84
resent the system of concern, and is the model to be calibrated. A linear approximate model can be
represented as:
y = b
0
+ b
1
X
1
+ b
2
X
2
+ . . .
β
j
X
j
. . .
β
n
X
n
+ e = y
’
+ e
(C2)
where
y is a measurement of the dependent variable (here, hydraulic heads, flows, and so on), as above;
b
j
are the estimated parameter values;
X
j
are the independent variables (here,
location, depth, time, etc.);
n
is the number of terms in the approximate model;
e is the true error; and
y
’
is the simulated equivalent of the measured dependent variable.
Approximate nonlinear model. As for the true model, the approximate nonlinear model
can not be represented as in C1, and requires the more general form presented after equation 1 --
that is,
using vector notation, y= f(b,
ξ
) + e, where f repesents the form of the unknown nonlinear
function,
ξ
represents the independent variables, and the other symbols are as defined for equation
C2.
Linearized approximate nonlinear model. The linearized approximate nonlinear model is
produced using a Taylor series expansion about a defined
set of parameter values, b*. Within an
additive vector that is constant for any b* (this vector is needed to derive the interative equation
4a, but is not important to the present discussion), the linearized approximate nonlinear model can
be expressed in the form of equation C2. In this situation however, the X
j
are no
longer simply in-
dependent variables, but equal the derivatives of the approximate linear model with respect to the
parameter values, evaluated at b*. These derivatives were defined for equation 8 and have the fol-
lowing characteristics:
1. Like the X
j
for linear problems, the derivatives include the independent variables; but they also
include the effects of other aspects of the nonlinear model.
2. Because of model nonlinearity, the values of the derivatives depend on
the parameter values in
b*.
3. The derivatives generally are called sensitivities because the represent the senstivity of the sim-
ulated value to a change in the parameter value.
Linearized models reproduce the same simulated value at b* as the nonlinear model, by
definition, and often closely mimick the nonlinear model for values of b near b*. As the linearized
model is evaluated for values further from b*, simulated values will vary from those of the approx-
imate nonlinear model depending on its degress of nonlinearity. This deviation
is apparent in the
sum-of-squared residuals surfaces of figure 2, which shows an objective-function surface calculat-
ed using the Theis equation as the approximate nonlinear model, and two objective-function sur-
85
faces calculated using a linearized approximate model. The linearized surfaces closely mimic the
nonlinear surface near the b* values,
marked by an x, and mimick it less well, and even poorly, for
increasingly different sets of parameter values.
Do'stlaringiz bilan baham: