Objective Function Definition in Conceptual Hydrological Modelling

Objective Function Definition
in Conceptual Hydrological Modelling
V. Guinot, B. Cappelaere, C. Delenne, D. Ruelland
HydroSciences Montpellier UMR 5569 (CNRS, IRD, UM1, UM2)
SimHydro 2010
1
Objective Function Definition in Conceptual Hydrological Modelling
Guinot, Cappelaere, Delenne & Ruelland
Reservoir‐based models
• discretized versions of balance equations written in differential form
dU
= R (U , ϕ , t ) − Q (U , ϕ , t )
dt
R
U n +1 = f (U n , Rn , ϕ ,...)
Qn = g (U n , ϕ )
• Predefined flux functions
U
Q
• Parameters to be calibrated
R (U , ϕ , t )
SimHydro 2010
Q (U , ϕ , t )
2
Objective Function Definition in Conceptual Hydrological Modelling
Guinot, Cappelaere, Delenne & Ruelland
The calibration process
R
• The primary variable can often
not be used for calibration
U
• A « secondary variable » is used
instead (e.g. the discharge Q) because it is easier to observe
Q
Q
Observed
Modelled
t
SimHydro 2010
3
Objective Function Definition in Conceptual Hydrological Modelling
Guinot, Cappelaere, Delenne & Ruelland
The calibration process
V
• Try to make the modelled signal as close as possible to the observed one BUT…
• What does « as close as possible » mean ? F(U, ϕ)
Observed
Modelled
t
Q
t
SimHydro 2010
4
Objective Function Definition in Conceptual Hydrological Modelling
Guinot, Cappelaere, Delenne & Ruelland
Classical measures of « closeness »
• Nash‐Sutcliffe Efficiency (NSE)
• Mean Square Error (MSE)
• Root Mean Square Error (RMSE)
J=
∑ [F (U
N
J=
n , ϕ ) − Vn
]
2
=
∑
V
en2
e : model error
e
Ω
t
N
2
2
[
]
F
(
U
(
t
),
ϕ
)
−
V
(
y
)
d
t
=
e
∫
∫ (t ) dt
Ω
SimHydro 2010
Ω
5
Objective Function Definition in Conceptual Hydrological Modelling
Guinot, Cappelaere, Delenne & Ruelland
Classical measures of « closeness »
V
• Generalized Nash‐Sutcliffe
Efficiency (NSE)
e : model error
e
Jp =
∑
en
∑
F (U (t ), ϕ ) − V ( y ) dt =
∫
F (U n , ϕ ) − Vn
p
=
∫
Ω
SimHydro 2010
Ω
N
N
Jp =
p
p
t
p
e (t ) dt
Ω
6
Objective Function Definition in Conceptual Hydrological Modelling
Guinot, Cappelaere, Delenne & Ruelland
Classical measures of « closeness »
V
• Generalized Nash‐Sutcliffe
Efficiency (NSE)
e : model error
e
Jp =
∑
en
∑
F (U (t ), ϕ ) − V ( y ) dt =
∫
F (U n , ϕ ) − Vn
p
=
∫
p
Ω
Ω
N
N
Jp =
p
t
p
e (t ) dt
Ω
Perfect model
SimHydro 2010
Jp = 0
7
Objective Function Definition in Conceptual Hydrological Modelling
Guinot, Cappelaere, Delenne & Ruelland
Classical measures of « closeness » verify
• Positivity
Jp =
∫
e(t )
p
dt ≥ 0
∀( p > 0, Ω)
Ω
• Triangle rule
• Zero property
J p = 0 ⇔ e(t ) = 0∀t
∫
p
⎤
dt ⎥
⎥
⎦
1/ p
⎡
⎢ e1 + e2
⎢Ω
⎣
⎡
≤ ⎢ e1
⎢Ω
⎣
∫
p
⎤
dt ⎥
⎥
⎦
1/ p
⎡
+ ⎢ e2
⎢Ω
⎣
∫
p
⎤
dt ⎥
⎥
⎦
1/ p
These are the properties of a norm (distance)
SimHydro 2010
8
Objective Function Definition in Conceptual Hydrological Modelling
Guinot, Cappelaere, Delenne & Ruelland
Another possible approach: the weak form
• Desired property: • Weak form:
e(t ) = 0 ∀t ∈ Ω
∫ e(t ) w(t ) dt = 0
∀w(t )
Ω
• Particular case: SimHydro 2010
w(t ) = e(t )
p −1
,
p ≥1
9
Objective Function Definition in Conceptual Hydrological Modelling
Guinot, Cappelaere, Delenne & Ruelland
Another possible approach: the weak form
• Desired property: • Weak form:
e(t ) = 0 ∀t ∈ Ω
∫ e(t ) w(t ) dt = 0
∀w(t )
Ω
• Particular case: w(t ) = e(t )
p −1
,
p ≥1
∫
J p = a + b e(t )
Proposed Objective Function
p −1
e(t ) dt
Ω
Jp = a+b
∑
en
p −1
en
N
SimHydro 2010
10
Objective Function Definition in Conceptual Hydrological Modelling
Guinot, Cappelaere, Delenne & Ruelland
Very small differences…
Distance-based
p=1
J1 =
∫Q
sim
Weak form-based
− Qobs dt
VE =
Ω
∫ (Q
sim
− Qobs ) dt
Ω
p=2
MSE =
2
(
)
Q
−
Q
∫ sim obs dt
Ω
SimHydro 2010
J2 =
∫Q
sim
− Qobs (Qsim − Qobs ) dt
Ω
11
Objective Function Definition in Conceptual Hydrological Modelling
Guinot, Cappelaere, Delenne & Ruelland
… yield very different objective function behaviours
0.5
cP
dU
= cP(t ) − kU
dt
k
U
kU
0
0
c
0.2
Distance‐based, p = 1
Observed variable: discharge
SimHydro 2010
12
Objective Function Definition in Conceptual Hydrological Modelling
Guinot, Cappelaere, Delenne & Ruelland
… yield very different objective function behaviours
0.5
0.5
k
k
0
0
0
c
0.2
Distance‐based, p = 1
Observed variable: Q
SimHydro 2010
0
c
0.2
Weak form‐based, p = 1
Observed variable: Q
N.B.: This is the Volume Error VE
13
Objective Function Definition in Conceptual Hydrological Modelling
Guinot, Cappelaere, Delenne & Ruelland
… yield very different objective function behaviours
0.5
0.5
k
k
0
0
0
c
0.2
Distance‐based, p = 2
Observed variable: Q
SimHydro 2010
0
c
0.2
Weak form‐based, p = 2
Observed variable: Q
14
Objective Function Definition in Conceptual Hydrological Modelling
Guinot, Cappelaere, Delenne & Ruelland
Response surface monotony
If the following assuptions hold
• Q > 0, R > 0
• R – Q is an increasing function of U
• There exists as steady state solution
• Initial condition: U(0) ≥ 0
Then (Guinot & al., submitted)
• Weak form‐based objective functions are monotone in ϕ (no local minimum)
• Distance‐based objective functions
are not (local minima may appear)
SimHydro 2010
dU
= R (U , ϕ , t ) − Q(U , ϕ , t )
dt
Jp
D
W
ϕ
15
Objective Function Definition in Conceptual Hydrological Modelling
Guinot, Cappelaere, Delenne & Ruelland
The calibration process
Distance‐based
• Optimization process
• Local minima
• Equifinality
• Gradient‐based methods
may fail
• Global search procedures
• Optimum depends on p
Weak form‐based
• Root finding problem
• Monotone function
• Gradient‐based methods
can be used
• Search the intersection between hypersurfaces
defined for different
values of p
Both types of objective functions lead to similar optimal parameter sets
SimHydro 2010
16
Objective Function Definition in Conceptual Hydrological Modelling
Guinot, Cappelaere, Delenne & Ruelland
Moreover…
(from Guinot & al, submitted to Journal of Hydrology)
• Using different p in Jp allows redundant parameters to be
identified
• In reservoir‐based models, applying the same objective function to several model variables is more efficient for calibration than applying different objective functions to the same model variable
• Correspondence (Parameters ÍÎ Variables) for optimal calibration efficiency
SimHydro 2010
17
Objective Function Definition in Conceptual Hydrological Modelling
Guinot, Cappelaere, Delenne & Ruelland
Conclusions
• Weak form‐based objective functions may prove
easier to use than distance‐based objective functions in practice
• It is well‐admitted that calibration and validation should be carried out using different data sets
• We may also admit that they should be carried out using different types of objective functions !
SimHydro 2010
18
Objective Function Definition
in Conceptual Hydrological Modelling
V. Guinot, B. Cappelaere, C. Delenne, D. Ruelland
HydroSciences Montpellier UMR 5569 (CNRS, IRD, UM1, UM2)
SimHydro 2010
19