5
The simulated values related to the prior information are restricted in this work to be of the
form =
Σ
a
pj
b
j
, which are linear functions of b. Most prior information equations have only
one term with a coefficient equal to 1.0, so the contribution to the objective function
is simply the
prior information value of a parameter minus its estimated value. Additional terms are needed
when the prior information relates to a linear function that includes more than one parameter value.
For example, additional terms are included in a ground-water inverse model to account for the fol-
lowing circumstances: seasonal recharge rates are estimated and measurements of annual recharge
are available, so that the P
p
value equals the seasonal recharge rate and the summation includes
terms for the seasonal recharge rates; or storage coefficients in two model layers are estimated and
an aquifer test was conducted that measured the
combined storage coefficient, so that the Pp value
equals the storage coefficient from the aquifer test, and the summation includes terms for the layer
storage coefficients.
A simple problem and its weighted least-squares objective function surface are shown in
figure 1. The figure was constructed by calculating equation 1 for this problem using different sets
of parameter values T1 and T2. The log of the resulting numbers were contoured to produce the
contour map of figure 1B. For a linear problem, the objective function surface would be a smooth
bowl, and the contours would be concentric ellipses or parallel straight lines symmetrically spaced
about a trough. The nonlinearity of Darcy’s Law with respect to hydraulic conductivity results in
the much different shape shown in figure 1B.
The
differences
and
are called residuals, and represent the
match of the simulated values to the observations. Weighted residuals are calculated as
and
and represent the fit of the regression in the context
of how the residuals are weighted.
P'
p
b
( )
y
i
y'
i
b
( )
–
[
]
P
p
P'
p
b
( )
–
[
]
ω
i
1 2
⁄
y
i
y'
i
b
( )
–
[
]
ω
p
1 2
⁄
P
p
P'
p
b
( )
–
[
]
6
Figure 1: Objective function surfaces for a simple example. (a) The sample problem is a one-di-
mensional porous media flow field bounded by constant heads and consisting of three
transmissivity zones and two transmissivity values. (b) Log of the weighted least-
squares objective function that includes observations of hydraulic heads h1 through h6,
in meters, and flow q1, in cubic meters per second. The observed values contain no error.
(c) Log of the weighted least-squares objective function
using observations with error,
and a three-dimensional protrayal of the objective function surface. Sets of parameter
values produced by modified Gauss-Newton iterations are identified (+), starting from
two sets of starting values and progressing as shown by the arrows. (from Poeter and
Hill, 1997)
7
In equation 1, a simple diagonal weight matrix was used to allow the equation to be written
using summations instead of matrix notation. More generally, the weighting requires a full weight
matrix, and equation 1 is written as:
S(b) =
(2)
where the weight matrix,
and the vectors of observations and simulated values, and
include terms for both the observations and the prior information,
as displayed in Appendix A, and
is a vector of residuals. Full weight matrices are supported for most types of observations and
prior information in MODFLOWP. With a full weight matrix, MODFLOWP calculates weighted
residuals as
, where the square-root of the weight matrix is calculated such that
ω
1/2
is symmetric.
An alternative derivation of the least-squares objective function involves the maximum-
likelihood objective function. In practical application, the maximum-likelihood objective function
reduces to the least-squares objective function (as shown in Appendix A),
but the maximum-like-
lihood objective function is presented here and its value is calculated and printed by UCODE and
MODFLOWP because it can be used as a measure of model fit (Carrera and Neuman, 1986; Loa-
iciga and Marino, 1986). The value of the maximum-likelihood objective function is calculated as:
S’(b) = (ND+NPR) ln2
π
- ln
(3)
where
is the determinant of the weight matrix, and it is assumed that the common error variance
mentioned in Appebdix A and C equals 1.
Do'stlaringiz bilan baham: