G. Thirel, L. Coron, V. Andréassian, C. Perrin

www.irstea.fr

Pour mieux affirmer ses missions, le Cemagref devient Irstea

G. Thirel, L. Coron, V. Andréassian, C. Perrin

22 July 2013

Application of several hydrological models (and objective functions) to the complete dataset of the workshop

2

Introduction

IAHS Hw15 22 July 2013

• What is the main issue when we fail on non-stationarity?

• Models?• Objective functions?• Something else?

• Application of 3 models• Application of 6 objective functions

3

Outline of this presentation

• Impact of using different models

• The models

• The results

• Impact of using different objective functions

• The objective functions

• The resultsIAHS Hw15

22 July 2013

4

Outline of this presentation

• Impact of using different models

• The models

• The results

• Impact of using different objective functions

• The objective functions

• The resultsIAHS Hw15

22 July 2013

5

GR4J and GR5J

Lumped conceptual models, resp. 4 and 5 parameters

GR5JGR4J

IAHS Hw15 22 July 2013

6

MORDOR6

Lumped conceptual model with 6 parameters (simplification of the MORDOR model).

IAHS Hw15 22 July 2013

7

The snow module

• No snow module: Axe Creek, Gilbert, Flinders, Wimmera and Bani Rivers.• CemaNeige: all the other basins.

CemaNeige = degree-day model, 2 free parameters.

IAHS Hw15 22 July 2013

8

The objective function

The Nash on root square of discharge is used in this part.

IAHS Hw15 22 July 2013

9

Outline of this presentation

• Impact of using different models

• The models

• The results

• Impact of using different objective functions

• The objective functions

• The resultsIAHS Hw15

22 July 2013

10

Rivers with T increase

IAHS Hw15 22 July 2013

• GR4J and GR5J are the best for the Kamp, except during P2

• No big difference for the Garonne

High flows

11

MORDOR6 misses the 2002

Kamp flood

Rivers with T increase

Observed peak value

MORDOR6 peak values

IAHS Hw15 22 July 2013

12

Rivers with T increase

IAHS Hw15 22 July 2013

• GR5J the best for the Kamp• MORDOR6 and GR5J the best for the Garonne

Low flows

13

Rivers with T increase

The model choice and calibration induce the same order of variability

VARIABILITY DUE TO MODEL AND CALIBRATION CHOICES

IAHS Hw15 22 July 2013

GR4J GR5J

MORDOR6Kamp

14

Rivers with discharge change or high variability

Wimmera Best performance for

GR5J

IAHS Hw15 22 July 2013

High flows

15

Rivers with discharge change or high variability

Wimmera No model has the « solution » for handling the

Millenium Drought

IAHS Hw15 22 July 2013

16

Rivers with T increase

The model choice and calibration induce the same order of variability

VARIABILITY DUE TO MODEL AND CALIBRATION CHOICES

IAHS Hw15 22 July 2013

GR4J GR5J

MORDOR6Wimmera

17

Rivers with discharge change or high variability

Bani

GR4J

Severe crash from GR4J due to high reactivity

Attempts to increase the reaction time or to better initialize the parameters all failed

The structure of GR4J (&GR5J) is to revise for such a basin

IAHS Hw15 22 July 2013 MORDOR6

18

Outline of this presentation

• Impact of using different models

• The models

• The results

• Impact of using different objective functions

• The objective functions

• The resultsIAHS Hw15

22 July 2013

19

The objective functionsFOR THIS PART ONLY THE GR4J MODEL IS USED

IAHS Hw15 22 July 2013

Inverse of discharge

Square root of discharge

Discharge

Nash NaIQ NaRQ NaQ

KGE KGEIQ KGERQ KGEQ

20

Outline of this presentation

• Impact of using different models

• The models

• The results

• Impact of using different objective functions

• The objective functions

• The resultsIAHS Hw15

22 July 2013

21

Rivers with T increase

Calibrating on IQ gives the lowest Nash(Q) -> of course!

IAHS Hw15 22 July 2013

22

Rivers with T increase

Calibrating on Q gives the lowest Nash(IQ) -> of course!

IAHS Hw15 22 July 2013

23

Rivers with P decrease

IAHS Hw15 22 July 2013

Kamp

NaQ NaRQ NaIQ

KGEQ KGERQ KGEIQ

VARIABILITY DUE TO CALIBRATION AND OBJECTIVE FUNCTION CHOICES

The objective functions impact the model bias more than the model choice

24

Rivers with discharge change or high variability

Wimmera Only NaRQ does not show disastrous results on P5 when calibrated on wet period.

KGERQ performs the best on wet periods when calibrated on P5. IAHS Hw15

22 July 2013

Prod. Store: NaRQ > KGERQLoss for P5: KGERQ > NaRQ

25

Rivers with P decrease

IAHS Hw15 22 July 2013

Wimmera

NaQ NaRQ NaIQ

KGEQ KGERQ KGEIQ

VARIABILITY DUE TO CALIBRATION AND OBJECTIVE FUNCTION CHOICES

The objective functions strongly impact the model bias

26

Conclusions

IAHS Hw15 22 July 2013

Attempts to quantify the (un-)stability induced by : -The model choice –> low impact

-The calibration period -> low impact on variability, high impact on bias

-The objective function -> huge impact

27

Thank you!