Introduction

1

Introduction

G. Thirel and V. Andréassian

IAHS Hw15 22 July 2013

5

Modelling is like painting

Catchments are hyper-dynamic systems: they change continuouslyFor the sake of our ‘portrating’, we need to make simplifying assumptionsThe risk: that the simplifying hypotheses cause a catchment non-stationarity artefact

6

Non stationarity makes the life of hydrologists miserable

Identifying parameters is already not an easy task within the stationarity hypothesis…

… it is much worse when changes which we have neglected turn out to have a significant impact on the calibration process

7

Goal of this workshop

To provide a factual diagnosis: describe / document the problembased on common catchments also on other datasetsCan we agree on the problem? On how to assess it, numerically and graphically? Investigate solutions

8

Workshop preparation group

Guillaume ThirelValérie Borrell-EstupinaSandra Ardoin-BardinJulien LeratOlga SemenovaFrancesco Laio

9

Outline of this presentation

• The dataset

• The calibration and evaluation protocol

• Some results


10


The dataset

The calibration and evaluation protocol

Some results


11

Need for a dedicated websiteFOR DEFINING THE COMMON FRAMEWORK AND FOR PROVIDING THE COMMON DATABASE

Website address: http://non-stationarities.irstea.fr/

• Description of the dataset for each basin• Description of the calibration and evaluation protocol

Possibility to download the data (password protected)


http://non-stationarities.irstea.fr/

12

The dataset14 RIVER BASINS SHOWING NON-STATIONARITIES


13

The dataset14 RIVER BASINS SHOWING NON-STATIONARITIES

Basins sizes from 0.2 km² to 100,000km²

Several types of non-stationarities encountered: • Temperature increase• Precipitation change or high variability• Urbanization• Forest cover modification

Period: variable according to the considered basin


14

Which data?DATA COLLECTED FROM MANY PARTNERS

Variables: precipitation (P), temperature (T), potential evapotranspiration (PE), discharge (Q).

What we provided: • basin-wide aggregated values of P, T and PE • Q at the outlet • (repartition of altitude within the basin if available)

Time step: daily


15

Temperature increaseTHE KAMP (622 KM²), ALLIER (2267 KM²), DURANCE (2170 KM²) AND GARONNE (9980 KM²) RIVERS

All located in Europe, impacted by snowmelt

Allier Kamp

Durance Garonne


16

The case of the Kamp RiverVERY LARGE FLOODS IN 2002

P

QIAHS Hw15

22 July 2013

(Komma et al., 2007; Blöschl et al., 2008; Reszler et al., 2008)

17

The case of the Allier RiverCONSTRUCTION OF A DAM IN 1983 FOR SUSTAINING LOW FLOWS


Impact on low flows

18

Precipitation change or high variabilityTHE AXE CREEK (237 KM²) AND THE WIMMERA RIVER (2000 KM²)

Millenium drought in Australia (1997-2008)

Wimmera River

Axe Creek


Q

Q

19

Decrease in rainfall and deep water recharge between before 1970 and after 1971

Precipitation change or high variabilityTHE BANI RIVER (100,000 KM²)

P Q


20

Precipitation change or high variabilityTHE GILBERT AND FLINDERS RIVERS (AROUND 1900 KM²)

Arid catchments under cyclonic heavy rainfall influence.

Major flood in 2002.

The Flinders RiverIAHS Hw15

22 July 2013

21

UrbanizationTHE FERSON (134 KM²) AND BLACKBERRY CREEKS (182 KM²)

Located in the USAThe urbanization modifies the hydrological response

1980

1983

1986

1989

1992

1995

1998

2001

2004

2007

2010

0

10

20

30

40

50

60

70

Ferson CreekBlackberry Creek

Percentage of urbanization


22

Forest cover modificationTHE FERNOW (0,2 KM²) AND MÖRRUMSÅN (97 KM²) RIVERS AND THE REAL COLLOBRIER (1,4 KM²)

The Fernow Experimental watershed: forest cut of the lower part of the basin, then forest cut of the upper part of the basin, then plantation of firtrees.

The Mörrumsån River: a severe storm (Gudrun), led to loss of forest in January 2005.

The Real Collobrier: forest fire in August 1990.


23

The dataset

River Country Area Period Change ProvidersFernow River USA 0.2km² 1956-2009 Forest USDA Forest Service

Real Collobrier France 1.4km² 1966-2006 Forest Irstea & Météo-France

Mörrumsan River Sweden 97km² 1981-2010 Forest SMHI

Ferson Creek USA 134km² 1980-2011 Urbanization USGS & DayMet

Blackberry Creek USA 182km² 1980-2011 Urbanization

Axe Creek Australia 237km² 1970-2011 P decrease Victoria data Warehouse

Kamp River Austria 622km² 1976-2008 T increase TU Wien, UFZ

Gilbert River Australia 1907km² 1963-1988 P variability Queensland GovernmentFlinders River Australia 1912km² 1967-2011 P variability

Wimmera River Australia 2000km² 1960-2009 P decrease Victoria data Warehouse

Durance River France 2170km² 1901-2010 T increase EDF

Allier River France 2267km² 1958-2008 T increase Météo-France & Banque HydroGaronne River France 9980km² 1958-2008 T increase

Bani River W Africa 103390km² 1959-1990 P decrease DMM & DNH

24


• The dataset


• Some results


25

The protocolA COMMON CALIBRATION AND EVALUATION FRAMEWORK

• Common calibration / evaluation periods

• Common minimum set of metrics

• Possibility that I produce a set of metrics and plots for the modellers (providing that they sent to me their simulations)

• Modellers are free to do more!


Complete periodWarm-up

TimeP1 P4P3P2 P5

Change

26

The protocolLEVEL 1: THE BEGINNER LEVEL

• Calibration has to be done on the “Complete period” or the model does not need calibration

• Models are run on the “Complete period”

• Evaluation is done on the “Complete period” + P1 to P5



TimeP1 P4P3P2 P5

Change

27

The protocolLEVEL 2: THE NORMAL LEVEL

• Calibration has to be done on each pre-defined sub-period P1 to P5

• Models are run on the “Complete period” for each calibration

• Evaluation is done on the “Complete period” + P1 to P5



TimeP1 P4P3P2 P5

Change

28

The protocolLEVEL 3: THE EXPERT LEVEL

The modellers found that their model failed at level 2 to deal with non-stationarities or could do better.

They want to try to solve this issue, or at least to try to test solutions that could solve this issue.

-> all solutions are allowed. Failing is fine, since it allows to discard a solution.


29

The protocolTHE METRICS

Participants were asked to produce the following statistics on each sub-period: • NSE and NSE on low flows (i.e. using 1/Q+ε instead of Q) • Bias (Qsim/Qobs)• Discharge quantiles: Q95, Q85, Q15 and Q05• Frequency of low flows (i.e. when Q<5% of mean Qobs)


30

The protocolTHE METRICS


Participants were asked to produce the following statistics on each sub-period: • NSE and NSE on low flows (i.e. using 1/Q+ε instead of Q) + their

decomposition• Bias (Qsim/Qobs)• Discharge quantiles: Q95, Q85, Q15 and Q05• Frequency of low flows (i.e. when Q<5% of mean Qobs)• KGE and its decomposition• Nash and bias on sliding windows• Flow regime• Ranked discharges

31

The protocolTHE GRAPHS

Used for the bias, the Nash criteria, the KGE, and their decompositions

For the quantiles and the frequency of low flows, the observed value is added

Two different ways of showing the same thingIAHS Hw15 22 July 2013

The criterion value

Six curves: one for each calibration

Six values: one for each evaluation period

The criterion value

Six columns: one for each calibration

Six lines: one for each evaluation period

32


Extension of some graphs for a 1-year frequency evaluation


33


The discharges regimes and the ranked discharges-> one graph for each evaluation period


34

The protocolTHE COMPARISONS BETWEEN MODELS


Not values, but differences

between the criteria of two

models

Blue values indicate that this model has a higher criterion

Red values indicate that this model has a higher criterion

Blue values indicate that this model has a higher criterion

Red values indicate that this model has a higher criterion

A column compares a single calibration on each evaluation period

A line compares each calibration on a single evaluation period

Mod 1 Mod 2

Mod 1

Mod 2

Over-estimation from model on top

Δ

Under-estimation from the model on top

35


• The dataset


• Some results


36

List of models used for the workshop

1k-DHM, AWBM, CLSM, COSERO, ECOMAG, GARDENIA, GR4J, GR5J, HBV, HYDROGEOIS, HYPE, HyMod, IHACRES, MISO, MORDOR, MORDOR6, SAFRAN-ISBA-MODCOU, SimHyd, SpringSim, TOPMODEL, Xinanjiang,…


37

Brief presentation of some results from people who participated but could not come

GARDENIA: D. Thiéry, BRGM, France

COSERO: H. Kling, Austria

SpringSim: A. Ramchurn, Australia


38

Lumped model with slow compo-nents reservoirs

Can take into account aquifer level measurements (not used here)

4 to 6 parameters

Calibration metrics: MSE(sqrt(Q))+5%(Qsim-Qobs)

Ran on 11 basins

Ref: Thiéry, D. (2010) Reservoir Models in Hydrogeology, in Mathematical Models, Volume 2 (ed J.-M. Tanguy), John Wiley & Sons, Inc., Hoboken, NJ, USA. doi: 10.1002/9781118557853.ch13

GARDENIAUSED BY: Dominique Thiery ([email protected], BRGM, France)


Soil storage

Vadose zone

Groundwater tank

Effective rainfall

Percolation (Recharge) Rapid flow

Grounwater flow

Rain Snow Evapotranspiration

(PET)

Aquifer level

mailto:[email protected]

39

Low bias on calibration periods, but high bias on contrasted periods.

The calibration on the complete period allows an « intermediary » solution but does not prevent from biased simulations.

Calibrations on dryer periods gave higher soil reservoir capacities: the model tries to allow more evapotranspiration for compensating the lower Q.



Bani

Regime change from

1971


40

High bias on contrasted periods.Calibrations on dryer periods gave higher soil reservoir capacities: try of the model to allow more evapotranspiration for compensating the lower Q.



Wimmera

MilleniumDrought


41

COSERO

Continuous, semi-distributed rainfall-runoff model.

• Snow processes• Soil moisture accounting (HBV-type)• Surface-flow, inter-flow, base-flow

(linear reservoirs)

Nachtnebel et al. (1993)


USED BY: Harald Kling ([email protected], Pöyry Energy GmbH, Austria)


42

COSERO

Objective function: KGE on Q.

Ran on 11 basins (i.e. all except US basins).

Dam module (affects low flows) added for the Allier River.Riparian zone (affects evaporation) added for Australian rivers.




43

COSERO

Allier



New dam built in 1983


44

COSEROUSED BY: Harald Kling ([email protected], Pöyry Energy GmbH, Austria)

Wimmera

COSERO GR4JSimilar behaviour: clear over-estimation for the Millenium DroughtDifference: no clear under-estimation of wet years for Cosero when calibrated on dry years


MilleniumDrought


45

SpringSIM

New model, implemented to deal specifically with incorporation of long term droughts in the routine response to rainfall/evaporation of rainfall-runoff models12 parameters


USED BY: Avijeet Ramchurn ([email protected], Bureau of Meteorology, Australia)

Interception store

Quick response layer

Uns

atur

ated

zone

Satu

rate

dzo

ne

Soil

profi

le

dept

h (D

)

Saturation flow

Quickflow

Interflow

Subsurface loss from saturated zone

Subsurface loss from unsaturated zone

Rainfall Evaporation

Outletheight (OH)

Spill

Leakage to saturated zone

Unavailableto quickflow

GWD

UNZD

Channel routing

Impervious area flow

Rainfall


46

SpringSIM



0

100

200

300

400

500

600

700

800

900

1000So

il M

oist

ure

Stor

e

Soil moisture content of unsaturated zone

Saturated zone

Outlet level


47

SpringSIM

Good simulations of the water volume



Bani


48

SpringSIM

Low bias in contrasted periods



Bani

Wet Dry

Wet

Dry


49

Thank you!

50

The protocolTHE COMPARISONS BETWEEN MODELS

Comparisons between the models

Comparisons with observations

Mod 1 Mod 2M

od 1M

od 2

Over-estimation from model on top

Δ

Under-estimation from the model on top

51


The 10-year sliding windows plots


Introduction

Documents

Transcript of Introduction