (when runoff is dominated more by precipitation than snowmelt).
Note, however, that the climatology period doesn’t include the recent end point, 1977, so if it did, the distributions might extend further down with respect to the forecast ensembles – not enough to change the significance of the results, however.

15. Downscaling Method Comparisons Streamflow
based on: downscaled vs. observed P & T ( averages)
important point(s):
The streamflow results are shown for The Columbia River at the Dalles as two ensembles (climatology and forecast) plotted alongside each other, for each forecast month.
The clear message from this plot is that the forecast (red plusses) distributions are clustered significantly below the climatology distributions, to a diminishing extent at the end of the forecast period (when runoff is dominated more by precipitation than snowmelt).
Note, however, that the climatology period doesn’t include the recent end point, 1977, so if it did, the distributions might extend further down with respect to the forecast ensembles – not enough to change the significance of the results, however.

16. Summary Comments about Approach
Climate-hydrology forecast model system has potential
only if model biases are addressed
should be compared with current forecast practices, and with other experimental approaches
performs as well as dynamical downscaling approach, and is simpler to implement
Critical needs
access to quality met data during spin-up period
ability to demonstrate / assess skill quantitatively, hopefully aided by what we learn from retrospective assessments (hindcasts)

17. Sample Results from Recent Work
Current Objectives
Implement climate-hydrology model forecast system over western U.S. domain
Assess skill of approach with respect to traditional standards such as ESP and climatology, using retrospective analysis

18. Recent results: hindcast analysis
Columbia R. Basin
basin-averaged anomalies
GSM 6-month hindcasts
JAN initiation date
(shown: last 5 years of
retrospective analysis period)
Here, the retrospective assessment (using GSM hindcast and ESP ensembles) was conditioned on there being strong SST anomalies in the Pacific. The analysis was of basin-averages of monthly total precip, and monthly average temperature, runoff (ROBF), soil moisture (SM) and snow water equivalent (SWE). The latter three variables were verified against retrospective hydrologic simulations driven by observations of the first two.
GSM temperature forecasts showed skill relative to ESP, but precip did not. Runoff and soil moisture were predicted worse than with ESP, but SWE, particular in late spring/early summer, was predicted better.
Based on analyses like these, we’d like to develop a method of mixing the skillfully forecast output variables from the climate models and ESP, to create a hybrid or multi-model forecast ensemble that potentially is more skillful than any single model or ESP alone. In this instance we would likely discard the GSM precip forecast and replace it with an unconditional one, while retaining the GSM temperature forecast. The method for doing this, however, is not straightforward.
ens. fcst
obs
ens. fcst median
obs terciles

19. Recent results: streamflow
RMSE-Skill Score
JAN forecast of FEB-JUL flow
Columbia R. Basin hindcast analysis
GSM- and ESP-derived ensembles
for , all years
using RMSE-skill score wrt.
climatology
Results
Both ensembles show skill (from initial conditions), but ESP outperforms GSM in most locations
(in figure, larger circle = higher skill)
Explanation
Poor precipitation simulation
in GSM JAN forecasts
Recently, we’ve been using the retrospective hindcasts to assess the skill of the climate model in forecasting streamflow and basin-averages of variables like soil-moisture. This slide and the next are examples of the types of results we’re getting.
Here, the relative skill of GSM & ESP based streamflow forecasts for 15 locations in the CRB are depicted with this “bubble map”. These are Jan 1 forecasts of cumulative Feb-Jul streamflow. Generally, ESP does as well or better than GSM for producing these forecasts – as can be seen when the red circles (ESP skill) are between the blue (GSM skill) and the black (perfect skill).
The skill measure is 1-rmse(GSM)/rmse(ESP).
If the result seems dismal, note this is just one time of the year, for one particular forecast statistic. There may be others, at other times of the year, for which GSM performs better.

20. THE FOLLOWING SLIDES MAY BE OF INTEREST DEPENDING WHAT YOU WANT TO DISCUSS

21. NRCS Collaboration ~30 Basins Forecast Target Period: Issue dates:
with WCC-NRCS-USDA
Comparison of retrospective forecast performance for:
~30 Basins
Forecast Target Period:
Arizona Jan-May
Central Apr-Jul
North Apr-Sept
Issue dates:
1st of Jan, Feb, Mar, Apr

22. Continuing Research Develop a framework for use with ESP, multiple models
GSM CCM3 NSIPP COLA ECHAM
ESP
seasonal forecast skill profiles
seasonal forecast skill profile
specific to basin & season
best model(s) forecast ensemble
develop logic for using/discarding/combining model/ESP forecasts
associate forecast with reliability discussion based on skill profiles of component model and/or ESP
multiple models can be used
- e.g., skill-weighted super ensembles
ESP is unconditional
resampling of observed climate

23. Progress and schedule task current FALL 2002 2003 domain
Columbia (CRB) California
Colorado, Great Basin, Rio Grande
hindcast ensemble analysis
NCEP, ESP
NSIPP, CCM, MPI
real time ensemble forecast
NCEP, ESP, NSIPP, CCM, MPI
multi-model ensemble
test for CRB, NCEP+ESP
all domain / all models
official tercile forecasts
NCEP, (ESP), NSIPP, CCM, MPI
This is just an overview of our progress and anticipated schedule for expanding the forecasting scope.
This fall, we’ll complete the NCEP GSM retrospective analysis and expand the domain to the rest of the west (California is nearly done now), prepare to start issuing real-time forecasts in December or January with GSM. We’ll also work on the method for multi-model ensemble formation, using GSM & ESP only, with the idea of including the other models later.
In the winter and spring of 2003, we will add ensembles from several other models, probably NSIPP, then CCM, and perform retrospective assessments as well.
The multi-model ensemble work will continue in 2003, and we will try to start producing hydrologic forecasts based on the tercile-based official forecast sets from, e.g., NCEP & IRI, so as to have a more direct connection to these forecast products.

24. Experimental Forecasting Approach Downscaling-Disaggregation Test
Process into the daily VIC-scale input time series
Force hydrology model to produce streamflow
Ohio R. Metropolis, IL
Start with GSM-scale monthly observed
T & P (“unbiased”) time series
Is simulated streamflow unbiased against observed streamflow?
Experimental Forecasting Approach Downscaling-Disaggregation Test