Energy 136 (2017) 32e44

Contents lists available at ScienceDirect

Energy
journal homepage: www.elsevier.com/locate/energy

Optimal management of a theoretical coastal aquifer with combined

pollution and salinization problems, using genetic algorithms
Y.N. Kontos a, b, *, K.L. Katsifarakis a
a
Department of Civil Engineering, Aristotle University of Thessaloniki, 54006, Thessaloniki, Greece
b
Department of Civil Engineering (Technological Education), Technological Educational Institute of Thessaly, Larissa, Greece

a r t i c l e i n f o a b s t r a c t

Article history: The paper discusses optimal management of a theoretical coastal aquifer, providing water for drinking
Received 17 November 2015 and/or irrigation purposes, which is threatened by seawater intrusion from the coast and by non-
Received in revised form conservative pollutant plumes from the inland. A new computational tool, able to address the com-
8 July 2016
bined pollution-salinization problem, is used. It optimizes the classic Pump-And-Treat and Hydraulic
Accepted 12 October 2016
Available online 25 October 2016
Control techniques without compromising aquifer's sustainability. Οptimization entails minimization of
pumping, pipe network construction and pumped polluted water's remediation costs. Practically, the
goal is: find the best distribution of total required flow rate to existing wells and the best locations and
Keywords:
Coastal aquifer
flowrates of additional abstraction wells, in order to protect the aquifer with minimal management costs.
Groundwater pollution The respective objective function includes a complex penalty function. The optimization technique used
Seawater intrusion is a binary genetic algorithm including elitism. In order to maintain a reasonable balance between
Boundary element method computational volume and accuracy, a simplified equivalent 2D groundwater flow field is simulated by a
Particle tracking boundary element method, while advective mass transport (pollution spread and seawater intrusion) is
Genetic algorithms simulated by a particle tracking code.
© 2016 Elsevier Ltd. All rights reserved.

1. Introduction
Water and energy are closely interrelated, both at the planet
scale and at the scale of human works. At the planet scale, the
hydrologic cycle, which produces renewable fresh water resources,
is fueled by the energy of the Sun. Water, on the other hand, due to
the energy that it carries, changes the shape of the Earth, either
abruptly, as during flood events, or imperceptibly, during long pe-
riods of time.
The relationship between water and energy resources is quite
close at the scale of human works, too. Quite often we correlate them
in a negative way, focusing on water pollution and consumption due
to energy production, use and transport (e.g. Ref. [1]). This is part of
the whole picture, though. For each water drop that reaches our
homes, energy has been used at different stages, e.g. to pump ground
water, to treat it, to render it potable and to transport it to our houses,
through water supply networks (e.g. Ref. [2]). No water deficit
problem would exist, if we had an abundant, cheap and safe energy
* Corresponding author. Department of Civil Engineering, Aristotle University of
Thessaloniki, 54006, Thessaloniki, Greece.
E-mail addresses: ykontos81@gmail.com (Y.N. Kontos), klkats@civil.auth.gr
(K.L. Katsifarakis).
source for sea water desalination. Moreover, energy (together with
chemicals) is required to run sewer treatment plants.
Quite often, problems of optimal water resources management
are mainly energy management problems. This is particularly true
with optimal management of aquifers, which are threatened by
natural or anthropogenic pollution sources, such as industrial ac-
tivity, energy production, intensive agriculture including system-
atic use of fertilizers and pesticides, oil leakage or leakage from
legal or illegal landfills, etc.
Optimal management of polluted aquifers requires complex
pollution control or remediation techniques [3], which include
installation and operation of a well network, in order either to
control the spread of a contaminant via manipulation of ground-
water levels and flow directions (Hydraulic Control-HC) or to
reduce the contaminant mass, pumping it partially or thoroughly,
in order to meet a target concentration or global mass fraction and
even treat it accordingly (Pump and Treat-PAT) [4].
Depending on the way the pollution plumes are captured or
contained [5], contaminant control methods can also be classified
as a) concentration control (maximum concentration levels
compliance at control points) [6], b) hydraulic control (predefined
head difference, gradient, or velocity constraints at specific points)
[7], and c) advective control [8]. Practically, the aforementioned

http://dx.doi.org/10.1016/j.energy.2016.10.035
0360-5442/© 2016 Elsevier Ltd. All rights reserved.
 Y.N. Kontos, K.L. Katsifarakis / Energy 136 (2017) 32e44 33

2
techniques, as implemented here, are different versions of the   X Z
advective control approach, specifically the Particle Tracking 1 6XN
Qw 1 N
r ij $n
hi ¼ $4 ln þ hj $ dGj
Method (PTM), where hydrodynamic dispersion is neglected 2p j¼1 T riw j¼1
rij2
[9e12]). Gj
3
The ultimate goal of the ongoing research is optimal manage-
X
N Z  
vhj 1 7
ment of polluted aquifers, with minimum total solution cost, þ $ ln dGj 5 (1)
without compromising aquifer's further sustainability. In all cases, j¼1
vn riw
Gj
energy is the main cost item. From the computational point of
view, the target is to decide the number of new wells and their
where, N is the number of boundary elements, Qw is the flow rate of
coordinates, together with the new and existing wells' flowrates.
well W, riw and rij are distances shown in Fig. 1a. The line integrals
The optimization process can vary from simple series of tests or
have been analytically evaluated [21] and applied to various
trial and error procedures [13], to linear [14] or non-linear pro-
groundwater modelling studies [12,20,22].
gramming [15,16], and even to heuristics or meta-heuristics and
In cases of aquifers with locally homogeneous zones of
modern evolutionary algorithms [9e12,16e18]. The latter often
different transmissivities, two assumptions are made: a) in each
require simplification of the flow field (equivalent 2D fields)
zone k transmissivity Tk is constant and the Poisson equation
due to the excessive computational load deriving from their
applies and b) at every point along the two zones' interface, the
iterative nature.
compatibility and continuity equations apply [21,22]. The
Coastal aquifers often face an additional form of water quality
next step is calculation of the velocities at internal points
degradation: salinization due to the water table drawdown near
within the different zones of the aquifer by means of Eqs. (2)
the shoreline, resulting from excessive pumping. In such cases,
and (3), in order to simulate advective transport of pollutant
re-planning of pumping schemes (flow rates or/and well loca-
and seawater particles, as implemented before in similar
tions) is required, in order to guarantee operational continuance,
problems [11,12,22].
without salinization of pumping wells or further seawater
intrusion.
K X
W
Qw $cw X N
l $ðl $c $cos a þ l3 $c1 $cos b
The general class of all the above problems is that of con- Vx ¼ $  hj 1 2 2
strained, nonlinear, stochastic, multi-objective optimization 2$p$n w¼1 T$riw
2
j¼1 ðl2 $l3 Þ2 $sin g
problems [3]. In that class, optimization of HC and PAT manage-
X   
N
vhj g$c2 c  c2 l
ment strategies has been extensively studied in inland polluted  $ þ 1 $ g$cot a þ ln 2
aquifers and coastal aquifers under sea intrusion separately, but j¼1
vn l3 $sin a l1 l3
there is little research concerning the combined problem of
(2)
optimal management of a coastal aquifer threatened by both
pollution plumes from the inland and sea intrusion from the
coastline. This paper presents a new software application “Opti- K X
W
Qw $dw X N
l $ðl $d $cos a þ l3 $d1 $cos b
Vy ¼ $  hj 1 2 2
Manage” (Visual Basic), which deals with the combined problem 2$p$n w¼1 T$riw
2
j¼1 ðl2 $l3 Þ2 $sin g
of the optimal management of a coastal aquifer, threatened
X   
simultaneously by pollution plumes from the inland and sea N
vhj g$d2 d  d2 l
 $ þ 1 $ g$cot a þ ln 2
intrusion from the coastline. When a given total fresh water flow
j¼1
vn l3 $sin a l1 l3
rate is required from an existing pumping scheme, “OptiManage”
can find the best flow rate distribution among the existing wells (3)
and decide on the suitable locations and flow rates of additional
wells in order to protect the existing ones from pollution and all where, c1 ¼ xj  xi, d1 ¼ yj  yi, c2 ¼ xj þ 1  xi, d2 ¼ yj þ 1  yi,
wells from pumping seawater (problem version A1) during the cw ¼ xw  xi, dw ¼ yw  yi, while a, b, g, l1, l2 and l3 are shown in
studied period of time (i.e. non-conservative pollutant's deacti- Fig. 1b.
vation period), with minimum management costs. Alternatively, The study period is discretized into equal timesteps. Particles of
salinization prevention can refer to the aquifer in general and not infinitesimal mass simulate pollutants or seawater. Their displace-
only to the wells. In this case, an inflow check in the coastline ments during each timestep DΤ are calculated by local velocity
boundary elements is used (problem version A2 [12]). components, which are assumed to be constant during each DT. In
this way, the trajectory of each particle is calculated as a crooked line.
2. Flow field simulation (BEM) and mass transport (PTM) The finer the discretization (larger number of DΤs), the more realistic
the trajectory is and the more it resembles a continuous curve.
The use of genetic algorithms as the optimization tool dictates
simplification of the hydraulic model in order to reduce the vast
computational load. Hence, an equivalent 2D flow field is studied
and a simplified advective PTM is used to simulate advective
pollutant transport only [19], while the flow field is simulated using
the BEM [12,20].
BEM is based on Green's 2nd law. The hydraulic head h and
velocity V at internal points of a flow field U are calculated from
the values of h and q ¼ vh/vn along the boundary S of U. During
numerical implementation, the boundary S is divided into N line
segments (boundary elements), where h and q are assumed
constant. Missing h and q values on the boundary elements are
produced first, through solving a system of N equations and
unknowns. Then h at any internal point of the field can be
calculated as: Fig. 1. Auxiliary figure related to the formulation of boundary elements.
34 Y.N. Kontos, K.L. Katsifarakis / Energy 136 (2017) 32e44
The theoretical confined, isotropic coastal aquifer (thickness
b ¼ 50 m, hydraulic conductivity K ¼ 104 m/s and porosity n ¼ 0.2)
with a plane, horizontal, single-phase, steady flow, which is studied
in this paper, consists of three discrete homogeneous zones of
different transmissivities (Fig. 2). The aquifer's southern boundary
is actually the coastline, hence it is a boundary of constant hydraulic
head h ¼ 0 m (Dirichlet boundary condition). Its northern boundary
is also a constant head boundary (h ¼ 50 m), while the western and
eastern ones are impermeable (q ¼ vh/vn ¼ 0, Neumann condition),
where n stands for the direction vertical to the boundary. Due to
intrinsic constraints of BEM, exhibiting reduced accuracy near the
boundary elements (distance smaller than half the element's
length), new wells can be located only in specific permitted zones
in secure distances from boundaries (here equal to boundary ele-
ments' length, as shown in Fig. 2).
3. Minimization of management cost scenario
The hypothetical current state of the theoretical coastal aquifer,
the management cost of which is to be minimized, is shown in
Fig. 3. The operation of 3 existing pumping wells (Х1 ¼ 501 m,
Y1 ¼ 631 m, Х2 ¼ 1321 m, Y2 ¼ 1012 m, Х3 ¼ 1081 m, Y3 ¼ 1351 m,
radius of 0.25 m), aimed to supply a constant flow rate of 250 l/s of
clean drinking or irrigation water, leads to: a) seawater intrusion
Fig. 2. Theoretical polluted coastal aquifer with permitted zones for installation of new wells.
through the coastline and salinization of well 2 and b) pollution of
well 3 by the initially circular plume (center: 752 m, 1952 m and
radius ¼ 70 m), before the end of the study period (1000 d).
The optimal locations of two additional wells (radius of 0.25 m),
of predefined minimum (10 l/s) and maximum (120 l/s) flow rate
each, together with the proper distribution of the total required fresh
water flow rate among the existing wells (250 l/s), is studied, in order
to ensure the uninterrupted provision of clean groundwater during
1000 d, with minimum management cost. The additional wells can
be used to retard or avert polluted water (HC), or even pump it
partially or thoroughly in order to be treated (PAT). The minimum
constraint of the 10 l/s for each additional well is set in order to direct
the algorithm away from otherwise often proposed solutions that
include additional wells of too small, practically unrealistic, flow
rates [11,12]. These unwanted solutions are a direct result of the
simplified mathematical simulation regarding the pollution of a well
criterion, which does not take into account the volume of polluted
water that is represented by each particle (see par. 3.1).
3.1. Particle tracking and boundary element method configuration
The authors' previous research [11,13] has offered guidelines or
investigation techniques to define best parameter values, such as
the number of particles for the plumes' simulation, the temporal
 Y.N. Kontos, K.L. Katsifarakis / Energy 136 (2017) 32e44
Fig. 3. Theoretical coastal aquifer with pollution and salinization problems.
discretization of the study period, the approach for simulating
wells' pollution, the value of the radius of wells' “pollution arrival”
zones, the BEM-related discretization of internal and external field
boundaries for the accuracy-efficiency balance and the prolonga-
tion percentage of the pollutant's active period (indirect inclusion
of dispersion), so as to be on the safety side of calculations.
Practically, 16 points symmetrically placed along the initially
circular pollution plume boundary, serving as starting points for an
equal number of particles, are adequate to represent the plume
spread via calculation of the particles' trajectories during the study
period. As far as the seawater front is concerned (problem version
A1), 3 particles, starting from 3 points symmetrically placed along
each coastline boundary element, can represent the seawater front
advance.
The substantial influence of the temporal discretization of the
study period, imposed by the PTM on the produced pollutant or
seawater particles' trajectories and thus on the prediction of the
pollution plume spread and seawater front advance, is vividly
showcased in the investigation of 4 TS-TP combinations (where TS
is the DT duration in days and TP is the number of DTs) in the
simplified homogeneous flow field of Fig. 4. It is obvious that the
temporal discretization of the study period does not only affect
the graphical quality and realism of the produced trajectories, but
it may also lead to inaccurate and false predictions about the
pollution spread and seawater front advance. Given the fact that
Fig. 4d discretization (1000 timesteps of 1 day each) is closer to
35
the reality, the current pumping scheme seems to lead to the
pollution of wells 2, 3 and 4 by 3, 5 and 2 plume particles
respectively. On the other hand, discretization “a” (Fig. 4a) fails to
even predict pollution of well 2, while combination “b” (Fig. 4b)
underestimates it, failing to simulate all pollution particles that
reach well 2. The combination of Fig. 4 (100 DTs of 10 d each)
seems to be a fair approximation of the pollutant and seawater
particles' transport, managing to predict the wells polluted,
slightly overestimating the arrival time. In other words, the dis-
cretization implemented in this simulations series is an attempt to
balance accuracy and computational load.
As far as pollution of a well is concerned, a pollutant particle P is
assumed to pollute a well W during a certain DT, if and only if the
line segment that simulates the displacement of P during DT in-
tersects the “pollution arrival” zone of well W (Fig. 5).
The most realistic assumption regarding the radius of a well's W
“pollution arrival” zone (rw) is to be equal to the radius of the
aquifer's cylindrical volume (Rw), which contains the water pum-
ped by W during one DT:
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
QW $103 $Dt
RW ¼ (4)
p$n$a
where QW is well's W flow rate, Dt the duration of a timestep in
days, n the effective porosity and a the aquifer's thickness.
36 Y.N. Kontos, K.L. Katsifarakis / Energy 136 (2017) 32e44

Fig. 4. Influence of the temporal discretization of PTM on the produced trajectories in a simplified flow field.

In all graphical results presented in this paper, the wells are indi-
cated by their “pollution arrival” zones, hence the different radii drawn.
The internal and external boundaries of the field are discretized
into 50 boundary elements (Fig. 3). The discretization is rather poor,
in order to understand the BEM accuracy limits and test it to the
limit. In realistic simulations the discretization should be at least 5
times finer, so that the additional wells' permitted zones (Fig. 1) can
cover the majority of the actual field.
3.2. Genetic Algorithm's basic parameters and chromosome
structure
As far as the optimization tool is concerned, simple binary GAs
(including elitist approach) are used. Each chromosome (potential
problem solution) consists of 82-digits (string length SL ¼ 82) and
represents the coordinates of the new wells and the flow rates of all
wells (Fig. 6).
 Y.N. Kontos, K.L. Katsifarakis / Energy 136 (2017) 32e44
Fig. 5. Simulation of pollution of a well.
The genetic operators used are selection (following the
tournament procedure), crossover and mutation (or antimeta-
thesis, a mutation-like operator) [23]. Based on earlier research
[11,12] of similar problems the following GA parameter values
are used: population size PS ¼ 60, number of generations
NG ¼ 1500, selection constant KK ¼ 3 and crossover probability
CRP ¼ 0.40. Especially concerning mutation probability MP,
previous research [11,12] suggested that the use of values
around 2/SL to 2.5/SL (SL ¼ chromosome length), rather than
the empirically proposed 1/SL, is more efficient in directing the
algorithm to the global optimum. Hence, the value MP ¼ 0.026
is used.
3.3. Objective function
Optimization of the pumping scheme implies minimization
of the evaluation function (fitness value FV of proposed solu-
tions) which represents the total solution cost. FV is actually the
sum of the following cost items: a) pumping cost (VB1), b) pipe
network cost (VB2) and c) pumped polluted water treatment
(remediation) cost. Items a and c are mainly energy costs.
Moreover, a penalty function (Penalty) is included in the eval-
uation function (added to the sum of all cost items) to reduce a
chromosome's fitness, in case the respective solution violates
the constraints.
The minimization problem can be outlined as follows:
Find Qi, Xj, Yj, i ¼ 1,5, j ¼ 4,5 so that
FV ¼ VB1 þ VB2 þ VB3 þ Penalty ¼ Min
Constraints:
Q1þQ2þQ3 ¼ 250 l/s
10 l/s  Q4,Q5  120 l/s
(Xi,Yi), i ¼ 4,5 ε Permitted Zones (Fig. 2)
Penalty if pollution particles reach an existing well (see par.
3.3.2)
Penalty if seawater particles reach any well (version A1, see
par. 3.3.2)
Penalty if there is seawater inflow to aquifer (version A2, see
par. 3.3.2)
Fig. 6. Typical 82-digit binary chromosome representing a proposed solution of the studied cost minimization problem.
37
In the above equations, wells 1 to 3 are the existing wells and 4,
5 are the additional ones, Qi, Xi, Yi are the flow rates and co-
ordinates of wells, respectively, FV is the chromosome or solution
fitness value, which is equal to the total management cost in case of
a penalty-free solution, VB1 is the pumping cost, VB2 is the pipe
network amortization cost and VB3 is the annual pumped polluted
water treatment cost. The permitted zones are shown in Fig. 2.
3.3.1. Pumping cost (VB1)
VB1 is the annual key cost item in groundwater management
problems [11,12,24,25]. It is calculated as:
X
Nw
VB1 ¼ A$ Qi $Dsi (5)
i¼1
where, A is a coefficient, that depends on the density of pumped
fluid, the electricity cost per kWh, the pump efficiency and the
pumping duration (here A ¼ 6.48), NW is the total number of wells
and Dsi is the hydraulic head difference between well i and the
maximum hydraulic head (h ¼ 50 m).
3.3.2. Penalty function (penalty)
Whenever a proposed solution entails a constraint violation,
a penalty is imposed to its FV. The violation can either be
pollution of existing wells by plume particles or salinization of
any well by seawater particles (problem version A1), or even
seawater inflow from the coastline (problem version A2). The
Penalty function depends both on the number of violated con-
straints (number of plume particles polluting existing wells and
number of seawater particles polluting any well in problem
version A1 or number of coastline boundary elements exhibit-
ing seawater inflow in problem version A2) and on the magni-
tude of the violation (number of timestep during which a plume
particle pollutes an existing well and number of timestep dur-
ing which a seawater particle pollutes any well in problem
version A1 or magnitude of seawater inflow per coastline
boundary element in version A2). Penalty value in problem
version A1 is given by:
38 Y.N. Kontos, K.L. Katsifarakis / Energy 136 (2017) 32e44

"
plume 
NX  NX techniques vary from linear programming [27], to non-linear pro-
ðTP  ti Þ coast

PenaltyА1 ¼ PC þ PV $ $100 þ C$PC gramming [28], dynamic programming [29] and enumeration
i¼1
TP j¼1 techniques [30], while novel heuristic artificial intelligence
  #
Seaj TP  tj methods (such as neural networks [31], ant algorithm [32] and
þ $C$PV $ $100 (6) genetic algorithms [33]) are also used.
Qj TP
Nonetheless, use of a computationally demanding optimization
technique inside another one's objective function is rather unre-
where, Nplume is the number of plume particles polluting an
alistic. Thus, minimum VB2 is indirectly approximated through
existing well, Ncoast is the number of seawater particles polluting a
minimization of the total length of the network, taking indirectly
well, PC is the constant part of Penalty function, PV is the coefficient
into account pipes' diameter, which are defined by each pipe's flow
of the variable part of the function, TP is the total number of DTs
rate. That way, the network optimization problem has been trans-
(here TP ¼ 100), ti is the DT during which plume particle i pollutes
formed into a minimum spanning tree problem, also solvable by
an existing well, tj is the DT during which seawater particle j is
modern optimization techniques [34]. If the nodes can be internal
pumped by an existing well, C is a coefficient of PC and PV in the
field points, other than the wells themselves, the problem is solved
seawater related part of the Penalty function expressing the relative
with Steiner minimal trees [35]. In the current optimization
weight between pollution and salinization of a well, Seaj is the
application version, the simplest minimal tree problem case is
inflow (in l/s) of seawater represented by particle j and Qj is the
studied, where only the wells themselves constitute possible node
flow rate of well that pumped particle j. Seaj is given by:
locations. The technique used in this paper [36], has already been
Intrk applied in previous similar optimization research [10e12,37,38].
Seaj ¼ (7) Hence, pipe network optimization in this paper refers to the
NppBE
calculation of the best location of the nodes of a network of pipes
that carry pumped water from the additional wells to a reservoir or
where, Intrk is the inflow (in l/s) of seawater from coastline
treatment unit (here XR ¼ 1400 m, YR ¼ 2500 m) with minimum
boundary element k and NppBE is the number of particles per
cost (VB2). VB2 actually represents the pipe network amortization
coastline boundary element (here NppBE ¼ 3).
cost, directly proportional to the initial cost. The latter depends on
Intrk is given by:
the total length and diameter of the pipes. Initial construction cost
vhj is assumed 45 V/m and 60 V/m for small and large pipe diameters,
Іntrk ¼ Tk $ $Lj (8) respectively. The threshold is set at Q ¼ 50 l/s, since pipe diameter is
vn
selected according to the flow rate Q. For an amortization period of
where, Tk is the transmissivity of zone k to which coastline's 10 years and an annual interest rate of 5%, the annual amortization
boundary element j belongs and Lj is the element's length. cost per meter Aak is 6 V/m and 8 V/m for small and large pipe
In problem version A2, Penalty is given by: diameters, respectively, calculated by:
 
plume 
NX  NXelem r$ð1 þ rÞn
ðTP  ti Þ Aai ¼ Ci $ (10)
PenaltyА2 ¼ PC þ PV $ $100 þ C$PC ð1 þ rÞn  1
i¼1
TP j¼1

þ C$PV $Intrj (9) where, i is the type of pipe (1 for small or 2 large diameter), Ci the
initial network cost for pipe type i, r the annual interest rate and n
where, Nelem is the number of boundary elements through which the amortization period.
seawater inflow occurs. VB2 is thus calculated by:
“Optimal” solutions entail Penalty ¼ 0, hence FV represents the XNadw
total cost. The calculation of PC and PV follows the minimum pen- VB2 ¼ i¼1
Aak $Li (11)
alty rule [26], which states that the fittest penalty function is the
lower one that can consecutively provide penalty-free optimal so- where, Nadw is the number of additional wells, Aak is either 6 V/m or
lutions. It is a simple rule, but difficult to implement, since the 8 V/m and Li is the length of the pipe that carries water away from
boundaries between acceptable and unacceptable solutions in the well i.
search field are originally unknown. The simplest but highly time-
consuming way is implementing an extended series of tests of
3.3.4. Pumped polluted water treatment cost (VB3)
increasing Penalty parameters' values. The PC/PV ratio is set to 10/1
VB3 is calculated similarly to the Penalty function, depending
in all scenarios, trying to blend the impact of straight forward, blind
both on the pumped pollutant's mass (number of pollution particles
degradation of a solution just for violating the constraints with the
arriving at additional wells) and the pollutant's “toxicity” (inversely
more sophisticated variable part of the penalty that is proportional
proportional to a particle's arrival time at an additional well for non-
to the magnitude of the violation.
conservative pollutants), except for the fact that the variable part
depends on the respective additional well's flow rate, too. VB3,
which refers only to pollution plume particles and not seawater
3.3.3. Pipe network cost (VB2)
particles, is given for both problem versions (A1 and A2) by:
Pipe network optimization is an extremely difficult optimization
plume  
process that theoretically leads to a loop-less tree-shaped network NX
(tree or branching network). The node location entails different ðTP  ti Þ
VB3 ¼ VC þ VV $Qi $ $100 (12)
flow rate for every pipe section, hence different diameter and i¼1
TP
market value, but also transport and assembly costs. One of the
most common modelling tools of these networks is graph theory. where Nplume is the number of plume particles polluting additional
Since Euler's solution of the seven bridges of Ko € nigsberg in 1736, wells, VC is the constant part of VB3 representing standard
actually laying the foundation of graph theory, there is a great (installation) costs and VV is the coefficient of the variable part of
amount of research in the field. Network design optimization VB3 representing operational costs. The number of remaining
 Y.N. Kontos, K.L. Katsifarakis / Energy 136 (2017) 32e44 39

Table 1 timesteps (TPti) after a particle pollutes an additional well is

Algebraic presentation of the optimal solutions for all four problem versions. multiplied by 100/TP so that VB3 absolute value is not influenced by
Result Problem version the temporal discretization of the study period. That way, the var-
A1low A1hi A2low A2hi
iable part of VB3 can vary from 0 to 100∙VV∙Qi regardless of the
temporal discretization and the results of different series of algo-
FV 488244 665589 92967528 45280745
rithm runs with various discretizations can be directly compared.
VB1 (V/yr) 468188 499115 565284 564918
VB2 (V/yr) 5280 8818 5312 5529 The values of VC and VV can vary, depending on the pollutant.
VB3 (V/yr) 14776 157657 27791 313972 High values imply a pollutant demanding high water treatment cost
Penalty 0 0 92369140 44396325 (“expensive pollutant”), while low values imply low treatment cost
Q1 (l/s) 25.229 21.505 0 0
(“cheap pollutant”). In this paper, for each version of the problem the
Q2 (l/s) 2.294 0 0 0
Q3 (l/s) 222.477 228.495 250 250 cases of both a “cheap” (versions A1low and A2low) and an “expen-
Q4 (l/s) 10.866 12.598 19.528 10.866 sive” pollutant are studied (versions A1hi and A2hi). The first is
Q5 (l/s) 10.866 12.598 0 10.866 described by values VC ¼ 10, VV ¼ 1, while the latter by VC ¼ 200,
X4 (m) 761 671 767 763 VV ¼ 20. All versions of the theoretical problem share a common VC/
Y4 (m) 1900 1748 1881 1932
VV ratio equal to 10/1, similarly to the PC/PV ratio in the Penalty
X5 (m) 764 962 1289 767
Y5 (m) 1917 1442 633 1864 function, in an attempt to scale the treatment unit standard cost to
operational cost. If these values exceed a certain threshold, the

Fig. 7. Optimal solution for problem version A1low: a) general aspect, b) zoom to plume.
40 Y.N. Kontos, K.L. Katsifarakis / Energy 136 (2017) 32e44
minimization process can lead to solutions that avoid pollution of the
additional wells (HC). The values of these parameters that represent
the type of pollutant, combined with the aquifer's geometrical and
hydraulic characteristics, will be the crucial factor for the algorithm
to direct the search for optimal solutions towards PAT or HC strate-
gies or even a combination of the two. Obviously, assigning
extremely high values to VC and VV on purpose is also an indirect but
very efficient way to exclude additional wells' pollution from the
proposed solutions and search only in the HC tank of solutions.
4. Results-conclusions
For every problem version, 5 runs of the GA are implemented,
lasting 5 h in a Intel(R) Core(TM) i7-4510U CPU @2.00 GHz pc. The
optimal solutions for all four theoretical problem versions are
algebraically presented in Table 1.
4.1. Problem version A1low - well salinization criterion & “cheap”
pollutant
Fig. 7 graphically presents the optimal proposed solution for
version A1low, where the 2 additional wells are located inside the
Fig. 8. Optimal solution for problem version A1hi.
plume, almost fully pumping it. The additional wells are located
really close together and relatively close to the treatment unit
exhibiting minimum pipe network length and thus, VB2 cost. They
both pump a flow rate of just about the minimum possible each
(z10 l/s), while the single plume particle escaping from their in-
fluence gets close but does not pollute existing well 1. On the other
front, that of sea intrusion, given the fact that A1 problem versions
define salinization as the pumping of sea water by any well, the
solution is considered acceptable, since seawater particles entering
the flow field get close but are not being pumped by the closest to
the coastline existing wells 1 and 2. This is achieved through: a)
placing the additional wells far away from the coastline and b)
proper distribution of the total required flow rate of 250 l/s among
the existing wells. Well 3, which is the most distant from the
coastline pumps nearly 90% of the total flow rate, while well 1 only
10%, leaving well 2 a mere 2.294 l/s. In A1low version, the algorithm
proposes a PAT solution as the optimal strategy as expected, since
the “cheap” pollutant proves to be cheaper to treat than be
manipulated through more costly HC solutions that would require
larger additional well flow rates. The annual cost for the solution is
around 488244 V/yr, with VB1's share at 95.89%, VB2's at only 1.08%
and VB3's at only 3.03%.
 Y.N. Kontos, K.L. Katsifarakis / Energy 136 (2017) 32e44 41

The unrealistic trajectories of the seawater particles originating
from coastline's boundary element 20 are due to their proximity to
the adjacent zones' 1e2 internal boundary and are a direct result of
the poor discretization of the boundary. A greater number of
boundary elements simulating the internal and external boundaries
would minimize these intrinsic BEM-related accuracy problems.
4.2. Problem version A1hi - well salinization criterion & “expensive”
pollutant
Fig. 8 graphically presents the optimal proposed solution for
A1hi, where 2 additional wells, located between the plume and its
closest existing well (3), share tasks: well 5 fully pumps the
pollutant, while well 4 has already slightly averted the course of
some of the plume spread, delaying its arrival to well 5, thus
weakening its “toxicity” and minimizing total treatment cost VB3.
The additional wells are located away from each other, raising the
VB2 cost by a 67% (3538 V/yr), compared to A1low. The additional
wells both pump 12.598 l/s each, slightly more compared to A1low,
while the scheme really explores the limits of the constraints,
setting up an additional well (5) that fully pumps the pollutant just
a step before it pollutes the most productive abstraction well (3).
This time, in order for the seawater not to reach existing wells,
the algorithm takes well 2 completely out of action, allowing south-
eastern (Zone 2) coastline particles to enter deeper in the aquifer,
but without reaching the precious existing well 3. The south-
western (Zone 1) coastline particles also get close but do not
reach existing well 1 during the 1000 d of study period, which
contributes slightly less fresh water compared to A1low. The dis-
tribution of the total required fresh water flow rate among existing
wells is the key for this optimal proposed strategy, too. Well 3
pumps nearly 91% of the total flow rate, leaving the rest to well 1
and decommissioning well 2. In A1hi version, the algorithm pro-
poses again a combination of PAT and HC solution as the optimal

Fig. 9. “Best” solution of problem version A2low runs: a) general aspect, b) zoom to plume.
42 Y.N. Kontos, K.L. Katsifarakis / Energy 136 (2017) 32e44

strategy. One would expect that the algorithm would prefer solu-
tions that hydraulically control the “expensive” pollutant spread
rather than pump and treat it, but the specific flow field configu-
ration and the high fresh water flow rate required from the existing
wells render a clean HC solution rather impossible, unless the
pollutant were extremely dangerous and its treatment cost pro-
hibitive. The annual cost for the solution is around 665589 V,
36.32% (177345 V) higher than the “cheap” pollutant scenario, with
VB1 contribution at 74.99%, VB2 at only 1.32% and VB3 at a signif-
icant 23.69%. The actual increase in pumping cost VB1 between the
theoretical “cheap”-“expensive” pollutant scenarios is only 6.61%
(30927 V/yr), while pipe network cost VB2 has a 67% increase, but
translating in just 3538 V/yr. The real cause of the total cost FV
breakout is the 966.98% increase of treatment cost VB3, which ac-
counts for an extra 142881 V cost.
4.3. Problem version A2low - sea intrusion criterion & “cheap”
pollutant
Fig. 9 graphically presents the optimal proposed solution for
problem version A2low, where only one additional well, located just
outside the initial plume's boundary, fully pumps the “cheap”
pollutant, with a flow rate around 20 l/s. The solution includes a high
Penalty value, as it is impossible for the studied aquifer to sustain
pumping 250 l/s of groundwater from the existing pumping well
system without additional seawater inflow. Nevertheless, the appli-
cation proposes the least harmful scheme, which at least manages to
keep the wells free of pollution and seawater. In order to achieve that,
operation of wells 1 and 2 is discontinued and well 3 alone provides
the total required fresh water flow rate. South-eastern (Zone 2)
coastline particles enter deep in the aquifer but never reach the well 3

Fig. 10. “Best” solution of problem version A2hi runs: a) general aspect, b) zoom to plume.
 Y.N. Kontos, K.L. Katsifarakis / Energy 136 (2017) 32e44
Table 2
Optimal flow rates for the unconstrained case, groundwater levels and their differences for the unconstrained case (ignoring pollution and salinization) and the undisturbed
flow field (no pumping).
Well (i) Qi si s1si
1 96.949 47.129 e 1.378
2 119.094 45.751 1.378
3 33.957 41.456 5.673
during the study period. In A2low version, the algorithm proposes a
PAT solution as the least harmful strategy, as expected after the A1low
results. If one ignored the Penalty imposed and tried to compare the
results with the ones from A1low, the total solution cost would be
598387 V/yr, 27.81% (130199 V/yr) higher. That is due to a 20.74%
(97096 V/yr) increase in VB1, 0.61% (32 V/yr) in VB2 and an 88.08%
(13015 V/yr) increase in VB3 compared to A1low.
4.4. Problem version A2hi - sea intrusion criterion & “expensive”
pollutant
Fig. 10 graphically presents the optimal proposed solution for
problem version A2hi, where two additional wells, one inside the
plume and the other just outside it, fully pump the “expensive”
pollutant, with flow rates around 11 l/s, near the minimum allow-
able. This solution, just like A2low, includes a high Penalty value, for
the same reasons. Still, the algorithm proposes the least harmful
scheme, which keeps well 3, the only operating at full rate (250 l/s)
existing well, free of pollution and seawater. The proposed least
harmful solution is also based on PAT technique. Ignoring the
Penalty, the total solution cost would be 884419 V/yr, 32.88%
(218830 V/yr) higher than A1hi. Despite the fact that there is a
37.30% (3289 V/yr) decrease in VB2, the increase with regard to
A1hi is due to a 13.18% (65803 V/yr) increase in VB1 and a 99.15%
(156315 V/yr) increase in VB3.
4.5. Accuracy of the optimization process
In order to check the accuracy and the efficiency of the proposed
optimization process, an optimality criterion for pumping cost
minimization has been used, which is derived analytically [39].
“OptiManage” was used to calculate the optimal distribution of the
total required flowrate (250 l/s) to the 3 existing wells for the un-
constrained case (ignoring pollution and salinization). Ground-
water levels were recorded, too. Results are shown in Table 2,
together with sui values, namely the groundwater level at the lo-
cations of the wells, if no pumping occurs. It can be seen that the
derived solution fulfills very satisfactorily the optimality criterion,
namely the differences between si values are half of the differences
between respective sui values. Actually:
ðs1  s2 Þ=ðsu1  su2 Þ ¼ 0:517; ðs1  s3 Þ=ðsu1  su3 Þ
¼ 0:505 and ðs2  s3 Þ=ðsu2  su3 Þ
¼ 0:501:
5. Conclusions
The method of GAs is a highly effective optimization tool that
can successfully address groundwater management problems. The
application created, based on the method, “Optimanage”, proves to
be a powerful user-friendly tool that solves the general problem of
optimal management of a coastal aquifer with pollution or/and
salinization problems, using GAs, PTM and BEM. Moreover, it can
graphically represent the proposed pumping schemes and
43
s2si s3si sui su1sui su2sui su3sui
5.673 5.657 e 2.664 11.229
e 4.295 8.321 2.664 e 8.565
4.295 e 16.886 11.229 8.565 e
pollutant/seawater transport through scaled maps and flow videos
in an automated fashion.
Supplementary video related to this article can be found at
http://dx.doi.org/10.1016/j.energy.2016.10.035.
The automation of the calculation procedure through “Optiman-
age” enables users-researchers to produce and process a large number
of ‘optimal’ solutions for a specific problem, so that they can discover,
identify and categorize many different strategies and their versions
available (PAT, HC or combination of the two) in their algebraically
optimal form. That could lead to the creation of a group of different
strategies, for a specific aquifer under pollution or/and sea intrusion
problem. This qualitative optimization, given the uncertainties
deriving from the many simplifying assumptions and accuracy com-
promises of the 2D problem, seems to be of far greater significance
compared to the strictly algebraic minimization of the total cost. The
users' important role is to adjust the algorithm and select the fittest
GA, PTM and BEM parameters and to find the largest number of
different flow profiles in their algebraic optimal version, in order to
finally select the most appropriate strategy version for the specific
problem, taking into account not only the criteria and constraints
included in the objective function, but possibly additional ones, too.
In order to keep total computational volume under control,
some simplification of the evaluation process (in flow and mass
transport models) is required. Such simplifications may raise
questions regarding the accuracy of the overall results. For this
reason, every proposed solution should be logically checked.
Moreover, it should be checked with a finer temporal discretization
of the study period and with a finer discretization of the internal-
external boundaries, in order to discard unrealistic solutions with
unpractical features (e.g. additional wells of low flow rates sup-
posedly pumping a disproportional amount of pollution or con-
struction of additional wells unrealistically close to each other).
Future research could include the optional use of additional
recharge wells instead of, or together with, abstraction wells, to
investigate whether the total cost can be further decreased and to
point out which cases are most likely to benefit from it. Moreover, a
series of brief additional well's pumping sudden failure incidents
can be studied, in order to stress test the existing solution profiles
and conclude on their fail-safe attributes, filtering the most pre-
carious of them. This research direction could easily lead to the
addition of targeted intermissions in the operation of the additional
wells and ideally the inclusion of intermittent pumping in the
objective function of the genetic algorithm. Last but not least,
“Optimanage” could also be programmed to solve the water supply
(WS) problem in complex coastal aquifers just like the studied one,
where optimization will mean maximization of fresh pumped
water, without further sea intrusion.
Acknowledgements
Presented research is part of the action “Research & Technology
Development Innovation Projects - AgroETAK”, MIS 453350, in the
framework of O.P. “Human Resources Development”, co-funded by
ESF and National funds (NSRF 2007e2014), coordinated by the
Hellenic Agricultural Organization “DEMETER” (Institute of Indus-
trial and Forage Plants, Scientific supervisor: Dr C. D. Tsadilas).
44 Y.N. Kontos, K.L. Katsifarakis / Energy 136 (2017) 32e44

References [19] Bayer P, Finkel M. Evolutionary algorithms for the optimization of advective
control of contaminated aquifer zones. Water Resour Res 2004;40(6):
W06506.
[1] Shang Yizi, Wang Jianhua, Liu Jiahong, Jiang Dong, Zhai Jiaqi, Jiang Shan.
[20] Katsifarakis KL, Petala Z. Combining genetic algorithms and boundary ele-
Suitability analysis of China's energy development strategy in the context of
ments to optimize coastal aquifers' management. J Hydrol 2006;327(1e2):
water resource management. Energy 2016;96:286e93.
200e7.
[2] Venkatesh G, Brattebø Helge. Energy consumption, costs and environmental
[21] Ingham DB, Heggs PJ, Manzoor M. The numerical solution of plane potential
impacts for urban water cycle services: case study of Oslo (Norway). Energy
problems by improved boundary integral equation methods. J Comput Phys
2011;36:792e800.
1981;42(1):77e98.
[3] Mayer AS, Kelley CT, Miller CT. Optimal design for problems involving flow
[22] Latinopoulos P, Katsifarakis KA. Boundary element and particle tracking
and transport phenomena in saturated subsurface systems. Adv Water Resour
model for advective transport in zoned aquifers. J Hydrol 1991;124(1e2):
2002;25(8e12):1233e56.
159e76.
[4] Medina Jr MA. Modeling ground water contamination and surface-subsurface
[23] Katsifarakis KL, Karpouzos DK. Minimization of pumping cost in zoned aqui-
interactions. In: Singh VP, Yadava RN, editors. Ground water pollution: proc.
fers by means of genetic algorithms. In: Katsifarakis KL, Korfiatis GP,
Of the international conference on water and environment (WE-2003). Bho-
Mylopoulos YA, Demetracopoulos AC, editors. Proc. Of the international
pal, India: Allied Publishers; 2003. p. 401e18.
conference “protection and restoration of the environment ІV”, Sani, Halki-
[5] Mulligan AE, Ahlfeld DP. Advective control of groundwater contaminant
diki, Greece; 1998. p. 61e8.
plumes: model development and comparison to hydraulic control. Water
[24] Sidiropoulos E, Tolikas P. Well locations and constraint handling in ground-
Resour Res 1999;35(8):2285e94.
water pumping cost minimization via genetic algorithms. Water Air Soil
[6] Guan J, Aral MM. Optimal remediation with well locations and pumping rates
Pollut Focus 2004;4(4e5):227e39.
selected as continuous decision variables. J Hydrol 1999;221(1e2):20e42.
[25] Kalwij IM, Peralta RC. Non-adaptive and adaptive hybrid approaches for
[7] Gorelick SM. Sensitivity analysis of optimal groundwater contaminant capture
enhancing water quality management. J Hydrol 2008;358(3e4):182e92.
curves: spatial variability and robust solutions. In: Proc. Of the national water
[26] Le Riche R, Knopf-Lenoir C, Haftka RT. A segregated genetic algorithm for
well association conference: solving groundwater problems with models.
constrained structural optimization. In: Proc. Of the 6th international con-
Denver, Colorado: National Water Well Association; 1987. p. 133e46.
ference on genetic algorithms. Morgan Kaufmann Publishers Inc.; 1995.
[8] Bayer P, Finkel M, Teutsch G. Reliability of hydraulic performance and cost
p. 558e65.
estimates of barrier-supported pump-and-treat systems in heterogeneous
[27] Sonak VV, Bhave PR. Global optimum tree solution for single-source looped
aquifers. In: KKaZ Hrkal, editor. Proc calibration and reliability in groundwater
water distribution networks subjected to a single loading pattern. Water
modelling: a few steps closer to reality. Prague: Czech Republic; 2002.
Resour Res 1993;29(7):2437e43.
p. 331e8. IAHS Publ. no 227.
[28] Brooke A, Kendrick D, Meeraus A. GAMS: a user's guide. Redwood City, CA,
[9] Kontos YN, Katirtzidou MI, Kizeridou MA, Katsifarakis KL. Optimal manage-
USA: The Scientific Press; 1998.
ment of a polluted fractured aquifer, using genetic algorithms. In: Christo-
[29] Walters GA. Dynamic programming approach to the optimal design of tree-
doulou, Stamou AI, editors. Proc. 6th international symposium on
like networks. In: International conference on optimization techniques and
environmental hydraulics, vol. 2. Athens, Greece: Taylor & Francis Group;
applications, Singapour, Singapore; 1987. p. 487e96.
2010. p. 685e90.
[30] Gessler J. Pipe network optimization by enumeration. Conference pipe
[10] Kontos YN, Katsifarakis KL. Optimization of pumping scheme in a polluted frac-
network optimization by enumeration, New York. ASCE, p. 572e581.
tured aquifer using genetic algorithms. In: Proc. International conf. “Protection
[31] Ze NT. Optimal lay-out of natural gas pipeline network. In: 23rd world gas
and restoration of the environment XІ”, Thessaloniki, Greece; 2012. p. 375e84.
conference. Amsterdam, Netherlands; 2006.
[11] Kontos YN, Katsifarakis KL. Optimization of management of polluted fractured
[32] Afshar MH. Application of ant algorithm to pipe network optimization. Iran J
aquifers using genetic algorithms. Eur Water 2012b;40:31e42.
Sci Technol Trans B Eng 2007;31(B5):487e500.
[12] Kontos YN. Optimal management of fractured coastal aquifers with pollution
[33] Shau H-M, Lin B-L, Huang W-C. Genetic algorithms for design of pipe network
problems (in Greek) [PhD]. Thessaloniki, Greece: Aristotle Univ. of Thessalo-
systems. J Mar Sci Technol 2005;13(2):116e24.
niki; 2013.
[34] Neumann F, Witt C. Ant colony optimization and the minimum spanning tree
[13] Glover EW. Containment of contaminated groundwater - an overview. Con-
problem. In: Maniezzo V, Battiti R, Watson J-P, editors. Learning and intelli-
ference containment of contaminated groundwater - an overview, Wor-
gent optimization. Berlin Heidelberg: Springer; 2008. p. 153e66.
thington, OH. p. 17e22.
[35] Gilbert E, Pollak H. Steiner minimal trees. SIAM J Appl Math 1968;16(1):1e29.
[14] Ahlfeld DP, Page RH, Pinder GF. Optimal ground-water remediation methods
[36] Prim RC. Shortest connection networks and some generalizations. Bell Syst
applied to a superfund site: from formulation to implementation. Ground
Tech J 1957;36:1389e401.
Water 1995;33(1):58e70.
[37] Katsifarakis KL, Tselepidou K, Konstantakos N, Stamati D, Mpletsa E,
[15] Sawyer C, Lin Y. Mixed-integer chance-constrained models for ground-water
Tzanakis I. Optimization of low-enthalpy geothermal heating schemes by
remediation. J Water Resour Plan Manag 1998;124(5):285e94.
means of genetic algorithms. J Sustain Dev Plan 2006;1(4):429e42.
[16] Huang C, Mayer AS. Pump-and-Treat optimization using well locations and
[38] Tselepidou K, Katsifarakis KL. Optimization of the exploitation system of a low
pumping rates as decision variables. Water Resour Res 1997;33(5):1001e12.
enthalpy geothermal aquifer with zones of different transmissivities and
[17] Yoon J, Shoemaker C. Comparison of optimization methods for ground-water
temperatures. Renew Energy 2010;35(7):1408e13.
bioremediation. J Water Resour Plan Manag 1999;125(1):54e63.
[39] Katsifarakis KL, Tselepidou K. Pumping cost minimization in aquifers with
[18] Erickson M, Mayer A, Horn J. Multi-objective optimal design of groundwater
regional flow and two zones of different transmissivities. J Hydrol
remediation systems: application of the niched pareto genetic algorithm
2009;377(1e2):106e11.
(NPGA). Adv Water Resour 2002;25(1):51e65.