Use of Multi-Model Super-Ensembles in Hydrology Lauren Hay George Leavesley Martyn Clark * Steven Markstrom Roland Viger U.S. Geological Survey Water Resources Discipline National Research Program * University of Colorado - Boulder
Hydrologic Simulation Inputs – Time series data Precipitation, Minimum + Maximum Temperature – Parameters (static information) Spatial characteristics Non-spatial characteristics Modeling Software
Sources of Error State of the system: observed != simulated Error in: – Inputs Time series data Parameters – Modeling Software
Optimization of Model Standard technique: – adjustment of parameters Spatial characteristics Non-spatial characteristics – “Fitting” simulated hydrograph to the observed hydrograph
Optimization of Model Standard technique: – adjustment of parameters Spatial characteristics Non-spatial characteristics – “Fitting” simulated hydrograph to the observed hydrograph Ignores numerous other sources of error!
Sources of Error Inputs – Time series data Weather Stations
Sources of Error Inputs – Time series data Weather Stations –Measurement inaccuracy –Measurement bias –Measurement drift
Sources of Error Inputs – Time series data Weather Stations –Measurement inaccuracy –Measurement bias –Measurement drift Global or Regional Climate Model inputs
Sources of Error Inputs – Time series data Weather Stations –Measurement inaccuracy –Measurement bias –Measurement drift Global or Regional Climate Model inputs –Model accuracy (timing, volume, extremes) –Spatial scale –Temporal scale
Sources of Error Inputs – Time series data Weather Stations –Measurement inaccuracy –Measurement bias –Measurement drift Global or Regional Climate Model inputs –Model accuracy (timing, volume, extremes) –Spatial scale –Temporal scale Representation & Distribution –Does this data describe what’s “hitting the ground”?
Sources of Error Inputs – Time series data – Parameters Spatial characteristics
Sources of Error Inputs – Time series data – Parameters Spatial characteristics –Quality of GIS layers –Quality of algorithms –Quality of GIS delineation techniques
Sources of Error Inputs – Time series data – Parameters Spatial characteristics –Quality of GIS layers (is my soil info accurate enough?) –Quality of algorithms (is my GIS using my soils data correctly?) –Quality of GIS delineation techniques (are my model’s geographic feature concepts appropriately represented in the GIS?)
Sources of Error Inputs – Time series data – Parameters Spatial characteristics Non-spatial characteristics
Sources of Error Inputs – Time series data – Parameters Spatial characteristics Non-spatial characteristics –adjustment factors for Time series data coefficients for measurement error & bias correction distribution of climate data to land surface units Modeling Response Units (MRUs)
Sources of Error Inputs Modeling Software
Sources of Error Inputs Modeling Software – Model concepts valid? – In setting of the application area? – Are selected processes successfully integrated?
Sources of Error Inputs Modeling Software Optimization technique – “fitting” the simulated hydrograph to the observed
Sources of Error Inputs Modeling Software Optimization technique – “fitting” the simulated hydrograph to the observed How is this measured? Is chosen statistic appropriate? Is a single statistic appropriate? Is this statistics appropriate for the entire cycle of hydrologic response?
Optimization of Model Standard technique: – adjustment of parameters Based on single statistic over entire period
Optimization of Model Standard technique: – adjustment of parameters Based on single statistic over entire period Seems incomplete!
Super-Ensemble Study Joint effort: – USGS – University of Colorado – Boulder Funded by: – NOAA – University of Colorado – USGS (barely)
Super-Ensemble Study: purpose Systematically evaluate alternative components for hydrologic modeling Develop optimized modeling configurations Produce map-based database of configurations to support field staff
Super-Ensemble Study: approach Specify approximately 15 different model permutations Select 2 watersheds from each Hydrologic Landscape Unit Develop input climate time series data Automate delineation & parameterization of geographic features Automate Sensitivity & Optimization Analyses
Super-Ensemble Study: tools Modular Modeling System (MMS) Climate processing methods GIS Weasel MOGSA & MOCOM – Multi-object sensitivity and optimization tools – University of Arizona
Super Ensemble Study: MMS PROCEDURESPROCEDURES # Modules in MMS X X X Input Data Climate Processing Solar Radiation Potential Evapotranspiration Snow Soil Subsurface Groundwater X X X X X X X X
Super Ensemble Study: MMS
Climate Processing Methods Produces time series values for each MRU Basin Average Inverse Distance Nearest Neighbor Thiessen Polygons XYZ Local Polynomial Regression Artificial Neural Networks
Basin Selection 2 basins from each HLU approximately 70 for first iteration Each basin part of Hydrologic Climate Data Network (HCDN) Drainage area > 50 km 2 < 3000 km 2
Land surface form Climate geology Hydrologic Landscape Units (HLUs)
Basin Selection
GIS Weasel Simplifies the creation of spatial information for modeling Provides tools to: Delineate Parameterize relevant spatial features
GIS Weasel Still have to insert a nice plug for da weasel…
GIS Weasel: Example Delineation Methodology
“Uncalibrated” Watershed Model METHODOLOGY Basin Setup Optimize Volume Optimize Timing
Basin Setup Optimize Volume Optimize Timing “Uncalibrated” Watershed Model METHODOLOGY Basin Setup 1.Data set compilation (temperature, precipitation, DEM, Q) 2.Basin delineation 3.GIS Weasel 4.XYZ parameterization
Basin Setup Optimize Volume Optimize Timing METHODOLOGY Basin Setup “Uncalibrated” Watershed Model Identify and calibrate the ET parameters by comparing “observed” and simulated monthly mean PET out of hydrologic model Get a Water Balance Calibrate ET and climate station choice August Monthly Mean PE
“Uncalibrated” Watershed Model METHODOLOGY Basin Setup Optimize Volume Optimize Timing Get a Water Balance Calibrate ET and climate station choice Find ‘best’ climate station sets
METHODOLOGY Uncalibrated Watershed Model “Uncalibrated” Watershed Model METHODOLOGY Basin Setup Optimize Volume Optimize Timing Developed at U. of AZ: MOGSA MO MOGSA – Multi Objective G Generalized S Sensitivity A Analysis Determines parameter sensitivity Identify and optimize sensitive parameters
METHODOLOGY Uncalibrated Watershed Model “Uncalibrated” Watershed Model METHODOLOGY Basin Setup Optimize Volume Optimize Timing Developed at U. of AZ: MOGSA MO MOGSA – Multi Objective G Generalized S Sensitivity A Analysis Determines parameter sensitivity Developed at U. of AZ: MOCOM MO MOCOM – Multi-Objective COM COMplex Evolution Solves the multi-objective optimization problem Identify and optimize sensitive parameters
FD: Driven FQ: Quick FS: Slow FD FS FQ FS Multi-Objective Peak/Timing Quick recession Baseflow (See Boyle et al., WRR, 2000)
Anticipated Products Linking of physical processes – Atmospheric – Watershed – Two-way interaction (eventually) Development of Super-ensemble approach Physically-based watershed models that need limited interactive calibration
Anticipated Products Regionalization (spatial maps) of: Climate: –recommended sources variables –processing methods –parameters Recommendations for place-specific model selection/configuration Pareto sets of optimized parameters Confidence and error figures
Limitations Study deals with limited modeling question – Volume & timing of streamflow – Watershed scale ( km2) – Daily time step Limited number of physical process algorithms tested Limited number of watersheds featured – Automation will enable broader (nationwide) application
Timeline Dare we make these predictions?
Work Completed Climate processing – 4 of 7 methods implemented – Station observations selected for all test basins Records clean – Regional and Global Climate Model outputs assembled GIS – Delineation of geographic features automated – Parameterization of geographic features automated – Spatial data layers assembled – Processing complete
Work Completed Hydrologic science modules assembled MOGSA & MOCOM established
Contact Information Staff – George Leavesley (project chief) – Lauren Hay – Steve Markstrom – Roland Viger – Martyn Clark URLs – –
Thank You
Climate Processing Need to be able to distribute: From: stations grid points To: individual Modeling Response Unit (MRU)
Multiple Linear Regression (MLR) equations – Developed for: Precipitation Temperature, Maximum Temperature, Minimum – Based on: X Y Z – Monthly – Explains variation in observation across stations Climate Processing: XYZ overview Same relationship between stations and MRUs (use MRU X,Y,Z in MLR)
Climate Processing: Statistical Downscaling (SDS) overview Output from Global Climate Model (GCM) – National Center for Environmental Prediction (NCEP) model Averaged to a point (e.g. basin centroid) Distributed to MRUs – XYZ methodology
Climate Processing: Dynamical Downscaling (DDS) overview Uses Regional Climate Model – RegCM2 – Seeded with NCEP output Averaged to a point (e.g. basin centroid) Distributed to MRUs – XYZ methodology