How does the choice/configuration of hydrologic models affect the portrayal of climate change impacts? Pablo Mendoza 1.

How does the choice/configuration of hydrologicmodels affect the portrayal of climate change impacts?

Pablo Mendoza

1

Subjectivity in model selection:•How does the choice of model equations impact simulations of hydrologic processes?•Missing processes, inappropriate parameterizations?

Subjectivity in selecting/applying models

• Define a-priori values for model parameters

• Decide what model parameters we adjust, if any

• Decide what calibration strategy we implement, if any

Choice of objective functionChoice of forcing data and calibration period

Model parameters

• Decide which processes to include• Define parameterizations for individual

processes• Define how individual processes combine

to produce the system-scale response• Solve model equations

Model structure

Subjectivity in parameter identification:•How does our choice of model parameters impact simulations of hydrologic processes?•Compensatory effects of model parameters (right answers for the wrong reasons)?

Climate change studies commonly involve several methodological choices that might impact the hydrologic sensitivities obtained. In particular:

Study area

Basins of interest for this study

The Colorado Headwaters Region offers a major renewable water supply in the southwestern United States, with approximately 85 % of the streamflow coming from snowmelt. Hence, we conduct this research over three basins located in this area:

-Yampa at Steamboat Springs

-East at Almont

-Animas at Durango.

3

Approach

How do our methodological choices impact the results that we obtain when we evaluate hydrologic sensitivities to climate change?

Master question!

Impact of hydrologic model choice and parameters

Key points-Different hydrologic model structures.-Uncalibrated vs. calibrated model sensitivities.-Model structures vs. parameters.

Impact of spatial forcing resolution on hydrologic sensitivities

Key points-Use of dynamical downscaling outputs at different spatial scales, generated with the same methodology.-Impact of the spatial aggregation of a 4 km gridded dataset on hydrologic sensitivities.

4

Model structure selection

Differences in both model architecture and model parameterizations

5

Part I:Impact of hydrologic model choice

and parameters

6

Research plan: impact of model choice and parameters

Why do different models have different sensitivities to climate change?Is it due to differences in model structure rather than parameter values?

KEY QUESTIONS

Evaluation of uncalibrated model performance and sensitivities to climate change (done)

Model calibration(almost done)

Calculation of calibrated model sensitivities and comparison with uncalibrated(ongoing)

Assessment of differences on hydrologic sensitivities among feasible parameter sets (ongoing)

Task 1

Task 2

Task 3

Task 4

APPROACH 7

Status

8

Uncalibrated model simulations(WRF@4km resolution)

How does model performance change with model calibration?

9

Calibrated model simulations(WRF@4km resolution)

Calibration process substantially improves streamflow simulation.

Results Model performance

Impact of model structure on climate sensitivity

Impact of model parameters on climate sensitivity

How are observed hydrologic signature measures reproduced by different models?Raw values

10 The choice of a particular objective function (e.g. RMSE) for model calibration does

not necessarily improve simulated signature measures! (example: FMS)

Calibrated

Uncalibrated




How does hydrologic model choice affect the partitioning of precipitation into ET and runoff?

11



Uncalibrated models: Climate change signal in Noah (↑ET and ↑Runoff) differs from the rest of models (↑ET and ↓Runoff).

Inter-model differences are larger that climate change, even after calibration process.




Changes in signature measures

12

Uncalibrated

Calibrated

Inter-model differences in signature measure sensitivities don’t necessarily decrease after calibration! (e.g. seasonality and flashiness, especially at East and Yampa).




How does the impact of model parameters compare with that of model choice?

13

Optimal parameter set for NSE (raw space)

Optimal parameter set for objective function

Parameter sets selected

Approach (analysis restricted to VIC):Randomly select 2 points in the parameter space located in the area of maximum values of the objective function (ie. 8 parameter sets in total)

The optimal parameter set may change significantly with the choice of objective function.




14

Impact of parameters (VIC)Impact of model choice (after calibration)

Inter-parameters differences (VIC) have similar magnitudes than inter-model differences when we look at monthly runoff.




15

Impact of parameters (VIC)Impact of model choice (after calibration)

Uncertainty in monthly sensitivities of internal states and fluxes is still substantial, even when evaluating a limited set of model parameters.




16

Impact of parameters on the change in hydrologic signature measures (VIC)

Impact of model choice on the change in hydrologic signature measures

Model parameters: larger impact on changes in runoff ratio. Model choice: larger impact on changes in runoff seasonality and flashiness.




1. Calibration of hydrologic models improves streamflow simulation, but does not necessarily:

i. Improve representation of hydrological processes.ii. Decrease inter-model differences in signature measures change (PGW - CTRL).

2. Inter-model differences in hydrologic sensitivities to climate change:i. Are less pronounced for calibrated models rather than uncalibrated models.ii. May be larger than climate change signals even after calibration.

3. Regarding the role of parameters:i. Model choice (after calibration) and parameter selection from “optimal zones”

provide similar uncertainty in impact of climate change on monthly runoff.ii. Preliminary analysis suggests that uncertainty in monthly variations of specific

fluxes and states (e.g. ET, Soil moisture, SWE) is model-dependent rather than parameter-dependent.

Conclusions: Part I

17

Part II:Impact of spatial forcing resolution

on hydrologic sensitivities

18

Research plan: impact of forcing resolution

What is the impact of forcing spatial resolution on signature measures?How does forcing spatial resolution affect climate change impact results?

KEY QUESTIONS

Generate forcing datasets for all models:-Raw WRF @4km, 12km and 36km (done).-WRF-4 km data aggregated to 12km and 36 km (done).

Experiment 1: evaluate climate change impact using raw WRF outputs- Uncalibrated model simulations (done). - Calibrated model simulations (ongoing).

Task 1

Task 2

Task 3

APPROACH

Experiment 2: evaluate climate change impact using aggregated WRF-4km outputs.- Uncalibrated model simulations (done). - Calibrated model simulations (ongoing).

19

Status

20

7-year average cool-season precipitation : 1 October – 31 May

36 km 4 km OBSERVATIONS1000

900

800

700

600

500

400

300

200

100

0

Prec

ipit

atio

n (m

m)

12 km

Fig. Kyoko Ikeda

Results: previous studies

7-year average warm-season precipitation:1 June – 30 September

36 km 4 km OBSERVATIONS700

600

500

400

300

200

100

0

Prec

ipit

atio

n (m

m)

SNOTELGHCN

12 km

Fig. Kyoko Ikeda

Results: previous studies

Results

23

How does forcing resolution affect signature measures of hydrologic behavior?RR: Runoff Ratio

Impact of forcing is model dependent (Noah is much more sensitive) Impact of forcing resolution on runoff ratio is also basin dependent.

Calibrated

Impact of forcing resolution on signature measures (historical)

Impact of forcing resolution on climate sensitivity

Y: Yampa River Basin; E: East River Basin; A: Animas River BasinUncalibrated

Raw WRF outputs WRF-4km aggregated

24

How does forcing resolution affect signature measures of hydrologic behavior?CTR: Runoff Seasonality

Calibrated

36km resolution datasets tend to produce earlier runoff (less clear for 12km). Aggregation reduces resolution differences. Results depend on basin/model.


Impact of forcing resolution on climate sensitivityResults

UncalibratedY: Yampa River Basin; E: East River Basin; A: Animas River Basin


Experiment 1: raw WRF output

25 Raw WRF outputs at 12km and 36km change direction of signal (↓Runoff and

↑ET), even after model calibration, to ↑Runoff and ↑ET.



Experiment 2: aggregated WRF output

26 Aggregated WRF-4km outputs at 12km and 36km don’t change signals

significantly, but may affect the amplitude (e.g. calibrated Noah).



27

How will signature measures change across models and forcing resolutions?RR: Runoff Ratio

RR clearly depends on forcing resolution! General decrease of RR in all cases.

Y: Yampa River BasinE: East River BasinA: Animas River Basin

Calibrated



Uncalibrated


28

How will signature measures change across models and forcing resolutions?CTR: Runoff Seasonality

Shift to earlier runoff in all cases (answer does not depend on forcing)


Calibrated



Uncalibrated


1. The impact of forcing resolutions on signature measures for historical simulations is reduced when we use spatially aggregated WRF-4km outputs. This implies that physics options in each WRF configuration (4km, 12km and 36 km) dominates hydrological responses.

2. Regarding climate change signal:• Raw WRF outputs at 12km and 36km change direction of signal, even after model

calibration, to ↑Runoff and ↑ET.• Aggregated WRF-4km outputs at 12km and 36km don’t change signals

significantly, but may affect the amplitude (e.g. calibrated Noah).

3. Under a future climate scenario, earlier runoff volumes and a general decrease in runoff ratios is obtained with all forcing datasets. However, results are still model dependent.

Conclusions: Part II

29

Thank you

EXTRA

31

• Useful information: SIGNATURE MEASURES!!!

• What do they represent?i. RR: overall water balance.ii. FMS: vertical redistribution of soil moisture.iii. FHV: watershed response to large precipitation events.iv. FLV: Long term baseflow.v. FMM: Mid-range flow levels.vi. CTR: runoff seasonality.

Casper et al. (2012)

EXAMPLES

RR: Runoff Ratio (Q/P)FMS: Slope of mid-segment in FDC (0.2 < Pexc < 0.7)FHV: High segment volume in FDC (0 < Pexc < 0.02)FLV: Low segment volume in FDC (0.7 < Pexc < 1)FMM: Median value of simulated streamflowCTR: Centroid of avg. water year daily hydrograph (days since Oct 1)

Approach: diagnostic signatures

32

Results: impact of model choice and parameters


How does model performance change with model calibration?

33Calibrated model simulations(WRF@4km resolution)


34

How are observed hydrologic signature measures reproduced by different models?Raw valuesUncalibrated

Calibrated


How are observed hydrologic signature measures reproduced by different models?CTRL - Observed

35

Uncalibrated

Calibrated


36

How are observed hydrologic signature measures reproduced by different models?CTRL - ObservedUncalibrated

Calibrated

Monthly total runoff values for different basins/models


37



Monthly differences (PGW-CTRL) in mm for specific fluxes/states


38Uncalibrated model simulations(WRF@4km resolution)


Current (CTRL) and Future (PGW) signature measures


39

Uncalibrated

Calibrated

Current (CTRL) and Future (PGW) signature measures


40

Changes in signature measures


41

Changes in signature measures (PGW vs. CTRL runs)


42

Changes in signature measures (PGW vs. CTRL runs)


43

Results: impact of model choice and parametersWhat is the impact of the objective function on the optimal parameter set?

44



Kling-Gupta Efficiency (KGE)

Nash-Sutcliffe Efficiency (NSE)


So… how will our hydrologic sensitivities change if we arbitrarily select parameter sets within the optimal region?

45



Parameter sets selected

Approach:Randomly select 2 points in the parameter space located in the area of maximum values of the objective function (ie. 8 parameter sets in total)


46

Model performance for CTRL simulations (Sep/2002 – Oct/2008)


47

Model performance for CTRL simulations (Sep/2002 – Oct/2008)East River Basin


48

How are observed hydrologic signature measures reproduced by different parameter sets?Raw values


49

How are observed hydrologic signature measures reproduced by different parameter sets?CTRL - Observations


50

Impact of parameters on partitioning of precipitation into ET and Runoff


51

How will hydrologic signature measures change in a future climate?PGW vs. CTRL values for 8 different parameter sets (VIC)


52

Impact of parameters on the change in hydrologic signature measures (VIC)

Impact of model choice on the change in hydrologic signature measures


53

How will hydrologic signature measures change in a future climate?Current and future raw values


54

How will hydrologic signature measures change in a future climate?Future - Current

Results: impact of model choiceHow are observed hydrologic signature measures reproduced by different

models/datasets?Raw values

55Experiment 1: raw WRF output & uncalibrated models

Results: impact of model choice

56Experiment 1: raw WRF output & uncalibrated models

How are observed hydrologic signature measures reproduced by different models/datasets?

Raw values

Results: impact of model choiceHow are observed hydrologic signature measures reproduced by different

models/datasets?Raw values

57Experiment 2: aggregated WRF output & uncalibrated models


58Experiment 2: aggregated WRF output & uncalibrated models

How are observed hydrologic signature measures reproduced by different models/datasets?

Raw values


59

How does forcing resolution affect signature measures of hydrologic behavior?FMS: Flashiness of runoff

Uncalibrated models Calibrated modelsY: Yampa River BasinE: East River BasinA: Animas River Basin


60

How does forcing resolution affect signature measures of hydrologic behavior?FHV: Response to large precipitation events



61

How does forcing resolution affect signature measures of hydrologic behavior?FLV: Long-term baseflow



62

How does forcing resolution affect signature measures of hydrologic behavior?FMM: Mid-range flow levels



63

How will signature measures change across models and forcing resolutions?FMS: Flashiness of runoff


Uncalibrated Calibrated


64

How will signature measures change across models and forcing resolutions?FHV: Response to large precipitation events




65

How will signature measures change across models and forcing resolutions?FLV: Long-term baseflow




66

How will signature measures change across models and forcing resolutions?FMM: Mid-range flow levels



Results: impact of forcing spatial resolution

Experiment 1: raw WRF output67

Total runoff (uncalibrated) Total runoff (calibrated)


Experiment 2: aggregated WRF output68

Total runoff (uncalibrated) Total runoff (calibrated)



Evapotranspiration (uncalibrated) Evapotranspiration (calibrated)


70

Evapotranspiration (uncalibrated) Evapotranspiration (calibrated)




SWE (uncalibrated) SWE (calibrated)


72

SWE (uncalibrated) SWE (calibrated)




Soil moisture (uncalibrated) Soil moisture (calibrated)


74

Soil moisture (uncalibrated) Soil moisture (calibrated)


NSE surfaces: PRMS

Nsim = 10,000 (100 x 100 points)

East River Basin

75

NSE surfaces: PRMS

Nsim = 2,500 (50 x 50 points)

East River Basin

76

NSE surfaces: PRMS

Nsim = 2,500 (50 x 50 points)

East River Basin

77

NSE surfaces: PRMS

Nsim = 2,500 (50 x 50 points)

East River Basin

78

NSE surfaces: PRMS

Nsim = 2,500 (50 x 50 points)

East River Basin

79

NSE surfaces: PRMS

Nsim = 2,500 (50 x 50 points)

East River Basin

80

NSE surfaces: PRMS

Nsim = 2,500 (50 x 50 points)

East River Basin

81

NSE surfaces: PRMS

Nsim = 2,500 (50 x 50 points)

East River Basin

82

NSE surfaces: PRMS

Nsim = 2,500 (50 x 50 points)

East River Basin

83

NSE surfaces: VIC

Nsim = 10,000 (100 x 100 points)

East River BasinIs this related to spatial parameterization or to the model?No, because thick2 is spatially constant in the basin

84

NSE surfaces: VIC

Default

Calibration strategy: One multiplier for each parameter (binfilt, Ds, Dsmax, Ws, depth2, depth3)

Before discontinuity

After discontinuity

The conflictive parameter is spatially constant!!

85

Results: impact of model choiceWhat is the impact of the objective function on the optimal parameter set?

86

NSE surfaces: VIC

Nsim = 2,500 (50 x 50 points)

East River Basin

87

NSE surfaces: VIC

Nsim = 2,500 (50 x 50 points)

East River Basin

88

EXTRA: the VIC experiment

89

How does the choice/configuration of hydrologic models affect the portrayal of climate change impacts? Pablo Mendoza 1.

Documents

model calibration

model parameterizations

model differences

model architecture

model selection

impact of model choice

calibrated model sensitivities

climate sensitivityhow