An experimental real-time seasonal hydrologic forecast system for the western U. S

An experimental real-time seasonal hydrologic forecast system for the western U.S.

• Andrew W. Wood and

• Dennis P. Lettenmaier

• Department of Civil and Environmental Engineering

• University of Washington

• Climate Diagnostics and Prediction Workshop

• Pennsylvania State University

• October 27, 2005

Outline

• Background – UW West-wide hydrologic forecasting system

• Preliminary multi-model ensemble work

• Final Comments

Background: UW west-wide system

• where did it come from?

• 1997 COE Ohio R. basin/NCEP ->
• -> UW East Coast 2000 (NCEP/ENSO) ->
• -> UW PNW 2001 -> UW west-wide 2003

• what are its objectives?

• evaluate climate forecasts in hydrologic applications
• seasonal: CPC, climate model, index-based (e.g., SOI, PDO)
• 16-day: NCEP EMC Global Forecast System (GFS)
• evaluate assimilation strategies
• MODIS snow covered area; AMSR-E SWE
• SNOTEL/ASP SWE
• evaluate basic questions about predictability
• evaluate hydrologic modeling questions
• role of calibration, attribution of errors, multiple-model use
• evaluate downscaling approaches

• what are its components?

CURRENT WEBSITE

Surface Water Monitor

Background: UW west-wide system

MAP LINKS TO FLOW FORECASTS

Background: UW west-wide system

Background: UW west-wide system

• what drives UW system activities?

• research goals:

• exploration of CPC & NCEP products
• data assimilation of NASA products
• Klamath Basin, Sacramento River (particularly Feather)

• collaborations:

• requests by WA State drought personnel
• Yakima-basin forecasts, Puget Sound
• SW Monitor type hydrologic assessment
• interests of Pagano, Pasteris & Co (NWCC):
• calibrated forecast points in Upper Colorado, upper Missouri R. basin, Snake R. basin
• spatial soil moisture, snow and runoff data
• one-off analyses
• other, e.g., U. AZ project with USBR in lower Colorado basin

Background: UW west-wide system

• research objectives include:

• climate forecasts

• data assimilation

• hydrologic predictability

• multi-model / calibration questions

LDAS models

Multiple-model Framework

Multiple-model Framework

• Models:

• VIC - Variable Infiltration Capacity (UW)

• SAC - Sacramento/SNOW17 model (National Weather Service)

• NOAH – NCEP, OSU, Army, and NWS Hydrology Lab

• Model Energy Balance Snow Bands

• VIC Yes Yes

• SAC No Yes

• NOAH Yes No

• Calibration parameters from NLDAS 1/8 degree grid (Mitchell et al 2004) – no further calibration performed

• Meteorological Inputs: 1/8 degree COOP-based, 1915-2004

Test Case - Salmon River basin (upstream of Whitebird, ID) - retrospective (deterministic evaluation): 25 year training 20 year validation

Individual Model Results

Individual Model Results

Individual Model Results

• VIC appears to be best “overall”

• Captures base flow, timing of peak flow
• Lowest RMSE except for June
• Magnitude of peak flow a little low

• SAC is second “overall”

• No base flow
• peak flow is early but magnitude is close to observed*

• NOAH is last

• No base flow
• peak flow is 1-2 months early and far too small (high evaporation)

Combining models to reduce error

• Average the results of multiple models
• Ensemble mean should be more stable than a single model
• Combines the strengths of each model
• Provides estimates of forecast uncertainty

Computing Model Weights

• Bayesian Model Averaging (BMA) (Raftery et al, 2005)

• Ensemble mean forecast = Σwkfk

• where
• fk = result of kth model
• wk = weight of kth model, related to model’s correlation with observations during training

Computing Model Weights

• We transform flows to Gaussian domain and bias-correct them before computing weights using the BMA software

• Western U.S. – many streams have 3-parameter log-normal (LN-3) distributions for monthly average flow

• Each month, for each model, is given distinct distribution, transformation, bias-correction

• Procedure

• monthly LN-3 transformation
• monthly bias correction based on regression
• BMA process to calculate monthly weights, statistics
• weights used to recombine models
• transform outputs back to flow units

Multi-model ensemble results

Multi-model ensemble results

Multi-model ensemble results

Multi-model ensemble results

Multi-model ensemble results

Multi-model ensemble results

Multi-model ensemble results

Multi-model ensemble results

• despite large biases, SAC had a stronger interannual correlation with observations than VIC

• post-processing fixes many of the biases

• BMA procedure only really uses the inter-annual signal supplied by the models

Follow-on questions

• Can we infer anything about physical processes from the ensemble weights?

• How will this work in the ensemble forecast context?

• in gaining forecast accuracy, might we lose the physical advantages of models?

• other ways of applying BMA? e.g., not monthly timestep; with different bias-correction & transformation

ongoing work

• RESEARCH -- RESEARCH -- RESEARCH

• assimilation of MODIS & other remote sensing

• climate forecast (CPC outlooks, climate model, index-based)

• downscaling

• shorter term forecasts (GFS-based)

• multiple-model exploration

• further development of SW Monitor

• generally, water / energy balance questions in face of climate change / variability

• HEPEX support

HEPEX western US/BC testbed

• Test Bed Leaders:

• Frank Weber (BC Hydro, Burnaby, British Columbia, Canada)

• Andrew Wood (University of Washington, Seattle, USA)

• Tom Pagano (NRCS National Water and Climate Center, Portland, OR)

• Kevin Werner (NWS/WR)

• focus:

• hydrologic ensemble forecasting challenges that are particular to the orographically complex, snowmelt-driven basins of the Western US and British Columbia…prediction at monthly to seasonal lead times (i.e., 2 weeks t0 12 months).

• snow assimilation & model calibration

• basins:

• Mica (BC), Feather (CA), Klamath (OR/CA), Yakima (WA), Salmon (ID), Gunnison (CO), others?

END