1
NCAR Networking DayBoulder, April 17, 2015
Advancing streamflow forecast science to support water management
Andy Wood NCAR Research Applications LaboratoryHydrometeorological Applications Program
Acknowledgements & Website
NCAR• Martyn Clark• Andy NewmanUniversity of Washington• Bart Nijssen
http://www.ral.ucar.edu/projects/hap/flowpredict/
Reclamation• Levi BrekkeUS Army Corps of Engineers• Jeff Arnold
Themes
• Motivation• An Evolution of Streamflow Forecasting Approaches• New Science Begets A New Challenge
Traditional Resources
1980s construct:• parsimonious watershed models
run on single PCs (or card decks + VAX)
• phone/mail transmission of data, forecast output bulletins
• manual synoptic weather analyses, rudimentary NWP
grey = inactive
Forecastprecip / temp
Wea
ther
and
Clim
ate
For
ecas
ts RiverForecastModelingSystem
parameters
Observed Data
Analysis &Quality Control
Calibration
modelguidance
Hydrologic Model Analysis
hydrologicexpertise &judgment
OutputsGraphics
River Forecasts
Decisions
Rules, values, other factors, politics
NWS River Forecast Process
Observed and Simulated not Tracking
Result: Improved observed period simulation
Runtime Modifications River Forecast Centers• MOD capability has been available in the NWS >30
years• Generic MOD capability implemented within FEWS• Extend capability to other users outside of OHD-
core models
MOD Interface
Calibration Climatology
Hydrologist render a Run-Timemodification to the SACSMA Modeland increases the lower zone primary and supplemental states
CHPS
Slide by H. Opitz, NWRFC
9
Examples: Nooksack RTypical situation during snowmelt: the simulation goes awry What can a hydrologist deduce from this simulation? As it is, blending simulation and obs gives an ‘unrealistic’ forecast
11
Examples: Nooksack RThe resulting simulation is better, hence the forecast is more confident Flows stay elevated, have diurnal signal of continued melt.
Paradigm Limitations
For major events• it is difficult to assess forecast
system skill
• influence of forecasters cannot be separated from influence of data, models methods
• forecast service scales with FTEs, not computers
For the development process• new science that requires
reproducibility cannot be integrated
• the value of science infusion is hard to quantify Gochis et al, BAMS, 2015
Forecasting resources have greatly expanded
1980s construct:• parsimonious watershed models
run on single PCs (or card decks + VAX)
• phone/mail transmission of data, forecast output bulletins
• manual synoptic weather analyses, rudimentary NWP
Since then:• supercomputing, desktop clusters• web data services and connectivity• GIS • high-res satellite DEMs & land cover• real-time remote sensing• dozens of complex land surface
schemes at fine scales• ESM at large scales• dozens of (better) NWP outputs &
ensembles
Since the late 1990s, this cornucopia of new resources has been applied toward increasingly extensive hydrologic analyses at increasingly fine scales.
grey = inactive
New resources create new capabilities
global meteorological and climate datasets, eg, precip timeseries & analyses (TMPA+GPCC)
global hydrologic modeling simulation (VIC)
global, yet fine resolution land cover and terrain analyses for routing and hydraulic features (SRTM)
eg. channel slope in Pakistan
A myriad of new ingredients exist for multi-scale hydrologic analysis (and forecasting)
Workflow/Data Management Platform
hindcasting, ensembles (uncertainty), benchmarking, real-time operations
Hydrologic Prediction Science now has a framework
Historical Forcings
We are now automating, to the extent possible, what forecasters did before
Spinup Forcings
Forecastand
Hindcast Forcings
Appropriate Hydro/Other
Models
Hydro/OtherObservations
Streamflow & Other Outputs
Products, Website
verification
post-processing,
forecast calibration
objective DA
(regional)parameter estimation
calibrated downscaling
feedback into component
improvements
auto QC
http://www.hepex.org/Since 2004, HEPEX has highlighted progress in key methods to make new systems work well
Continental-domain fine-scale flood forecasting systems have arrived
NCAR/Gochis
EFAS
Scottish Flood
Forecasting Service
NSSL
‘End Member’ Forecast System Approaches
Operational, 1980s => • Provide tailored, but limited
(mostly deterministic) forecasts that are inputs to water management
• Simple conceptual models that can be adjusted manually, in real-time
• Heavy use of model calibration• Reliance on human expertise at the
model/data system level.• Non-reproducible, non-scalable
forecast process• Run on small number of
workstations
Research/experimental 2000s =>• Provide non-tailored, centrally
produced forecasts, typically in form of percentile or frequency analyses
• ‘Physically-based’ high-dimensional models
• Little or no calibration• Automated forecast process,
reproducible• Ensemble outputs desired, leverage
supercomputing
grid2grid
End Member Forecast System Philosophies
Operational, 1980s => • “The models, data and systems will
always be inadequate, thus human expertise is needed to fix performance on the fly”
• “If the decisions using your model outputs require a certain answer, the models must be simple so that they can be adjusted to provide that answer”
Research/experimental 2000s =>• “The superior physics in new models
and datasets will yield good quality results”
• “Most problems can be fixed with higher resolution and even more detailed process representation”
• “Research-grade results should be operationally useful”
grid2grid
Flow Forecasting Catch-22
There is a key tradeoff in forecast system design- as the model resolution (time/space) becomes finer, the
uncertainty at the model scales increases … but the ability to characterize uncertainty falls
system scale/complexity
ability to assess uncertainty
localuncertainty
lowlow
high
high
low complexitycan run ensembles,calibrate, hindcast, post-process,run many thousands of variations
hyper complexitycannot calibrate,
no ensembles, hindcasting, or full verificationcan only run tens of variations
Short Range Flow Forecasting ObservationsThe ‘new forecasting’ in fact encounters many of the traditional paradigm challenges, e.g.,
- uncertain initial conditions (watershed moisture and energy, amt & distribution- depends on quality of spinup forcing, the model, flow obs, regulation info
- inconsistent real-time and retrospective forcings and analysis- uncertain future forcings (quality of met forecasts)
MOD name Count Descriptionaescchng 190 Snow areal extent changechgblend 578 Blend simulation with last observationignorets 8 Throw out timeseries input datamfc 133 Melt factor correction (change melt rate)sacco 529 Soil moisture content changessarreg 921 Reservoir regulation changetschng 8554 Alter a timeseries (ie, redraw a flow
forecast or obs)tschng_MAP 2136 Change precipitation forcings (obs,
forecast)tschng_MAT 461 Change temperature forcingsuadj 7 Change threshold for rain on snow to
cause melttotal/watershed/day
~1.25 For ~360 watersheds, for ~30 days
NWRFC Mods for 1 Month
parameter issues
input data issues
water regulation issues
and so on…
model issues
Sophisticated systems are not immune to the same broad range of uncertainty
system by Amy Sansone, Matt Wiley, 3TIER
New Paradigm Limitations
For major events• it is difficult to assess forecast system skill (also cannot be run for long
enough to see track record)
• the complex system may be off track for reasons that cannot be easily detected or corrected (science, data and technology gap)
For the development process• options may now be limited due to computational requirements
• a cumbersome system may make it hard to test new variations andinfuse new science
For extreme events, we likelystill can’t diagnose forecasterrors and describe forecast skill!
R2O/O2R – a trek requiring tradeoffs
Research
Operations
How can I carry my hyper-resolution
ESM data across this valley?
Still haven’t seen anything better over on
that far side … why cross over?
Two water agencies (USACE and Reclamation) are supporting NCAR (A. Wood) to assess and demonstrate (in
real time) the adequacy of new hydrologic prediction science for operational forecasting
R2O/O2R – elements of the path forward
Research
Operations
• Flexible, intermediate-complexity approaches are needed• Resolution choices must support key methods in hydrological prediction science• Development efforts must integrate people and knowledge from both R and O• International communities of practice can provide insight on choices