using ground-water model predictions and the ppr and opr statistics to guide data collection
TRANSCRIPT
Using Ground-Water Model
Predictions and the ppr and opr
Statistics to Guide Data Collection
Motivation
Parameter Distribution
Parameter
UncertaintyPrediction Uncertainty
Hydrogeologic Data
Calibrated Model & Predictions
Incomplete Data
• Ground-water model predictions are always uncertain.
• What hydrogeologic data could be collected to reduce this prediction uncertainty most effectively?
Motivation
Parameter Distribution
Reduced Parameter
Uncertainty Reduced Prediction Uncertainty
Hydrogeologic Data
Calibrated Model & Predictions
Additional Data
Approach
Use calibrated model to identify parameters important to predictions.
Parameter Distribution
Calibrated Model & Predictions
Approach
Collect hydrogeologic data: Parameter values Flow system characteristics
Use calibrated model to identify parameters important to predictions.
Hydrogeologic Data
Parameter Distribution
Calibrated Model & Predictions
Approach
Collect hydrogeologic data: Parameter values Flow system characteristics
Incorporate these data into the model to reduce parameter and prediction uncertainty.
Use calibrated model to identify parameters important to predictions.
Parameter-Prediction Statistic (PPR)
1. Calculate prediction uncertainty (sZ) using the calibrated model.
2. Assume improved information on one or more parameters, and recalculate sZ.
3. PPR statistic equals the percent decrease in sZ from step 1 to step 2.
Prediction
Uncertainty(standarddeviationssZ)
ParameterUncertainty
PredictionSensitivities
ObservationSensitivities
DVRFS Model Parameters
9 Hydraulic Conductivities
4 Recharge Parameters
Predictions: Advective-Transport Paths
• Advective transport used as a surrogate for regional contaminant transport.
• Advective transport paths are 10 km.
• Predictions are the distances traveled in the N-S, E-W, and vertical directions.
Black bars: Prediction standard deviations calculated using calibrated model.
Uncertainty in Path Position
Advectivepath
Black bars: Prediction standard deviations calculated using calibrated model.
Red bars: Prediction standard deviations calculated with improved information on a parameter.
Uncertainty in Path Position
Advectivepath
PPR Statistic:Individual Parameters
• Specify improved information on one parameter, so that its uncertainty decreases by 10 percent.
• Calculate resulting decrease in prediction standard deviations sZ.
• Repeat for all model parameters.
Parameter with Improved Information
HydraulicConductivity Recharge
PPR: Individual Parameters
East-WestR4K2
K1K3
0
5VOII (percentdecrease
in sZ)
PPR
East-WestR4K2
K1K3
0
5VOII (percentdecrease
in sZ)
Vertical R1
0
6
12
North-SouthR4K2
K1
K3
0
5
Parameter with Improved Information
HydraulicConductivity Recharge
PPR: Individual Parameters
PPR
PPR:Multiple Parameters
• Specify improved information on three parameters.
• Calculate PPR statistic (decrease in prediction standard deviation sZ).
• Repeat for all possible sets of three model parameters.
PPR: Multiple Parameters
East-West
K5
K1,K3,K5
R4K2K1 K3
0
10
Percentdecrease
in sZ
VOII on Individual Parameters VOII on 3 Parameters
PPR PPR
VOII: Multiple Parameters
East-West
K5
K1,K3,K5
R4K2K1 K3
0
10
Percentdecrease
in sZ
VOII on Individual Parameters VOII on 3 Parameters
Vertical
K4
R1R1,K4,A3
0
13
PPRPPR
Advective Path from Yucca Flat Site
K zones, layer 1 Recharge zones
K1
K5
R1
10 km
R4K3
layer 2
UsingthePPRResults
Collect hydrogeologic data related to important
parameters
System State Observations
Improved Predictions, Reduced Uncertainty
Societal decisions
Improve Model & Parameters
Recalibrate Model
Observation-Prediction (opr) Statistic
1. Calculate prediction uncertainty (sZ) using the calibrated model and all 517 observations.
2. Add or omit one or more observations, and recalculate sZ.
3. opr statistic equals the change in sZ from step 1 to step 2.
Prediction
Uncertainty(standarddeviationssZ)
ParameterUncertainty
PredictionSensitivities
ObservationSensitivities
Predictionsevaluatedforassessingobservations
Hill and Tiedeman, 2007, fig. 15.7. p. 366
Whichexistingobservationsareimportant(ornot)topredictions?
Useopr(-1)torankthe501existingobservationlocationsbytheirimportancetopredictions
• Averagedvaluesofopr(-1)forallthepredictionsareused,toobtainameasureindicatingtheimportanceofasingleobservationtoallthepredictionsofinterest.
• Calculateopr(-100) byremovingthe100leastimportantobservations
• opr(-100) =meanpredictionuncertaintyincrease=0.6%
Hill and Tiedeman, 2007, fig. 15.9. p. 368
Consideronepotentialnewheadobservationineachcellofmodellayer1.
Determineweightsforthepotentialobservations.
Here,sameweightingstrategyusedasforweightingexistingobservations–weightssmallerforheadsinhigh-gradientareas.
Calculateopr(+1)foreachcellinthelayer,eventhosewithanexistingobservation,sothatopr(+1)iscontinuousoverthewholemap.
Whatnewobservationswouldbeimportant(ornot)topredictions?
Hill and Tiedeman, 2007, fig. 15.10. p. 369
Improve Model & Parameters
Recalibrate Model
Hydrologic and Hydrogeologic Data
Collect additional observation data
Improved Predictions, Reduced Uncertainty
Societal decisions
UsingtheOPRresultsforpotentialnewobservations
Summary
• Parameters and observations most important to the predicted advective-transport paths do not necessarily lie near the paths themselves.
• Best to not use ppr and opr results alone for making decisions about future data collection – consider other criteria such as geologic insight about important subsurface units to investigate, maintaining good geographic and depth coverage of monitoring network, etc.
• The ppr and opr results are only as good as the model itself!