Sunteți pe pagina 1din 8

8th International Congress on Advances in Civil Engineering, 15-17 September 2008

Eastern Mediterranean University, Famagusta, North Cyprus

Combined Simulation-Optimization of an
Excavation Site for Dewatering Purpose

K. Demirbaş1, A. B. Altan-Sakarya1 and H. Önder1


1
Department of Civil Engineering, METU, Turkey

Abstract

Optimal decision of dewatering of an excavation area in an unconfined aquifer is


searched by using the combined optimization-simulation approach. In combining
groundwater model with the optimization model, response matrix approach,
generated on the basis of linear superposition, is used. MODMAN that calls
MODFLOW repeatedly to generate response matrix is utilized to transform the
groundwater management problem into linear integer programming problem. A
hypothetical dewatering area is considered to examine the effects of nonlinearity
due to pumpage. Two different iterative methods, revised unit rates and revised
coefficients, are applied to account for the nonlinear response of the unconfined
aquifer to pumpage. Genetic Algorithm based on selection, crossover and
mutation is also used as an optimization tool to compare the results. The study
has shown that modified linear optimization procedure used in conjunction with
MODFLOW provides a useful and computationally efficient tool in the analysis
of groundwater dewatering problems in unconfined aquifers.

Keywords: Groundwater; Optimization; Nonlinear; Excavation; Genetic


Algorithm.

1 Introduction

Construction works require excavations that more than often reach beyond
groundwater. In order to obtain suitable working conditions, water levels should
be lowered to certain heights at specific locations. The most common objective
during the dewatering process is to minimize the total amount of water extracted.
Combined simulation-optimization method incorporates aquifer response to
stress, directly in the constraints.

1
ACE2008

The first application of response matrix in groundwater started with Deninger


(1970). Aguado and Remson (1974) first combined the optimization approach
with the numerically simulated unconfined aquifer with one dimensional flow.
Methods to utilize response matrix approach for the solution of a nonlinear
management problem were studied by Rosenwald and Green (1974). Heidari
(1982) used drawdown correction factor proposed by Jacob (1944) to generate
response matrix for an unconfined aquifer. Danskin and Gorelick (1985)
described an iteration procedure to remove nonlinearity in unconfined aquifers.
Jonoski et al. (1997) applied response matrix approach for the design and
optimization of an artificial recharge-pumping system

The present study concentrates on the management of dewatering of unconfined


aquifers by using linear response matrix approach. The aim is to introduce and
study different iterative methods to deal with the errors resulting from nonlinear
response of unconfined aquifers to pumpage. For the simulation of groundwater
model, MODFLOW is used (Mcdonald and Harbaugh, 1988). To transport the
pumpage-drawdown relation into the optimization formulation, response matrix
approach is utilized. MODMAN (Greenwald, 1998a and b) is used to generate
response matrix for the optimization. The linear management formulation is then
solved by optimization program, LINDO. Additionally, the problem is solved by
using Genetic Algorithm and the results are compared with each other.

2 Hydraulic System Simulation

2.1 Simulation Model

In the case where principle axes of conductivity are parallel to the principle
coordinate axes, the governing groundwater flow equation for two dimensional
horizontal flow in unconfined aquifer is as follows,

∂ ∂h ∂ ∂h ∂h
( K x ( h − σ) ) + ( K y ( h − σ) ) + R = S y (1)
∂x ∂x ∂y ∂y ∂t

where Kx and Ky are the components of the hydraulic conductivity tensor in x and
y coordinate directions, respectively, h is the hydraulic head, R is the vertically
averaged source or sink term representing flow rate per horizontal area, Sy is the
specific yield, σ is the elevation of the bottom of the aquifer , (h-σ) is the
saturated thickness. Eq. 1, together with the specification of flow and/or head
condition at the boundaries of a flow domain and specification of initial-head
condition, constitutes a mathematical representation of groundwater flow
water system.

2.2 Generation of Response Matrix

Unit response matrix is formed by running the simulation model repeatedly with
unit discharge at each specified well. The response to the stress is assumed to be
linear. Hence, if the well rate is doubled, the drawdown induced by that well will

2
K. Demirbaş, A.B. Altan-Sakarya, H. Önder

also be doubled. Response of the simulation to different impulses is also added


by the principle of superposition, that is, the total drawdown induced by two
wells is equal to the sum of the drawdowns induced by each well. The basic
formula that enables the use of linear theory in groundwater management
problems in discrete form is given by Maddock (1972). For one stress period,
after dropping the time related terms, the equation takes the following form,

n
s j = ∑ Qiβ j,i (2)
i =1

where, sj is the drawdown at j, Qi is the pumpage at ith well location, βj,i is the
steady state system unit response function or unit drawdown at j as a result of unit
pumpage at well i, and n is the total number of wells.

2.3 Management Model

The general form of objective function, f, in linear form is,

n
Minimize f = ∑ λ i Qi (3)
i =1

where, λi is the coefficient for the pumpage at the ith well.

The model is subjected to set of constraints which may be listed as follows;


a) Water level should be less or equal to the value to obtain dry excavation area
during construction.

h j ≤ h uj (4)

where, hj is the head at jth location and hju is the maximum head allowed at jth
location. In order to include the pumpage-drawdown relation in the optimization
formulation, hj should be defined in terms of decision variable, pumpage rate.
Using Eq. 2, head at location j for a steady state solution can be written as,

n
h j + ∑ Q iβ j,i = h 0 (5)
i =1

where, h0 is the initial steady state water level.


b) The general form of stress constraint can be written as,

Qil ≤ Qi ≤ Qiu (6)

where, Qil and Qiu are the minimum and maximum pumpage rates allowed at ith
well location, respectively.

3
ACE2008

3 Applications and Discussion of Results

A hypothetical construction dewatering example which was also studied by


Degirmenci (1997) Altan-Sakarya and Önder (2003), and Demirbaş (2003) is
used (Fig. 1). The aquifer is considered to be unconfined, homogenous and
isotropic. Hydraulic conductivity, K is assumed to be constant, and is 25 m/day.
There are 21 potential wells and 10 control points (Fig. 2). Maximum capacities
of the wells are 5000 m3/day. The head at the excavation area is to be lowered by
4 m in 5 days. The initial water table is horizontal and at 20 m. The objective is to
minimize the total discharge through the wells while satisfying the 4 m
drawdown at control points.

The problem is simulated by Visual MODFLOW. MODMAN is utilized to


generate response matrix. Then, the developed linear optimization model is
solved by LINDO.

Figure 1. General Plan View of the Aquifer and Boundary Conditions.

Initially, the nonlinear response of unconfined aquifer to pumpage is tried to be


overcome with two different iteration methods. In the first approach, the revised
unit rate method is utilized to solve the optimization problem. In the second one,
the optimal drawdowns are tried to be equalized with the linear responses by
applying an iteration method based on modifying the coefficients of response
matrix. Afterwards, the problem is solved by using Genetic Algorithm for
comparison purposes.

In the revised unit rates method, initially, a unit rate is assigned to each well and
corresponding response matrix coefficients are calculated by MODMAN. Then,
LINDO is used to obtain optimal well rates. Afterwards, these optimal rates were
entered as unit rates for the next iteration. When the unit rates and optimal rates
become identical, iteration is stopped. For an unconfined aquifer, drawdown due

4
K. Demirbaş, A.B. Altan-Sakarya, H. Önder

to discharge is a nonlinear function. Response matrix approach assumes this


relation to be linear and to find the drawdown corresponding to a specific
discharge it multiplies the drawdown induced by the unit rate with the discharge.
Since response matrix consists of coefficients which are equal to drawdown due
to unit rates, as the unit rates gets closer to optimal rates, error due to linear
assumption decreases. And when unit rates become identical to optimal rates, that
error vanishes.

The revised unit rates method is repeated for three different initial unit rates (500,
1000 and 4000 m3/day). Each initial unit rate gives similar optimal rates after 4
iterations. For different initial unit rates, the same wells are chosen as the
operating wells. The optimal rate for each well is also similar. Total optimal rates
are computed as 5927.87, 5927.94 and 5927.87 m3/day for initial unit rates of
500, 1000 and 4000 m3/day, respectively.

Figure 2. Locations of the wells and control points

Trying to equalize the unit rate to the optimal pumping value for a well is an
improvement for the solution. However, summation of the responses due to each
well to find the response of multiple wells introduces additional error. To deal
with this error, an iteration method proposed by Tokgöz et al. (2002) which is
based on the method brought by Danskin and Gorelick (1985) is applied to the
optimization problem.

The procedure is based on revising the response matrix coefficients of the current
iteration by using the results of the previous iteration. In MODMAN,
management goals are input as constraints and the response matrix coefficients
are generated by running MODMAN. The unit rate is chosen and left as 1000
m3/day throughout the solution. The optimal rates and optimal drawdowns due to
linear responses to optimal rates are found by the optimization program, LINDO.
Then, MODMAN is executed once again in order to perform simulation based on

5
ACE2008

optimal rates. It is obvious that, due to nonlinearities, the simulated drawdowns


obtained after the simulation of optimal rates will not be same as the optimal
drawdowns (linear responses to optimal rates) obtained by LINDO. In the next
iteration, response matrix coefficients are updated by multiplying the
corresponding ratio of the simulated drawdowns with optimal drawdowns
obtained by LINDO. Afterwards, LINDO is executed again. The iteration process
continues till this ratio is converged to one.

For this method, the optimal rates are converged after three iterations for the
problem considered. An increase in the number of iterations may be expected, if
number of stress periods or wells increase. Throughout the iteration procedure,
the total pumpage is decreased by almost 7% and is found to be 5504.4 m3/day at
the end of 3rd iteration.

Revised coefficient technique seems to be superior to the revised unit rates


method. Not only it removes the error resulting from drawdown-pumpage
relation in a single well, but it also eliminates the error due to the linear
summation of overlapping cones of depression. Table 1 gives the optimal well
rates found by using revised unit rates and revised coefficients together with the
optimal results obtained by using a third approach, genetic algorithm.

To test the validity of the solutions, a third method which is not using the
response matrix approach is utilized. Genetic algorithm which is based on
selection, crossover and mutation, is applied to the management problem.
MODFLOW is called whenever the optimization module (genetic algorithm)
needs the drawdowns for the pumpages considered. It is known that genetic
algorithm is a powerful search and optimization technique for optimal solutions
when conventional techniques are not adequate (Goldberg, 1989). For this
problem it is only used for comparison purposes. The genetic algorithm reached
the optimal solution at 315th generation of a total generation of 400 and with a
population number of 100. The optimal total rate found by genetic algorithm is
5593.8 m3/day which is approximately 2% greater than the optimal result found
by the revised coefficient method. It can be seen form the results that, for genetic
algorithm approach, all the wells have a nonzero value of pumpage, which in
reality is not practical. This problem might be overcome by adding a limit to the
number of wells. For this problem, another disadvantage of the genetic algorithm
is the duration of simulation time which is 10 hours and 16 minutes with a Core
Duo processor, while the solution for both revised unit rates and revised
coefficient lasted only a few seconds.

4 Summary and Conclusion

The applications of response matrix approach for the optimization of the


excavation problems in unconfined aquifers are presented. The accuracy of using
response matrix approach in unconfined aquifers is observed and iterative
methods are applied to remove the errors occurred due to nonlinear character of
unconfined aquifers. Studies are carried out to predict the optimal well rates and
well locations to minimize either total pumpage or capital cost, to obtain a certain
amount of drawdown in an unconfined aquifer.

6
K. Demirbaş, A.B. Altan-Sakarya, H. Önder

Table 1. Optimal well rates corresponding to three optimization methods.

Revised Unit Revised Genetic


Rates Coefficients Algorithm
Optimal Well Optimal Well Optimal Well
Wells
Rates (m3/day) Rates (m3/day) Rates (m3/day)
Q1 0.0 0.0 70.9
Q2 0.0 0.0 156.8
Q3 0.0 0.0 484.4
Q4 1463.6 1083.4 34.5
Q5 0.0 0.0 18.5
Q6 0.0 0.0 101.5
Q7 374.4 411.2 147.4
Q8 0.0 0.0 1305.4
Q9 908.9 842.6 93.9
Q10 804.0 737.6 727.0
Q11 0.0 0.0 43.2
Q12 615.9 531.3 59.6
Q13 0.0 0.0 20.5
Q14 0.0 0.0 674.6
Q15 0.0 229.1 161.9
Q16 621.0 559.6 215.2
Q17 0.0 0.0 131.5
Q18 0.0 0.0 17.8
Q19 0.0 0.0 131.4
Q20 1140.1 1109.5 909.4
Q21 0.0 0.0 88.3
TOTAL 5927.9 5504.4 5593.8

It is shown that the results of the optimization problem depend on the unit rates
chosen for the generation of the response matrix. Revised unit rate technique
gives identical results for a range of different starting unit rates and since the
errors coming from the nonlinear pumpage-drawdown relation in unconfined
aquifer diminishes; it gives a good approximation to the exact optimal rates for
small excavation problems.

As the number of wells and their extraction rates increase, the effect of
overlapping cones of depression, which is not considered in revised unit rates
method, increases. And for above cases nonlinearities may significantly affect
management results. The iterative method which is based on revising the
coefficients of response matrix not only eliminates the error due to linear
summation of the drawdowns induced by separate wells but also the error due to
nonlinear response of unconfined aquifer to pumpage induced by one well. And
as the results from genetic algorithm confirms, we believe that we found the near
global optimum results.

7
ACE2008

The linear response matrix approach in conjunction with MODFLOW is also


computationally efficient method. The genetic algorithm reached optimal solution
with a simulation time lasting more than 10 hours, while the solution for both
revised unit rates and revised coefficient lasted only a few seconds.

References
Aguado, E., and Remson, I. (1974). Ground-water Hydraulics in Aquifer
Management, Journal of Hydraulics Division, ASCE, Vol.100 pp. 103-117.
Altan Sakarya, A. B., and Önder, H. “İnşaat Kazı Sahasının Optimum Drenajı”
(Optimal Dewatering of an Excavation Site), I. Ulusal Su Mühendisliği
Sempozyumu, 22-26 September 2003, Gümüldür, İzmir, pp.169-180 (in Turkish)
Danskin, W. R., and Gorelick, W. R. (1985). A Policy evaulation tool: Management
of a multiaquifer system using controlled stream charge, Water Resour. Res.,
21(11), 1731-1747.
Değirmenci, M. (1997). Optimal dewatering system for an excavation of construction
foundation, MS thesis, Middle East Technical Univ., Ankara, Turkey.
Demirbaş, K. (2003). Combined optimization-simulation of an excavation site for
dewatering purposes, MS Thesis, Middle East Technical Univ., Civil Eng. Dept.,
Ankara, Turkey.
Deninger, R. A. (1970). System Analysis of Water Supply Systems, Water Resources
Bulletin, 6(4), pp. 573-379.
Goldberg D. E. (1989), Genetic Algorithms in Search, Optimization and Machine
Learning, Addison-Wesley-Longman, Reading,Mass.
Greenwald, R. M. (1998a). MODMAN 4.0 Windows-Based Preprocessor,
Preprocessor Version 1.02 (HSI Geotrans, Frehold, New Jersey).
Greenwald, R. M. (1998b). MODMAN An Optimization Module for Modflow
Version 4.0, Documentation and Users’s Guide, (HSI Geotrans, Frehold, New
Jersey).
Heidari, M. (1982). Application of linear’s system theory and linear programming to
groundwater management in Kansas, Water Resour. Bull., 18(6) 1003-1012.
Jacob, C.E. (1944). Notes on determining permeability by pumping tests under
watertable conditions, USGS Open File Report, In: USGS Water Supply Paper
1536-I, 1963, pp. 245-271.
Jonoski, A., Yangxiao, Z., and Nonneri J. (1997). Model-aided design and
opimization of artificial recharge-pumping systems, Hydrological Sciences, 42(6),
937-953.
Lindo Systems (1996). LINDO User’s Manual (Lindo Systems, Chicago, Illinois).
Maddock, T. (1972). A Ground-water Planning Model: A Basis for a Data Collection
Network, International Symposium on Uncertanities in Hydrologic and Water
Resources Systems, Int. Assoc. Sci. Hydrol., Univ. Of Ariz., Tuscon.
McDonald, M. G., and Harbaugh A. W. (1988). A modular three-dimensional finite
difference groundwater flow model, Modelling Techniques, Book 6, Chapter A1,
586 p.
Rosenwald, G. W., and Green, D. W. (1974). A Method for Determining the
Optimum Location of Wells in a Reservoir using MIP, Soc. Pet. Eng. J., pp. 44-54.
Tokgöz, M., Yılmaz, K. K., and Yazıcıgil, H. (2002). Optimal aquifer dewatering
schemes for excavation of a collector line, Journal of Water Resour. Planning and
Management, 248-261.

S-ar putea să vă placă și