Eric Rothwell (Reclamation), Levi Brekke (Reclamation), Andy Wood (NCAR), and Jeff
Arnold (USACE)
Co-Investigators:
(NCAR) AJ Newman, K Sampson, TM Hopson, M Clark, (Univ. WA) B Nijssen
Reclamation PN – NWS NWRFC Coordination Call, January 15, 2015
• Developing solutions for water & power management
– Traditional categories
• environmental stewardship for water delivery and management
• water and power infrastructure reliability
• water operations decision support
• conserving or expanding water supplies
– Priority themes intersecting research categories:
• advanced water treatment
• climate change and variability (CCV)
• invasive quagga mussels
• renewable energy and energy conservation
• Sustainable water infrastructure and safety
– S&T Portfolio: ~100 to 150 projects, ~8-9M per year
• recent CCV activity: ~1.5 M/yr; competed and facilitated projects
• projects led by internal Pis (Regions, TSC) and/or external collaborators
Portfolio Categories
1.
Long-term Climate Change
Impacts
2.
Short-term Climate
Variability, from Floods to
Droughts (focus today)
3.
Data, Tools, and Training
Resources
4.
Scoping
• Defining User Needs, Developing Research Strategy
• Fostering collaborative R&D
• Developing Training Resources
• Hosting Workshops on Emerging Topics (e.g.,
Nonstationarity, Portfolio of Assessemtn Approaches)
www.ccawwg.us
S&T Project 2264
Application of a Physically-based Distributed Snowmelt Model in Support of Reservoir Operations and Water Management
Objectives Graphic
Testing the application of a physically-based distributed snow model in an operational forecast setting. Are these modeling techniques appropriate to operational needs and do the deliverable products improve reservoir management outcomes?
Approach
1.
Application of Isnobal in the Boise River
Basin to provide maps of SWE and snowcover energy state
2.
Coupling of Isnobal to soil storage or routing model
3.
Proof of concept – using the coupled models with short-term weather forecasts to forecast reservoir inflows
Partners NRCS, BSU
Milestones
1.
Spring of 2014 ARS provided weekly SWE maps
2.
August 2015 present POC, comparison of historic results, and review of integrating model outputs into operations.
• …
S&T Project 9682
Intermediate-range Climate Forecasting to Support Water Supply and Flood Control with a Regionally Focused Mesoscale Model
Objectives
Develop mesoscale weather prediction models, tailored to regional characteristics, to provide hydroclimate forecast data to test accuracy and spatiotemporal coverage.
Graphic
Approach
1.
Develop WRF simulation model for a region that includes portions of the PN, Upper
Colorado, and Great Plains.
2.
Perform historical re-forecasts for a range of years.
3.
Quantify and characterize the accuracy of data products from WRF, and how could they enhance run-off forecast products.
Partners USDA-ARS
Milestones
1.
Development of the WRF simulation model for a domain encompassing portions of the headwaters of the PN, UC, and GP Regions.
(July 2015)
2.
Hold a stakeholder meeting to illicit feed back on initial WRF simulation model
(August 2015).
S&T Facilitated Research:
Airborne Snow Observatory (ASO): Value of Information for supporting Snowmelt Reservoir Operations
Objectives Graphic
ASO technology provides unique wall-to-wall monitoring of basin snowpack and dust conditions.
Goal is to assess value of ASO monitoring in the context of Reclamation’s snowmelt management during late Spring to early Summer, focusing on western Colorado.
Approach
1.
Focus on two reservoir systems and basins
(Gunnison, San Juan)
2.
Emulate ASO: (a) synthetic “truth” hydrology at high-resolution, (b) simulated snowpack observation using point and ASO schemes
3.
Hydrologic Hindcasts: (a) current model + point sensing, (b) improved model + point sensing, (c) improved model + ASO
4.
Operations Hindcasts: informed by 3.
5.
Assess value of operational effects
Partners NASA JPL, Reclamation
UC, Reclamation TSC
Milestones
1.
(FY15) Tasks 1-3 Q2, Task 4 Q3, Task 5 Q4
2.
Plan to have results meeting and next-steps discussion during Q4 timeframe; include
RFCs.
S&T Facilitated Research: (FY13-15) The Predictability of
Streamflow across the Contiguous United States
(FY15-17) Experimental Demonstration and Evaluation of
Objectives
Real-time, Over-the-Loop Streamflow Forecasting
Graphic
Evaluate work-stage opportunities for improving streamflow forecasts (days to seasons).
Assess how these opportunities perform under historical hydroclimate variability (hindcasts) and, subsequently, within a real-time forecasting environment. In the latter, feature forecaster over-the-loop workflow and assess its benefits and disadvantages.
Approach
1.
Implement CONUS-wide watershed simulation framework with automated model application and forcing generation.
2.
Assemble building blocks for forecasting improvement (described later)
3.
Conduct hindcast experiments and operations impact evaluations using these building blocks.
4.
Evaluate building blocks in an experimental, real-time forecasting evaluation with forecaster over-the-loop workflow.
Partners NCAR, USACE, Reclamation,
University of Washington
Milestones Opportunities to engage operators and RFCs to review/discuss:
1.
3/15/15: Hindcasts on effects of alternative historical and future forcing generation.
2.
6/15/15: Hindcasts on effects of data assimilation and post-processing.
3.
9/1/15: (a) Hindcasts on effects of alternative model and calibration approaches, (b) draft real-time system specs
More detailed briefing on:
(FY13-15) The Predictability of Streamflow across the
Contiguous United States
(FY15-17) Experimental Demonstration and Evaluation of Real-time, Over-the-Loop Streamflow Forecasting
Andy Wood, NCAR
Hydrologic Model Analysis
Analysis products model outputs
Forecast precip / temp
River
Forecasting
System
+
Update
Model states
Observed
Data
Analysis &
Quality
Control
Outputs
Graphics
River
Forecasts parameters
Models
Calibration
Support
W.M.
Decisions
1) alternative hydrologic model(s),
2) new forcing data/methods (eg, QC) to drive hydrologic modeling
3) new calibration tools to support hydrologic model implementation
4) Improved data assimilation to specify initial watershed conditions for hydrologic forecasts
5) new data and methods to predict future weather and climate
6,7) methods to post-process streamflow forecasts and reduce systematic errors
8) benchmarking / hindcastsing / verification system / ensembles (not shown)
Streamflow Prediction System Elements
Science Questions & Approach
• Questions:
– For different types of forecasts and user needs, what method or data improvement opportunities are most promising?
– How do these opportunities fare under historical hydroclimate variability ?
– How do these opportunities fare in a real-time forecasting environment ?
• To develop answers to above question, we’ve:
– Implemented a CONUS-wide watershed simulation framework with automated model application and forcing generation.
•
Opportunity
: Create “many basins” platform for forecasting application and evaluation
• Benefit : Permits efficient study of forecasting elements (model, forcing, data assim, etc.) under a variety of basin and climate conditions
•
Specs : Newman et al. 2014
– Base model: National Weather Service operational Snow-17 and Sacramento-soil moisture accounting model (Snow-17/SAC)
… more models to be added
– Locations: 670 basins from GAGES-II, Hydro-climatic data network (HCDN)-2009
– Forcings: DAYMET ( http://daymet.ornl.gov/ ), NLDAS, and Maurer et al. (2002) for (a) lumped (Snow-17/SAC apps), (b) hydrologic response unit (from PRISM), and (c) elevation band
– Calibration: automated Shuffled Complex Evolution (SCE) global optimization routine: 15 years, validation on remaining data for all lump forcing types; areas with seasonal snow, frequent precipitation perform best; high plains, desert SW perform worse
Newman, A. J., et al.
2014: “Development of a large-sample watershed-scale hydrometeorological dataset for the contiguous USA:
Dataset characteristics and assessment of regional variability in hydrologic model performance,” HESS, in press.
Science Questions & Approach
• Questions:
– For different types of forecasts and user needs, what method or data improvement opportunities are most promising?
– How do these opportunities fare under historical hydroclimate variability ?
– How do these opportunities fare in a real-time forecasting environment ?
• To develop answers to above question, we’ve:
– Implemented a CONUS-wide watershed simulation framework with automated model application and forcing generation.
– Assembled building blocks for forecasting improvement (next slides)
• Opportunity: Develop ensemble historical CONUS forcing dataset
• Benefits : supports (1) more robust historical calibration, (2) ensemblebased data assimilation to initialize forecasts
• Specs : 1/8º grid, informed by 12,000+ stations, 100 ensemble members
• Example: Example June 1993 precipitation
• two example members (a-b)
• ensemble mean (c)
• ensemble standard deviation (d)
Newman, A. J., M. P. Clark, J. Craig, B. Nijssen, A. Wood, E. Gutmann, N. Mizukami, L. Brekke, and J.R. Arnold, 2015:
“Gridded
Ensemble Precipitation and Temperature Estimates for the Contiguous United States ,” in development.
• Opportunity: automated assimilation of observed streamflow (flood forecasts) and snow water equivalent (seasonal forecasts)
• Benefits : improves initialization of watershed states, replaces manual modifications in forecasting process
• Specs : apply particle filter (PF) with uncertainty from ensemble forcings
Figure: Particle Filter based ensemble DA with 6-hour update cycle automatically adjusts SAC model to correct for model and forcing errors
Figure: RMSE of forecasts with DA using PF, EnKF and AEnKF, versus the raw forecast
• Opportunity: downscaled ensemble met forecasts enable estimation of prediction uncertainty
• Benefits : supports risk-based approaches for forecast use
• Specs : use locally-weighted multi-variate regression to downscale
GEFS (reforecast) atmospheric predictors to watershed precipitation and temperature
Figures: Case study hindcast of
15-day ensemble forecast including
7 days of downscaled GEFS as met forecast
(Snow17/SAC model)
• Opportunity: contrast ability of different modeling approaches to capture hydrologic variability and response
• Benefits : provides broader array of modeling options for forecasting
• Specs : baseline is NWS models (Snow17/Sac/UH/etc, lumped); alternatives include gridded VIC, SUMMA in various configurations
Figures: Exploring various model configurations and physics (HRU Snow17/SAC, band SUMMA)
Snow17-lump
Snow17-hru
Snow17-band
SUMMA-band
• Opportunity: seasonal climate forecasts can add information to seasonal streamflow predictions
• Benefits ?
increased skill benefits water supply forecasts and associated applications
• Specs : use ESP trace-weighting approaches based on likelihood from principle component regression of predictors including climate system indices and climate forecasts
• Opportunity: Apply statistical adjustments to raw streamflow forecasts based on past forecast performance or observable error at initiation time
• Benefits ?
reduces systematic forecast errors to improve forecast reliability
• Specs : use linear damping of error at forecast start; other approaches to be added
Science Questions & Approach
• Questions:
– For different types of forecasts and user needs, what method or data improvement opportunities are most promising?
– How do these opportunities fare under historical hydroclimate variability ?
– How do these opportunities fare in a real-time forecasting environment ?
• To develop answers to above question, we’ve:
– Implemented a CONUS-wide watershed simulation framework with automated model application and forcing generation.
– Assembled building blocks for forecasting improvement (next slides)
– Initiated hindcast experiments and operations impact evaluations using these method improvement (FY15)
• Objectives:
• Evaluate alternative process variations
• Specify hindcast experiments to address specific questions
• Inform future real-time system design
• Forecast Types
• Flood: run 5-10 years of daily updating, ensemble flood hindcasts with leads 1-7 days, for different process variations.
• Seasonal: run 30+ years of weekly updating ensemble seasonal hindcasts with lead time 1 year
Reference:
RFC Archived
Forecasts
Area
Model
Calib / Spinup
Forcing
Calib. Param.
Future
Forcing
Data Assim.
Post-Process.
FF1
FF2
DA1
Alternative
Future
Forcings?
DA2
Alternative
Data
Assimilation?
HF2
HF1
Alternative
Historical
Forcings?
Reference:
RFC Archived
Forecasts
MC1
Alternative
Model or
Calibration?
MC2
MC3
Area
Model
Calib / Spinup
Forcing
Calib. Param.
Future Forcing
Data Assim.
Post-Process.
March 15
HF1
SMA/Snow
17
Daymet
SCE
HF2
SMA/Snow
17
Newman
Ens v0+
SCE
FF1 FF2 DA1
SMA/Snow
17
SMA/Snow
17
SMA/Snow 17 best forcing best forcing best ens forcing
SCE SCE SCE
GEFS control none
Linear blend
GEFS control none
GEFS DS ens v0 none
Linear blend Linear blend
GEFS DS ens v1 none
Linear blend best GEFS particle filter
Linear blend
June 15 September 1
DA2 MC1
SMA/Snow 17 SMA/Snow
17 best ens forcing best forcing
SCE MOCOM best GEFS best GEFS
MC2 MC3
VIC SUMMA band/hru best forcing best forcing
MOCOM SCE best GEFS best GEFS
EnKF
Linear blend best DA best DA
Linear blend Linear blend best DA
Linear blend
CF3
CF2
Alternative
Climate
Forecasts?
CF1
DP1
DP2
Alternative
DA or Post-
Processing?
Reference:
RFC Archived
Forecasts or
ESP
DP3
MC1
MC2
Alternative
Model and
Calibration?
March 15 June 15
MC3
September 1
Area
Future
Forcing
Data Assim.
CF1 CF2
Model
Calib / Spinup
Forcing
SMA/Snow 17 SMA/Snow
17 best forcing
(see HF)
Calib. Param.
best calib best forcing
(see HF) base calib
CF3 statistical
(eg, PCR) mixed obs
NA
ESP + indexbased wgts.
none
ESP + CFSbased wgts.
none pred. clim.
fields none
Post-Process.
ST blend ST blend NA
DP1 DP2
SMA/Snow
17 best forcing
(see HF) best calib
SMA/Snow
17 best forcing
(see HF) best calib best clim.
forecast
SWE (PF or EnKF)
ST blend best clim.
forecast none
ST blend + regr./analog
DP3 MC1
SMA/Snow 17 Alt model / calib #1 best forcing
(see HF) best calib best forcing
(see HF) best clim.
forecast
Alt model / calib #1 best clim.
forecast
SWE (PF or
EnKF)
ST blend + regr./analog best DA +
PP combo
MC2
Alt model / calib #1 best forcing
(see HF)
Alt model / calib #1 best clim.
forecast best DA +
PP combo
MC3 multi-model as configured as configured as configured as configured
• 50 watersheds (and growing), chosen for varying hydro-climates & regions, being relatively unimpaired, and supplying reservoir inflows http://www.ral.ucar.edu/staff/wood/case_studies/
• Basins & Offices:
– Basins: See 50 case study watersheds
– Offices: at least Reclamation PN & GP; and USACE NWS (Seattle
District); aiming to include more …
• Milestone #1) Late March
– NCAR briefing on Flood (HF, FF) and Seasonal (CF); include RFCs
– Operators review, react, provide feedback on mgmt relevance
• Milestone #2) Late June
– NCAR briefing on Flood (DA) and Seasonal (DP); include RFCs
– Operators review, react, provide feedback on mgmt relevance
• Milestone #3) Early September
– NCAR briefing on Flood (MC) and Seasonal (MC); include RFCs
– Operators review, react, provide feedback on mngt relevance
– Operators / RFCs provide suggestions on real-time forecasting workflow, products stream, etc.
• Basins & Offices:
– Columbia-Snake Headwater Basins, tbd;
– Reclamation PN (Boise) and USACE NWS (Seattle District)
• Emulate how operators…
– (1) use forecasts (which are obtained?), (2) plan operations (how do obtained forecasts lend influence?), and (3) operate (roll ahead system states between forecasts)
• Forecast Process Variants:
– Reference
– tbd, likely Reference, DA1 (for Flood event hindcasting) and DP3 (for
Seasonal event hindcasting)
– Focus on set of past difficult events, floods to droughts
• Why only a set of events?
We can’t do full period analysis because we don’t have built models that emulate short-term ops process.
• Share preliminary findings at Milestone #3) Early September
Science Questions & Approach
• Questions:
– For different types of forecasts and user needs, what method or data improvement opportunities are most promising?
– How do these opportunities fare under historical hydroclimate variability ?
– How do these opportunities fare in a real-time forecasting environment ?
• To develop answers to above question, we’ve:
– Implemented a CONUS-wide watershed simulation framework with automated model application and forcing generation.
– Assembled building blocks for forecasting improvement (next slides)
– Initiated hindcast experiments and operations impact evaluations using these method improvement (FY15)
– Scoped a follow-on, experimental, real-time forecasting evaluation
(FY16-17).
Hindcasts FY15; Real-Time Forecasting effort FY15Q4 - FY17
Objectives:
1) Test advanced techniques in “forecaster over-the-loop” system.
2) Generate real-time flow forecast products similar to those from the
RFCs, as well as other information; display/disseminate on website
– daily to subdaily update flood forecasts; monthly to seasonal forecasts
– verification (real-time + long-term), trailing forecasts, uncertainty (spread); other water balance variables (forcings, snow/soil moisture), retrospective climatologies, archived hindcasts for past events
– real-time reliability less than RFC
– transparency on data & methods
3) Interact with operational partners regularly.
– feedback on products and guidance on development
– gain insight into user decisions, tailoring product formulation
– mode of interaction & frequency TBD as project evolves
• subject to partner interest & availability
• Proof-of-concept Research
– Aim to assess and evaluate forecasting methods relevant to
RFC practices and short-term water management.
• Two-way Education Opportunity
– (1) Reclamation & partners hear from RFCs about how projects and findings resonate with their practices, to what degree
– (2) RFCs learn about projects, potentially inform future workflow planning and/or more targeted collaborations w/ Reclamation & partners.
• Data Sharing during implementation
• 2264:
– Spring of 2015 ARS will make weekly SWE maps available
– August 2015 present POC, comparison of historic results, and review of integrating model outputs into operations.
• 9682:
– July 2015: Development of the WRF simulation model for a domain encompassing portions of the headwaters of the PN, UC, and GP Regions.
– August 2015: Stakeholder meeting to get feedback on initial WRF simulation model.
• ASO Value of Information
– Summer 2015: Results meeting and next-steps discussion
• NCAR-led effort: Meetings to review
– (late March) hindcasts on effects of alternative historical and future forcing generation methods
– (mid June) hindcasts on effects of data assimilation and post-processing.
– (early September) (a) Hindcasts on effects of alternative model and calibration approaches, (b) draft real-time system specs
– Andy Wood ( andywood@ucar.edu
), Martyn Clark (NCAR, mclark@ucar.edu
), Andy Newman ( anewman@ucar.edu
),
Pablo Mendoza ( medoza@ucar.edu
)
– Jeff Arnold (USACE, Jeffrey.R.Arnold@usace.army.mil
)
– Levi Brekke (Reclamation, lbrekke@usbr.gov
)
– University of Washington (Bart Nijssen)
– Agencies (e.g. RFCs, USACE & Reclamation field offices)
– More welcome!
• Newman, A. J., M. P. Clark, J. Craig, B. Nijssen, A. Wood, E. Gutmann, N.
Mizukami, L. Brekke , and J.R. Arnold, 2014: “Gridded Ensemble Precipitation and
Temperature Estimates for the Contiguous United States,” in development.
• Newman, AJ, MP Clark, K Sampson, AW Wood , LE Hay, A Bock, R Viger, D
Blodgett, L Brekke, JR Arnold, T Hopson, and Q Duan, 2014, Development of a large-sample watershed-scale hydrometeorological dataset for the contiguous USA: dataset characteristics and assessment of regional variability in hydrologic model performance, Hydrol. Earth Syst. Sci. Discuss ., 11, 5599-5631, doi:10.5194/hessd-
11-5599-2014 (in press)
• Wood, AW, T Hopson, A Newman, L. Brekke, J. Arnold, M Clark, 2014, quantifying streamflow forecast skill elasticity to initial condition and climate prediction skill. J.
Hydromet.(in review)
•
Clark, MP, B Nijssen, JD Lundquist, D Kavetski, DE Rupp, RA Woods, JE Freer, ED
Gutmann, AW Wood, LD Brekke, JA. Arnold, DJ Gochis, and RM Rasmussen,
2014, A unified approach to hydrologic modeling: Part 1. Model structure, Wat. Res.
Rsrch (submitted)
• Clark, MP, B Nijssen, JD Lundquist, D Kavetski, DE Rupp, RA Woods, JE Freer, ED
Gutmann, AW Wood, DJ Gochis, and RM Rasmussen, DG Tarboton, V Mahat, GN
Flerchinger, and DG Marks, 2014, A unified approach to hydrologic modeling: Part
2. Comparison of alternative process representations, Wat. Res. Rsrch (submitted)
EXTRA SLIDES
Real-Time Forecasting Scope
Forecast
Element
1. Hydrologic Model
Operational
Practice
Legacy single-physics, spatially coarse and conceptual models from
1970s-1980s, support forecasting at limited number of river locations
2. Model Forcings
Mean areal averages of station-based precipitation and temperature, spatially and/or temporally disaggregated by radar
3. Model
Development
–
Parameter
Estimation
Manual calibration oriented toward reproducing daily streamflow; single parameter set
Contrast
Modeling system permitting multiple models and alternative physics portrayals, with spatially distributed multivariate predictions.
Probabilistic forcings at varying spatial scales with full meteorological forcing suite provides richer information base
Automated calibration
(multiple techniques, multivariate focus, multiple parameter sets
(e.g., wet/dry)
Impacts of adopting automated technique
Pros : Address modeling uncertainty; surpass conceptual model limitations for process representation
Cons : Model variations difficult or costly to maintain in single system, and support with training
Pros : Finer spatial discrimination represents more controls on watershed processes; estimates of watershed condition uncertainty possible
Cons : Spatially distributed parameters more difficult to estimate and probabilistic forcings costly to run
Pros: Represent parameter uncertainty, inform conditional model application (wet/dry), bring consistency and speed to calibration process
Cons : Possible loss of skill aspects perceived by forecasters to be important at individual locations; difficulty handling individual station data variations
Real-Time Forecasting Scope
Forecast
Element
4. Forecasting –
Data
Assimilation and
Initial Basin
Condition
Estimation
5. Forecasting
–
Estimating
Future Weather
6. Forecasting – post-processing of streamflow forecast to reduce errors
Operational
Practice
Contrast
Manual adjustment of model states to reflect station snow (SWE) observations and reduce for streamflow (Q) simulation errors
Manual adjustment of single-value streamflow forecasts based on forecaster intuition and awareness of impact thresholds.
Automated assimilation of multiple observed conditions (SWE, Q) to adjust model states via multiple statistical techniques (such as the particle filter)
Manually merged met. forecast grids from models and other NWS weather forecast offices to yield single-value mean areal meteorological forecasts; no use of climate predictions;
HEFS 1 and MMEFS 2
Automated downscaling and calibration of GEFS and CFSv2 or NMME ensembles, drawing from larger predictor suite; multiple techniques (e.g., analog, hybrid analog,
HEFS).
Automated application of multiple ensemble streamflow forecast calibration techniques to reduce systematic bias, spread and timing errors.
Leverages retrospective simulations and hindcasts.
Impacts of adopting automated technique
Pros : Supports reproducible updates for efficient, scalable forecast generation, avoids labor-intensive state modification
Cons : Performance of automated DA is still less well-understood than other forecast method areas; vulnerable to observed data errors if not caught
Pros : Automated process allows for rapid realtime updates; ensembles support quantification of forecast uncertainty; reproducibility enables hindcasting and verification, as well as method benchmarking
Cons : Nowcast range (1-12 hour) predictions may not integrate as many data sources as
RFC forecasters consider, and be less accurate.
Pros : Reproducible techniques that can be assessed and improved through verification, supporting a quantification of forecast uncertainty. Hindcastable.
Cons : Individual events may have regimerelated errors that can be perceived by forecasters but are difficult to detect from longterm statistical analysis.
FY15-17 Effort: aiming to kick off the experimental operational forecast system withing year 1