Vous êtes sur la page 1sur 5

GROUND2000 International Conference on Grounding and Earthing

Belo Horizonte - Brazil June 18-21, 2000

TWO LAYERS SOIL STRATIFICATION USING NUMERICAL OPTIMZATION METHODS

Jos Tavares de Oliveira Alexandre Magno de Almeida Correia Manoel Firmino de Medeiros Jnior
DEENTUFRN PPGEEUFRN DEELECAUFRN

AbstractThis article has the objective of demonstrating where:


the viability of using numeric methods of commercial (a) = apparent resistivity to spacing a;
libraries, for the parameters adjustment in the apparent 1 = first layer resistivity;
resistivity function. The adopted investigation universe
2 = second layer resistivity;
was the toolbox Optimization of MatLab, in order to explore
the numeric resorts already implemented in this k = reflection coefficient [k = (2-1)/ (1+2)];
computational tool. All the optimization methods contained H = first layer depth;
in the package were analysed, comparing their a = spacing between the rods for the of Wenner method;
performances by the stratification of several soils. The
method that presented better results, applied to two layers Equation (1) expresses a direct relationship between the
soil stratification, was the Gauss-Newton optimization spacings of the rods of the configuration of Wenner and
method, whose development will be presented in the paper. the respective value of (a). Therefore, the objective is to
The optimization method was applied to the mathematical obtain the values of 1, 2 and H, such that the
model of apparent resistivity of the soil, with measured expression reproduce, with the minimum possible error,
made them by the method of Wenner. The results obtained the set of measured values. In this way, one tries to
for different soils types were satisfactory, in terms of cancel the deviations among the measured and
computational effort and numeric values, what is proven
calculated values, minimizing the function:
through the comparison with other methods, as Newton
Method and the Uniformly Accelerated Search Method.
a j medido a j
2
Using the methodology, as proposed, presents good
1 , 2 , H
N

results and simplifies the solution of stratification problem,


j 1 a j medido
(2)
becoming unnecessary a detailed development of specific
computer programs.
where N is the number of accomplished measurements.

The analysis was led to determine the best method that


1 - INTRODUCTION
is adapted to minimization equation (2), under the
numerical methods implemented in the toolbox
Design of mesh grids requires a soil stratification
Optimization of MatLab [2] (more specifically using the
procedure. The results of stratification are the resistivities
function leastsq). The analyzed methods were:
and the thicknesses of the layers. The solution process
is modelled as a adjustment of the resistivity function
Levenberg-Marquardt method with quadratic
parameters, in order to represent the measured real
interpolation and extrapolation (LMM2);
values. As usual, the adjustment is carried out by
application of an optimization method. In this way, some Levenberg-Marquardt method with cubic
numerical methods were investigated, all of them interpolation and extrapolation (LMM3);
belonging to optimization toolbox of MatLab. The aim Gauss-Newton method with quadratic interpolation
was to test usually implemented methods of commercial and extrapolation (GNM2);
sotware packages, as well to recommend the best one, Gauss-Newton method with cubic interpolation and
for the purposes of two layers stratification. extrapolation (GNM3).

2 TWO LAYERS SOIL STRATIFICATION


3 OPTIMZATION TECHNIQUES
Two layers stratification is, in most cases, the best model
to represent its resistivity variations of soil. By this model, 3.1 - GAUSS-NEWTON METHOD [2]
apparent resistivity can be represented as a function of
distances between rods, in agreement with measures The Gauss-Newton method consists in a variation of
accomplished by Wenners Method, according to: Newtons method. Let f(x) be a real function, where x a
vector of n variables. Expanding the function f(x) in
Taylor series and taking its quadratic approximation, the
minimum of f(x) can be reached using eq. (3):



(1)
a 1 1 4 k n
1 1

x k 1 x k 2 f ( x k ) 1 f ( x k ) T
2 2
n 1
2 n H 2 n H (3)
1 a 4

a

GROUND2000 International Conference on Grounding and Earthing
Belo Horizonte - Brazil June 18-21, 2000

In order to improve the convergence of the method a 3.1 - LEVENBERG-MARQUARDT METHOD


parameter should be introduced, resulting in:
As previously stated, the Gauss-Newton iteration can fail
xk1 xk k2f ( xk )1f ( xk )T (4) since the second order term in eq. (8) is often not
negligible, especially in the cases when the residuals are
large. However, if instead of replacing x by x+ one uses
where k is a positive search parameter, selected to
x +e, where e is a step-length parameter, the
minimize the function f, 2 f ( x )
1
is the (symmetrical) convergence of the method can be improved ([2]). Two
inverse of Hessian matrix (furthermore named Gk) and further problems for convergence are differences in scale
between elements of the Jacobian A and possible
gk f ( xk )T . Thus, the function can be minimized using T
singularity of A A. In order to avoid these problems, the
the following search: following algorithm was proposed by Marquardt (1963):

xk1 xk Gk gk **** (5) 1 given xk and vk (weight parameter), g(xk) e G(xk);


2 factorise G(xk) + akI: if it is not positive defined, reset
The mathematical problem to be considered here is that vk = 4vk, and repeat (I is the identity matrix);
of minimizing: 3 solve of the following expression: (G(xk) + vI) = -
g(xk);
4 evaluate f ( xk k ) and hence rk;
r ( x)
m

f ( x) rT r
2
(6)
i 1
i
5 if rk 0.25 set vk+1 = 4vk
if rk 0.25 set vk+1 = vk/2
where r =r(x) is the vector of residuals, i.e., the difference otherwise set vk+1 = vk;
between measured and calculated values.
6 if rk 0 set xk 1 xk else xk 1 xk k
For a least square problem the derivative of f(x) is given
by: Here must be pointed out that, when the function is
sufficiently well behaved for the unmodified Gauss-
f ' ( x ) g( x ) f ( x ) 2Ar (7) Newton algorithm to work successfully, the modifications
m
described above introduce some waste of computational
f ' ' ( x ) G( x ) 2AA T 2 r r
i 1
i
2
i (8)
effort.

4 - RESULTS

where A( x) r1, r2 ,...,rm (Jacobian Matrix) Table 1 shows measurement data for some examples of
soil used to verify the performance of the methods. All
As r(x) is the function to be minimized, its components ri measurements were done by Wenners method. In the
become very small in relation to the other terms, so that present work the method of Schlumberger was not used
G(x) is approximated to: because of lack of data in the literature. Further
investigation with this method, using self obtained
2 T
G(x)= f(x)2AA ******* (9) measurement data, is in course.

Considering the value of G(x) defined in (9), equation (5) Table 2 shows a comparison between the methods
becomes: implemented by using the commercial library
Optimization of MatLab. It can be observed that Gauss-
x k 1 x k AA T 1
Ar (10) Newton Method with cubic interpolation and extrapolation
(GNM) is the one that is the most adequate to soil

AA T 1
Ar stratification in two layers.

Comparing now the performance of Gauss-Newton


Method (with cubic interpolation and extrapolation) with
Of Newton's basic method, it can be made a search of
both Newtons Method (MN) and the Uniformly
the parameter (and to define the new value of the
Accelerated Search Method (UASM) [4], proves the
iteration in the following way
efficiency of the former. The results are shown in table 3.
1 solves AkAk =-Akrk for =k
T

2 set xk+1=xk+k

Soil A B C D E
Depth Resist. Measure Resist. Measure Resist. Measure Resist. Measure Resist. Measure
GROUND2000 International Conference on Grounding and Earthing
Belo Horizonte - Brazil June 18-21, 2000

a (m) med (.m) med (.m) med (.m) med (.m) med (.m)
1 82.1 17.0 - - -
2 100.3 30.2 167.6 25.5 2471.4
4 102.6 42.4 68.7 19.7 2199.1
8 113.9 55.7 26.1 23.3 1072.3
16 149.1 69.5 31.2 43.6 576.4
32 274.8 106.9 - 80.0 415.5
Table 1: Data of the Used Soils

A B C D
Method LMM2 LMM3 GNM2 GNM3 LMM2 LMM3 GNM2 GNM3 LMM2 LMM3 GNM2 GNM3 LMM2 LMM3 GNM2 GNM3
Error 0.0338 0.0338 0.0338 0.0338 0.0683 0.0683 0.0683 0.0683 0.0409 0.0409 0.0409 0.0409 0.0526 0.0526 0.0526 0.0526
Iterations 162 116 73 56 94 64 47 32 276 88 52 36 804 424 76 64
Time (*) 17.50 10.19 7.94 5.88 7.25 6.81 9.63 3.44 23.00 6.50 6.88 2.75 69.38 35.69 6.88 5.50
1(.m) 93.24 93.24 93.24 93.24 13.29 13.29 13.29 13.29 282.32 282.32 282.32 282.32 20.86 20.86 20.86 20.86
2(.m) 1386.5 1386.5 1386.5 1386.5 93.72 93.72 93.72 93.72 26.01 26.01 26.01 26.01 8.6E+5 4.5E+5 8.9E+5 2.5E+5
H(m) 12.59 12.59 12.59 12.59 1.06 1.06 1.06 1.06 1.61 1.61 1.61 1.61 11.52 11.52 11.51 11.51

Table 2: Results for the modelling of the two layers soil with the methods implemented in the toolbox Optimization

A B C D E
Method MN USAM GNM MN USAM GNM MN USAM GNM MN USAM GNM MN USAM GNM
Error 0.0338 0.0338 0.0338 0.0681 0.0682 0.0683 0.0426 0.0426 0.0409 0.0525 0.0525 0.0526 0.0144 0.0144 0.0144
Iterations 10 117 56 4 220 32 5 56 36 31 176 64 5 54 32
Time (*) 4.08 469.39 5.88 2.04 146.94 3.44 2.04 16.33 2.75 161.22 210.20 5.50 2.04 28.57 3.04
1(.m) 93.24 93.20 93.24 13.18 13.41 13.29 278.44 277.61 282.32 20.87 20.87 20.86 2632.0 2632.4 2632.6
2(.m) 1385.5 1339.9 1386.5 93.58 93.96 93.72 26.11 26.07 26.01 5.1E+6 8.4E+6 2.5E+5 418.48 418.51 418.47
H(m) 12.59 12.52 12.59 1.05 1.08 1.06 1.62 1.62 1.61 11.56 11.56 11.51 4.26 4.26 4.26

Table 3: Results of Gauss-Newton Method for the modelling in the two layers soil compared with other numerical optimzation methods.
(*) The results of time are given in terms of a base, in other words, the numeric value given in the table is corresponding
10000
the how many times the computer used for the calculations it would take to calculate 1.05 .

presented as in the number of iterations in the


5 ANALYSIS OF THE RESULTS convergence process.

Is verified that the use of the optimization Gauss-Newton That shows that the use of commercial packages of
and Levenberg-Marquardt numerical methods are methods of numeric optimization is viable once the user
perfectly applied to the problem of two layers soil doesn't need to develop methods many times
stratification, once the errors resulting by application of compounds or to implement them in computer programs.
the methods are small and you tolerated, even when
compared with the errors obtained by using other method 7 - REFERENCES
already used for numeric optimization, as Newton's
method and the Uniformly Accelerated Search Method. [1] FLETCHER, R., Pratical Methods of Optimization, 2
Edition. Wiley, 1987
[2] NASH, J. C., Compact Numerical Methods for Computers,
6 - CONCLUSIONS 2 Edition, Adam Hilger, 1990.
[2] GRACE, ANDREW, Optimization Toolbox For Use with
MATLAB, The MathWorks, Inc, 1995
The use of methods for the process two layers soil [3] MEDEIROS, M. FIRMINO JR.; OLIVEIRA, J. TAVARES;
stratification facilitates a wide study of the behaviour of a VILAR, F. J. V., Mtodo de Busca Uniformemente Acelerada
soil, avoiding some considerable errors of other para Estratificao do Solo em n camadas, GROUND98
methods. International Conference on Grounding and Earthing, pp.131-
134, Belo Horizonte, Brazil, April 12-16 1998
All the methods of the toolbox of Optimization were [4] TAVARES, J. OLIVEIRA, Apostila do Curso de Proteo e
tested and applied to the problem, especially had good Aterramento de Sistemas Eltrico de Potncia. Apostila do
actings Gauss-Newton's Method, that presented good PPgEE-UFRN - Natal-RN, Brazil
results compared with Newton's method and Evenly [5] SILVA, EMERSON G. L.; MEDEIROS, M. FIRMINO JR. &
OLIVEIRA, J. TAVARES. Estratificao do Solo em duas
Search Accelerated method, so much in the error
GROUND2000 International Conference on Grounding and Earthing
Belo Horizonte - Brazil June 18-21, 2000

Camadas, Utilizando o mtodo de Newton como Tcnica de


Otimizao, XI CBA, Vol. 2, pp. 1179-1183, 1996
[6] CORREIA, ALEXANDRE M. DE A.., Relatrio Final de
Bolsista do CNPq PIBIC Universidade Federal do Rio Grande
do Norte, Natal-RN, Brazil, 1999

Main author
Name: Jos Tavares de Oliveira
Address: Departamento de Engenharia Eltrica
Campus Universitrio s/n
59072-970 Natal-RN Brazil
Fax: ; Phone: (55) 84 215-3738
E-mail: tavares@leca.ufrn.br
GROUND2000 International Conference on Grounding and Earthing
Belo Horizonte - Brazil June 18-21, 2000

Vous aimerez peut-être aussi