Computers and Electronics in Agriculture 125 (2016) 74–80

Contents lists available at ScienceDirect

Computers and Electronics in Agriculture

journal homepage: www.elsevier.com/locate/compag

Original papers

A new predictive model for the filtered volume and outlet parameters
in micro-irrigation sand filters fed with effluents using the hybrid
PSO–SVM-based approach
P.J. García Nieto a,⇑, E. García-Gonzalo a, G. Arbat b, M. Duran-Ros b, F. Ramírez de Cartagena b,
J. Puig-Bargués b
a
Department of Mathematics, Faculty of Sciences, University of Oviedo, 33007 Oviedo, Spain
b
Department of Chemical and Agricultural Engineering and Technology, University of Girona, Carrer de Maria Aurèlia Capmany 61, 17071 Girona, Spain

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

Article history: Filtration is a key operation in micro-irrigation for removing the particles carried by water that could clog
Received 8 December 2015 drip emitters. Currently, there are not sufficiently accurate models available to predict the filtered vol-
Received in revised form 29 April 2016 ume and outlet parameters for the sand filters used in micro-irrigation systems. The aim of this study
Accepted 29 April 2016
was to obtain a predictive model able to perform an early detection of the filtered volume and sand filter
outlet values of dissolved oxygen (DO) and turbidity, both related to emitter clogging risks. This study
presents a novel hybrid algorithm, based on support vector machines (SVMs) in combination with the
Keywords:
particle swarm optimization (PSO) technique, for predicting the main filtration operation parameters
Support vector machines (SVMs)
Particle swarm optimization (PSO)
from data corresponding to 769 experimental filtration cycles in a sand filter operating with effluent.
Drip irrigation This optimization technique involves kernel parameter setting in the SVM training procedure, which sig-
Clogging nificantly influences the regression accuracy. To this end, the most important physical–chemical param-
Regression analysis eters of this process are monitored and analyzed: effective sand media size, head loss across the filter and
filter inlet values of dissolved oxygen (DO), turbidity, electrical conductivity (Ec), pH and water temper-
ature. The results of the present study are two-fold. In the first place, the significance of each physical–
chemical variables on the filtration is presented through the model. Secondly, a model for forecasting the
filtered volume and sand filter outlet parameters is obtained with success. Indeed, regression with opti-
mal hyperparameters was performed and coefficients of determination equal to 0.74 for outlet turbidity,
0.82 for filtered volume and 0.97 for outlet dissolved oxygen were obtained when this hybrid PSO–SVM-
based model was applied to the experimental dataset, respectively. The agreement between experimen-
tal data and the model confirmed the good performance of the latter.
Ó 2016 Elsevier B.V. All rights reserved.

1. Introduction 2012; Tripathi et al., 2014; Wen-Yong et al., 2015). Furthermore,

only few studies to predict sand filter outlet dissolved oxygen val-
Sand media filters are considered the standard for filtration in ues using techniques as artificial neural networks (ANN) and gene
micro-irrigation systems, especially when irrigation waters such expression programming (GEP) (Puig-Bargués et al., 2012; Martí
as biological effluents that present increased emitter clogging haz- et al., 2013) have been carried out. However, and as far as the
ards are used (Trooien and Hills, 2007). Sand filters are commonly authors know, support vector machines (SVMs) approach has not
used for their hardware simplicity and large capacities (Burt, 2010) been applied yet in filtration for micro-irrigation systems. The
but they are more expensive than other filter types (Pujol et al., use of this methodology could help both irrigation engineers and
2011) and their optimal operation needs high technical back- farmers to properly manage sand filters achieving a good micro-
ground (Capra and Scicolone, 2007). The high efficiency achieved irrigation system performance and targeted crop yields.
by sand filters for removing suspended solids is well known SVM models are based on the statistical learning theory and are
(Puig-Bargués et al., 2005; Duran-Ros et al., 2008; Elbana et al., a new class of models that can be used for predicting values from
very different fields (Vapnik, 1998; Cristianini and Shawe-Taylor,
2000; Schölkopf et al., 2000). SVMs can be used for classification
⇑ Corresponding author.
and regression because they possess the ability of being universal
E-mail address: lato@orion.ciencias.uniovi.es (P.J. García Nieto).

http://dx.doi.org/10.1016/j.compag.2016.04.031
0168-1699/Ó 2016 Elsevier B.V. All rights reserved.
 P.J. García Nieto et al. / Computers and Electronics in Agriculture 125 (2016) 74–80 75

approximators of any multivariate function to any desired degree ple of granular media of 0.32 and 0.47 mm) were tested. A MP-400-
of accuracy (Negoita and Reusch, 2005). The statistical learning CB (Comaquinsa, Llinars del Vallès, Spain) electromagnetic flow
theory and structural risk minimization are the theoretical founda- meter measured water flow, two MBS 4010 (Danfoss, Nordborg,
tions for the learning algorithms of SVMs (Hastie et al., 2003; Denmark) pressure transmitter allowed to determine pressure loss
Hansen and Wang, 2005; Steinwart and Christmann, 2008; Li across the filter and several Endress + Hauser (Nesselwang, Ger-
et al., 2008). many) sensors (Orbisint CPS11D, OxyMax W COS61, TurbiMax W
In order to carry out the optimization mechanism correspond- CUS31 and ConduMax W CLS21) and transmitters (CPM253,
ing to the kernel optimal hyperparameters setting in the SVM COM253, CUM253 and CLM253) were used for measuring pH
training, the particle swarm optimization (PSO) technique can be and temperature, dissolved oxygen (DO), turbidity and electrical
used. Particle swarm optimization (PSO) is one of the oldest swarm conductivity (Ec) at the filter inlet, respectively. At filter outlet, only
intelligence (SI) based bio-inspired algorithms and it was intro- DO and turbidity were analyzed with the same sensors and trans-
duced by Kennedy and Eberhart (1995). The PSO technique is a mitters installed at the filter inlet for both parameters. All these
population-based search algorithm inspired in the simulation of data were collected every minute in a previously developed super-
the bird flocking (Eberhart et al., 2001; Clerc, 2006; Olsson, visory control and data acquisition system (SCADA) (Duran-Ros
2011). Similar to other evolutionary computation SI-based algo- et al., 2008) that was modified for acquiring water quality param-
rithms such as the ant colony optimization (Dorigo and Stützle, eters. Further details of the experiment are described in Puig-
2004; Panigrahi et al., 2011), and artificial bee colony (ABC) tech- Bargués et al. (2012).
nique (Karaboga and Akay, 2009; Karaboga and Gorkemli, 2014), A first task in the model development was the selection of the
PSO exploits the model of social sharing of information (Fister Jr. input and output variables. This variable selection was based on
et al., 2013; Fister et al., 2015). For the above-mentioned purpose, the previous knowledge (Puig-Bargués et al., 2005) of the sand fil-
hybrid PSO optimized SVM models (Simon, 2013; Yang et al., 2013) tration process resulting in the following group of input variables:
were used in this study to predict the filtration performance from effective sand media size; head loss; flow rate; water temperature
operation parameters in micro-irrigation sand filters fed with efflu- at the filter outlet; inlet pH; inlet electrical conductivity; inlet dis-
ents. According to previous research, the SVM technique has been solved oxygen; and inlet turbidity, which is related with sus-
proved to be an effective tool to predict natural parameters, being pended solids concentration and has been adopted as an easy
successfully used in a wide range of environmental fields: forest and reasonably accurate measure of overall water quality. The out-
modeling (Shrestla and Shukla, 2015), solar radiation estimation put variables are three parameters used to evaluate the filtration
(Chen et al., 2013, 2015; Zeng and Qiao, 2013), prediction of the performance in micro-irrigation sand filters: outlet dissolved oxy-
air quality (Ortiz-García et al., 2010), study of water properties gen, outlet turbidity and filtered volume.
(Pal and Goel, 2007; Nikoo and Mahjouri, 2013; Xu et al., 2015), The average, the standard deviation and the range of each
applications in soil physics (Lamorski et al., 2008; Haghverdi parameter are shown in Table 1.
et al., 2014) and so on.
The objective of this study was to evaluate the application of 2.2. Support vector machine (SVM) method
support vector machines (SVMs) approach in combination with
the evolutionary optimization technique known as Particle Swarm The support vector regression (SVR) technique is based on the
Optimization (PSO) to identify the filtered volume and sand filter statistical learning theory and structural risk minimization
outlet values of dissolved oxygen (DO) and turbidity, both related (Vapnik, 1998; Pal and Goel, 2007; Chen et al., 2013; Abbaszadeh
to emitter clogging risks which negatively affect micro-irrigation et al., 2016). Based on SVR, for the given training sample
system performance and therefore crop yields. D ¼ fðxi ; yi Þjxi 2 Rd ; yi 2 R; i ¼ 1; 2; . . . ; Lg; through the nonlinear
mapping function W(x), the sample data x is mapped to another
2. Materials and methods high dimensional feature space. Therefore, a regression function
can be defined in this feature space as follows (Nikoo and
2.1. Experimental dataset Mahjouri, 2013; Zeng and Qiao, 2013; Shrestla and Shukla, 2015):
X
L
The experimental dataset used for the PSO–SVM analysis was y ¼ f ðxÞ ¼ wi Wi ðxÞ þ b ð1Þ
collected from 769 filtration cycles carried out during 1620 h of fil- i¼1
ter operation with a wastewater treatment plant (WWTP) tertiary
effluent. Two parallel sand filters (Regaber, Parets del Vallès, where fwi gLi¼1 are the coefficients of the vector w normal to the
Spain), both with 500 mm inlet internal diameter and with a filtra- maximum-margin hyperplane, fWi ðxÞgLi¼1 is the set of mappings of
tion surface of 1963 cm2, were used. Two effective sand media size input vectors and y is the predicted output value or objective func-
b
(defined as the size of screen opening which retains 90% of a sam- tion. The parameter kwk determines the offset of this hyperplane

Table 1
Set of measured physical–chemical variables used in this study with their mean, standard deviation and range of values (Puig-Bargués et al., 2012).

Parameter Name of the variable Average Standard deviation Range

Effective sand media size (mm) de 0.39 0.08 0.32–0.47
Head loss (kPa) DP 55.7 7.12 26.0–107.0
Flow rate (m3 h1) Q 9.73 1.11 7.16–26.0
Filtered volume (m3) V 16.0 26.7 0.56–267.0
Water temperature (°C) T 23.9 3.81 15.0–29.9
Inlet pH pH 8.39 0.60 7.10–9.38
Inlet electrical conductivity (dS m1) Ec 5.21 0.64 2.33–7.10
Inlet dissolved oxygen (g m3) DOi 2.83 1.02 0.13–5.58
Outlet dissolved oxygen (g m3) DOo 3.10 0.95 0.46–6.13
Inlet turbidity (FTU) Turbi 21.19 16.03 3.13–50.0
Outlet turbidity (FTU) Turbo 4.67 3.76 0.86–45.6
76 P.J. García Nieto et al. / Computers and Electronics in Agriculture 125 (2016) 74–80

The velocity of each particle, i, at iteration k, depends on three

components stated as:

 The previous step velocity term, v ki , affected by the constant

inertia weight, x.
 The cognitive learning term, which is the difference between
k
the particle’s best position so far found (called li , local best)
and the particle current position xki .
 The social learning term, which is the difference between the
global best position found thus far in the entire swarm (called
gk, global best) and the particle’s current position xki .
Fig. 1. Regression with e – insensitive tube.
In this study, the Standard PSO 2011 (Clerc, 2012) has been
used. It contemplates some improvements in the implementation
from the origin along the normal vector w. If the insensitive loss (Eberhart et al., 2001; Olsson, 2011; Clerc, 2006, 2012) and the
function e is adopted (see Fig. 1), the aim of SVR is to look for a f PSO parameters are set to the values:
(x) which can make the difference between the true and the train-
1
ing values less than the given error e. x¼ and c1 ¼ c2 ¼ 0:5 þ ln 2 ð8Þ
Thus, the function f solution can be expressed as the following 2 ln 2
quadratic programming problem (Cristianini and Shawe-Taylor,
2000; Gu et al., 2006; Steinwart and Christmann, 2008; García 2.4. The goodness-of-fit of this approach
Nieto et al., 2012):
The following criteria were considered here (Hastie et al., 2003;
1 XL
Min kwk2 þ C ðni þ ni Þ ð2Þ Wasserman, 2003; Freedman et al., 2007):
ðw;b;nÞ 2
i¼1
 The coefficient of determination R2: A dataset takes observed
subject to
8 9 values ti, each of which has an associated modeled value or pre-
< yi  hw; Wðxi Þi  b 6 e þ ni >
> = dicted value yi. Then, it is possible to define:
P
hw; Wðxi Þi þ b  yi 6 e þ ni i ¼ 1; . . . ; L ð3Þ o SStot ¼ n ðti  tÞ : the total sum of squares, proportional
2
>
: >
;
i¼1
ni ; ni P 0 to the sample variance.
P
o SSerr ¼ ni¼1 ðti  yi Þ2 : the residual sum of squares.
where C is the error penalty parameter or regularization constant and
h,i denotes the ordinary dot product. In order to make sure of the
In the previous sums, t is the mean of the n observed data. Bear-
existence of solutions, it is necessary to relax the conditions on the
ing in mind the above sums, the general definition of the coeffi-
maximum-margin hyperplane introducing slack variables ni and ni
cient of determination is:
(soft margin approach). Additionally, we can define an inner or scalar
product by means of a positive definite function k (kernel function) SSerr
R2  1  ð9Þ
(Cristianini and Shawe-Taylor, 2000; Schölkopf et al., 2000; Shawe- SStot
Taylor and Cristianini, 2004; Hansen and Wang, 2005) as:
 The root mean square error (RMSE): represents the sample
X
kðx; x0 Þ ¼ hWðxÞ; Wðx0 Þi ¼ Wi ðxÞ  Wi ðx0 Þ ð4Þ standard deviation of the differences between predicted values
i and observed values. The RMSE is computed for n different pre-
dictions as (Freedman et al., 2007):
The selection of the kernel function depends on both the prob-
rffiffiffiffiffiffiffiffiffiffi
lem nature and type of data (Li et al., 2008; Wu, 2009; de Cos Juez SSerr
et al., 2010; Ortiz-García et al., 2010). In this study, RBF (radial RMSE ¼ ð10Þ
n
basis function) has been selected as the kernel function
(Kanevski et al., 2004; Shawe-Taylor and Cristianini, 2004):
3. Results
kðxi ; xj Þ ¼ erkxi xj k
2
ð5Þ
The particles xi are vectors that contain the parameters that we
Indeed, it allows to analyze high dimensional data sets appro-
want to tune, that is to say, xi = (Ci, ei, ri) for the RBF kernel. In this
priately (Nikoo and Mahjouri, 2013; Zeng and Qiao, 2013;
study, we had 20 particles. In the first iteration, we initialized them
Shrestla and Shukla, 2015).
randomly. Following the PSO algorithm, as described in Section 2.3.,
the particles for the next iteration were calculated. In each step, the
2.3. The particle swarm optimization (PSO) algorithm
objective function value for the particles was computed. The objec-
tive function value was the cross-validation coefficient of determi-
In the PSO algorithm (Kennedy and Eberhart, 1995; Eberhart
nation for each particle. When the termination criteria were met,
et al., 2001; Clerc, 2006; Olsson, 2011), the parameters, or possible
the global best xi contained the optimized parameters. Fig. 2 shows
set of solutions, are contained in a vector xi, which is called a par-
the flowchart of this new hybrid PSO–SVM-based model developed
ticle of the swarm and represents its position in the search space of
in this paper.
possible solutions. The initial particle position x0i and its velocity v 0i Cross validation was the standard technique used here for find-
are chosen randomly. The algorithm updates the positions and the ing the real coefficient of determination (R2) (Picard and Cook,
velocities of the particles following the equations: 1984). The data set was randomly divided into l disjoint subsets
v kþ1
i ¼ xv ki þ /1 ðgk  xki Þ þ /2 ðIki  xki Þ ð6Þ of equal size, and each subset was used once as a validation set,
whereas the other l  1 subsets were put together to form a train-
xkþ1
i ¼ xki þ v ikþ1 ð7Þ ing set. In the simplest case, the average accuracy of the l validation
 P.J. García Nieto et al. / Computers and Electronics in Agriculture 125 (2016) 74–80 77

2012). The searching in the parameter space were made taking into
account that the SVM algorithm changed its results significantly
when its parameters increased or decreased a power of 10. For
instance, we worked with powers of ten and the searched param-
eters were the exponents, being the search space three-
dimensional. For the RBF kernel, the search space was [2, 2] 
[1, 3]  [10, 0]. That is, C values in [102, 102], r values in
[101, 103] and e values in [1010, 100] were used in the optimiza-
tion phase. The stopping criterion was fulfilled if there was no
improvement in R2 after ten iterations, along with a maximum
number of iterations equal to 500.
Table 2 shows the optimal hyperparameters of the three fitted
SVM-based models found using the particle swarm optimization
(PSO) technique for the outlet dissolved oxygen (DOo),
respectively.
Table 3 shows the coefficient of determinations, root mean
square error and overall index (OI) of model performance (Galavi
et al., 2013) for the three PSO–SVM-based models fitted here for
the outlet dissolved oxygen (DOo), outlet turbidity (Turbo) and fil-
tered volume (V), respectively.

4. Discussion

The PSO–RBF–SVM-based approach is an excellent model for

estimating the outlet dissolved oxygen (DOo) in order to evaluate
the sand filter performance. Indeed, the fitted SVM with RBF kernel
Fig. 2. Flowchart of the new hybrid PSO–SVM-based model.
function has a testing coefficient of determination R2 equal to
0.9714. Therefore, a very good agreement is obtained between our
sets was used as an estimator for the accuracy of the method. The
model and the observed data. Additionally, the RMSE value was also
combination of the hyperparameters with the best performance
very close to zero (see Table 3). The goodness of fit achieved with the
was chosen (Schölkopf et al., 2000; Wasserman, 2003; Hansen
PSO optimized SVM approach was better than those obtained with
and Wang, 2005; Freedman et al., 2007; Li et al., 2008; García
multiple linear regression (MLR) (Puig-Bargués et al., 2012), and
Nieto et al., 2012).
ANN and GEP (Martí et al., 2013). Table 4 shows the weights that
Firstly, a 20% of the samples were selected randomly for testing
correspond to the three fitted models. The weight absolute value
purposes and the remaining 80% of the data was used to build the
is an estimation of the importance of the independent variable
optimized model via cross-validation. As it has been previously
within the model. The higher the value, the more important is the
pointed out, in order to guarantee the prediction ability of the
variable. As it could be anticipated, inlet dissolved oxygen (DOi)
PSO–SVM-based model, an exhaustive 5-fold cross-validation algo-
was the variable that had more importance on DOo. Martí et al.
rithm was used with this last 80% of the data (Picard and Cook,
(2013) obtained with GEP an equation for computing DOo which
1984; Efron and Tibshirani, 1997). Therefore, all the possible vari-
included also DOi, pH, Ec and DP and excluded de and temperature.
ability of PSO–SVM model parameters were evaluated in order to
Although these two variables have, with flow and pH, the smallest
get the optimum point, looking for those parameters that minimize
relative importance, the overall correlation coefficient obtained
the average error. With these optimal hyperparameters, the model
with our approach has been improved regarding previous works.
was built using 80% of the sample and tested with the remaining
20%. The fitness factor was the coefficient of determination (R2).
The regression modeling was performed with SVR-e using the Table 4
Weights in the fitted PSO–SVM-based model, which are related to the variable
LIBSVM library (Chang and Lin, 2011) and the parameters were importance in the model, for the outlet dissolved oxygen (DOo), outlet turbidity
optimized with PSO, using the standard PSO 2011 version (Clerc, (Turbo) and filtered volume (V), respectively.

Table 2 Input variable Weights for DOo Weights for Turbo Weights for V
Optimal hyperparameters of the three fitted SVM-based models with the RBF kernel, DOi 2.1345 0.3518 1.4611
found using the PSO technique for the outlet dissolved oxygen (DOo), outlet turbidity Turbi 0.6314 1.0407 1.9632
(Turbo) and filtered volume (V). DP 0.4252 0.5859 0.3180
Dependent variable Values of optimal hyperparameters Ec 0.2228 0.1891 0.1975
T 0.2177 0.5010 0.9179
DOo C = 6.9839  101, e = 2.0516  109, r = 2.0730  100 pH 0.1765 0.4178 1.8537
Turbo C = 1.4070  100, e = 1.4371  107, r = 4.1153  100 Q 0.1199 0.3057 0.5872
V C = 1.2184  100, e = 1.2933  107, r = 1.5428  101 de 0.0008 0.4113 2.1255

Table 3
Cross-validation and testing coefficient of determination (R2), cross-validation and testing root mean square error (RMSE) and testing overall index (OI) of model performance for
the three hybrid PSO–SVM-based models with the RBF kernel fitted in this study for the outlet dissolved oxygen (DOo), outlet turbidity (Turbo) and filtered volume (V).

Dependent variable Cross-val. R2 Testing R2 Cross-val. RMSE Testing RMSE Testing OI

DOo 0.9275 0.9714 1.7417  101 3.6211  102 2.3353  104
Turbo 0.6782 0.7381 1.1621  101 5.3445  102 1.9186  104
V 0.7936 0.8210 1.1621  101 8.4885  102 3.3067  104
78 P.J. García Nieto et al. / Computers and Electronics in Agriculture 125 (2016) 74–80

The negative weight of turbidity, pressure drop and tempera- PSO–SVM with RBF kernel function had a testing coefficient of
ture (see Table 4) in the fitted PSO–SVM model for the DOo is determination R2 of 0.7381, which was clearly higher than those
explained because these variables are related to microbial growth obtained with MLR (R2 = 0.278) and ANN (R2 = 0.630) (Puig-
within the filter therefore reducing DO concentrations (Elbana Bargués et al., 2012). Furthermore, the RMSE value was very close
et al., 2012). Thus, higher turbidity is related with higher influent to zero for this fitting (see Table 3), which confirms the good per-
pollutant load, which means also that the filtration media will formance of this approach.
retain more microorganisms which will consume more DO. Filter The importance ranking of the eight input variables in order to
pressure loss increases as inlet effluent particles are retained in predict the outlet turbidity (Turbo) in this high nonlinear complex
the filter, yielding more microbial growth and smaller DO just problem can also be inferred from the absolute weight values
before filter backwashing (Elbana et al., 2012). Higher temperature shown in Table 4. In this case, the outlet turbidity was more
causes faster microbial growth but also it holds less oxygen than affected by the inlet equivalent parameter, the inlet turbidity
cold water. (Turbi), as it was to be expected. The higher Turbo at higher head
Furthermore, the PSO–SVM with the RBF kernel function was loss is a bit surprising. Higher filter pressure loss means that filter
also a good model for estimating the outlet turbidity (Turbo) vari- is clogged, having additional particle removal capacity (Elbana
able in order to predict the sand filtration performance. The fitted et al., 2012) but in some cases it could produce solid detachment

Fig. 3. Comparison between the values observed and predicted by using the PSO–SVM-based models for: (a) the outlet dissolved oxygen (DOo) (R2 = 0.9714); (b) the outlet
turbidity (Turbo) (R2 = 0.7381); and (c) filtered volume (V) (R2 = 0.8210).
 P.J. García Nieto et al. / Computers and Electronics in Agriculture 125 (2016) 74–80 79

(Adin and Alon, 1986) or the deformation of biological particles oped to predict the sand filtration performance from the other
which therefore can pass through the filter (Puig-Bargués et al., measured input operation variables, in order to lower costs in
2005). Smaller effective sand media size are related with higher the water quality assessment.
turbidity removals in sand filter, yielding less Turbo (Duran-Ros  Thirdly, a coefficient of determination greater than those
et al., 2008) as the PSO–SVM-based model predicted. On the other obtained with other approaches have been obtained with this
hand, sand filters operating at lower flow can remove more parti- hybrid PSO–SVM-based model when a RBF kernel function
cles and therefore they reduce more turbidity. was applied to the experimental dataset corresponding to the
Additionally, the PSO–RBF–SVM-based model performed very outlet dissolved oxygen (DOo), outlet turbidity (Turbo) and fil-
well for estimating the filtered volume (V), since the model had tered volume (V). Moreover, the predicted results for this model
an R2 equal to 0.8210 which was higher than those obtained with have been proven to be consistent with the historical dataset of
MLR (R2 = 0.447) and ANN (R2 = 0.693) (Puig-Bargués et al., 2012) these three parameters.
and RMSE was close to zero (see Table 3).  Fourthly, the significance order of the input variables involved
Similarly, the importance ranking of the eight input variables in in the prediction of the sand filtration performance was set. This
order to predict the filtered volume (V) is inferred from Table 4. is one of the main findings in this work. Specifically, the opera-
While with DOo and Turbo there was only a predominant variable, tion variable input dissolved oxygen (DOi) could be considered
with the V there were four variables with a weight value greater the most influential parameter in the prediction of the outlet
than 1.4 in absolute value. Higher V were obtained with higher dissolved oxygen (DOo) (dependent variable). It is followed by
de (which yields less small particle removal and, thus, less filter the inlet turbidity (Turbi) and total head loss (DP). Similarly,
clogging), with lower Turbi (which means better water quality the operation variable input turbidity (Turbi) could be consid-
and, thus, less sand media fouling), less pH (acid conditions that ered the most influential parameter in the prediction of the out-
have a negative effect on microbiological growth that cause sand let turbidity (Turbo) (dependent variable), followed by the total
media fouling) and higher DOi (which is related to better water head loss (DP). In this way, the operation variable effective sand
quality). These results agree with other previous research media size (de) could be considered the most influential param-
(Duran-Ros et al., 2008; Elbana et al., 2012). eter in the prediction of the filtered volume (V) (dependent vari-
In summary, this research work was able to predict the outlet able), followed by the input turbidity (Turbi).
dissolved oxygen (DOo), outlet turbidity (Turbo) and filtered vol-  Finally, the results verify that the hybrid PSO–SVM regression
ume (V) in agreement to the actual experimental values observed method significantly improves the generalization capability
using the PSO–SVM-based models with greater accuracy and suc- achievable with only the SVM-based regressor.
cess than previous research works (Puig-Bargués et al., 2012;
Martí et al., 2013). Indeed, Fig. 3(a)–(c) shows the comparison In summary, this innovative methodology could be applied to
between the DOo, Turbo and V values observed and predicted using other filtration processes with similar or different sources of pollu-
the PSO–SVM-based model with RBF kernel, respectively. There- tants with success, but it is always necessary to take into account
fore, it is necessary the use of a SVM model with RBF kernel in the specificities of each particular process. Consequently, an effec-
order to achieve the best effective approach to nonlinearities pre- tive PSO–SVM-based model is a practical solution to the problem
sent in this regression problem. Obviously, these results coincide of the estimation of the filtration performance in micro-irrigation
again with the outcome criterion of ‘goodness of fit’ (R2) so that sand filters fed with effluents.
the PSO–SVM-based model with a RBF kernel function has been
the best fitting. It can be observed that the models fit quite well
almost all the values. Nevertheless, some isolated extreme values Acknowledgements
present a higher discrepancy with the observed values which it
is to be expected as we have constructed interpolation models. Authors wish to acknowledge the Spanish Ministry of Economy
and Competitiveness for its financial support of this study through
Grant CGL2012-31180 as well as the computational support pro-
5. Conclusions
vided by the Department of Mathematics at University of Oviedo.
Additionally, we would like to thank Anthony Ashworth for his
Based on the experimental and numerical results, the main
revision of English grammar and spelling of the manuscript.
findings of this research work can be summarized as follows:

 Firstly, sand media filters are specially used to avoid emitter References
clogging when water with large amount of organic pollutants
like effluents are used in micro-irrigation systems. In this way, Abbaszadeh, M., Hezarkhani, A., Soltani-Mohammadi, S., 2016. Proposing drilling
locations based on the 3D modeling results of fluid inclusion data using the
estimation of water quality parameters is a very common and support vector regression method. J. Geochem. Explor. 165, 23–34.
serious problem for irrigation engineers. The diagnostic tech- Adin, A., Alon, G., 1986. Mechanisms and process parameters of filter screens. J.
niques commonly used based on the traditional methods are Irrigation Drainage Eng. 112 (4), 293–304.
Burt, C.M., 2010. Hydraulics of Commercial Sand Media Filter Tanks used for
expensive from both the material, economic and human points
Agricultural Drip Irrigation. Irrigation Training & Research Center, California
of view. Consequently, the development of alternative diagnos- Polytechnic State University, San Luis Obispo, California.
tic techniques is necessary. In this sense, the new hybrid PSO– Capra, A., Scicolone, B., 2007. Recycling of poor quality urban wastewater by drip
irrigation system. J. Clean. Prod. 15 (16), 1529–1534.
SVM-based method with a RBF kernel function used in this
Chang, C.-C., Lin, C.-J., 2011. LIBSVM: a library for support vector machines. ACM T.
work is a good choice to evaluate the sand media filter Int. Syst. Technol. 2, 1–27.
performance. Chen, J.-L., Li, G.-S., Wu, S.-J., 2013. Assessing the potential of support vector
 Secondly, the hypothesis that sand filtration diagnosis can be machine for estimating daily solar radiation using sunshine duration. Energy
Convers. Manage. 75, 311–318.
accurately modeled by using a hybrid PSO–SVM-based model Chen, R., Liang, C.-Y., Hong, W.-C., Gu, D.-X., 2015. Forecasting holiday daily tourist
with a RBF kernel function in micro-irrigation sand filters fed flow based on seasonal support vector regression with adaptive genetic
with effluents was confirmed. Indeed, a hybrid PSO–SVM- algorithm. Appl. Soft. Comput. 26, 435–443.
Clerc, M., 2006. Particle Swarm Optimization. Wiley-ISTE, London.
based model with a RBF kernel function was successfully devel- Clerc, M., 2012. Standard Particle Swarm Optimisation: From 2006 to 2011,
Technical Report, <http://clerc.maurice.free.fr/pso/SPSO_descriptions.pdf>.
(Accessed on 5th November 2015).
80 P.J. García Nieto et al. / Computers and Electronics in Agriculture 125 (2016) 74–80

Cristianini, N., Shawe-Taylor, J., 2000. An Introduction to Support Vector Machines estimating outlet dissolved oxygen in micro-irrigation sand filters fed with
and Other Kernel-based Learning Methods. Cambridge University Press, effluents. Comput. Electron. Agric. 99, 176–185.
Cambridge. Negoita, M.G., Reusch, B., 2005. Real World Applications of Computational
de Cos Juez, F.J., García Nieto, P.J., Martínez Torres, J., Taboada Castro, J., 2010. Intelligence. Springer, Berlin.
Analysis of lead times of metallic components in the aerospace industry Nikoo, M.R., Mahjouri, N., 2013. Water quality zoning using probabilistic support
through a supported vector machine model. Math. Comput. Model. 52, 1177– vector machines and self-organizing maps. Water Resour. Manage. 27 (7),
1184. 2577–2594.
Dorigo, M., Stützle, T., 2004. Ant Colony Optimization. Academic Press, San Diego. Olsson, A.E., 2011. Particle Swarm Optimization: Theory, Techniques and
Duran-Ros, M., Puig-Bargués, J., Arbat, G., Barragán, J., Ramírez de Cartagena, F., Applications. Nova Science Publishers, New York.
2008. Definition of a SCADA system for a microirrigation network with Ortiz-García, E.G., Salcedo-Sanz, S., Pérez-Bellido, A.M., Portilla-Figueras, J.A., Prieto,
effluents. Comput. Electron. Agric. 64 (2), 338–342. L., 2010. Prediction of hourly O3 concentrations using support vector regression
Eberhart, R.C., Shi, Y., Kennedy, J., 2001. Swarm Intelligence. Morgan Kaufmann, San algorithms. Atmos. Environ. 44 (35), 4481–4488.
Francisco. Pal, M., Goel, A., 2007. Estimation of discharge and end depth in trapezoidal channel
Efron, B., Tibshirani, R., 1997. Improvements on cross-validation: the.632 + by support vector machines. Water Resour. Manage. 21 (10), 1763–1780.
bootstrap method. J. Am. Stat. Assoc. 92 (438), 548–560. Panigrahi, B.K., Shi, Y., Lim, M.-H., 2011. Handbook of Swarm Intelligence: Concepts,
Elbana, M., Ramírez de Cartagena, F., Puig-Bargués, J., 2012. Effectiveness of sand Principles and Applications. Springer, Berlin.
media filters for removing turbidity and recovering dissolved oxygen from a Picard, R., Cook, D., 1984. Cross-validation of regression models. J. Am. Stat. Assoc.
reclaimed effluent used for micro-irrigation. Agric. Water Manage. 111, 27–33. 79 (387), 575–583.
Fister Jr., I., Yang, X.-S., Fister, I., Brest, J., Fister, D., 2013. A brief review of nature- Puig-Bargués, J., Barragán, J., Ramírez de Cartagena, F., 2005. Filtration of effluents
inspired algorithms for optimization. Elektrotehniski Vestnik 80 (3), 1–7. for microirrigation systems. Trans. ASAE 48 (3), 969–978.
Fister, I., Stranad, D., Yang, X.-S., Fister Jr., I., 2015. Adaptation and hybridization in Puig-Bargués, J., Duran-Ros, M., Arbat, G., Barragán, J., Ramírez de Cartagena, F.,
nature-inspired algorithms. In: Fister, I., Fister, I., Jr. (Eds.), Adaptation and 2012. Prediction by neural networks of filtered volume and outlet parameters in
Hybridization in Computational Intelligence, vol. 18. Springer, New York, pp. 3– micro-irrigation sand filters using effluents. Biosyst. Eng. 111 (1), 126–132.
50. Pujol, J., Duran-Ros, M., Arbat, G., Ramírez de Cartagena, F., Puig-Bargués, J., 2011.
Freedman, D., Pisani, R., Purves, R., 2007. Statistics. W.W. Norton & Company, New Private micro-irrigation costs using reclaimed water. Span. J. Agric. Res. 9 (4),
York. 1120–1129.
Galavi, H., Mirzaei, M., Shui, L.T., Valizadeh, N., 2013. Klang river-level forecasting Schölkopf, B., Smola, A.J., Williamson, R., Bartlett, P., 2000. New support vector
using ARIMA and AFIS models. J. Am. Water Works Assoc. 105 (9), E496–E506. algorithms. Neural Comput. 12 (5), 1207–1245.
García Nieto, P.J., Martínez Torres, J., Araújo Fernández, M., Ordóñez Galán, C., 2012. Shawe-Taylor, J., Cristianini, N., 2004. Kernel Methods for Pattern Analysis.
Support vector machines and neural networks used to evaluate paper Cambridge University Press, Cambridge.
manufactured using Eucalyptus globulus. Appl. Math. Model. 36 (12), 6137– Shrestla, N.K., Shukla, S., 2015. Support vector machine based modeling of
6145. evapotranspiration using hydro-climatic variables in a sub-tropical
Gu, T., Lu, W., Bao, X., Chen, N., 2006. Using support vector regression for the environment. Agric. Forest Meteorol. 200, 172–184.
prediction of the band gap and melting point of binary and ternary compound Simon, D., 2013. Evolutionary Optimization Algorithms. Wiley, New York.
semiconductors. Solid State Sci. 8, 129–136. Steinwart, I., Christmann, A., 2008. Support Vector Machines. Springer, New York.
Karaboga, D., Akay, B., 2009. A survey: algorithms simulating bee swarm Tripathi, V.K., Rajput, T.B.S., Patel, N., 2014. Performance of different filter
intelligence. Artificial Intelligence Rev. 31 (1), 68–85. combinations with surface and subsurface drip irrigation systems for utilizing
Karaboga, D., Gorkemli, B., 2014. A quick artificial bee colony (qABC) algorithm and municipal wastewater. Irrigation Sci. 32 (5), 379–391.
its performance on optimization problems. Appl. Soft Comput. 23, 227–238. Trooien, T.P., Hills, D.J., 2007. Application of biological effluent. In: Lamm, F.R.,
Kanevski, M., Parkin, R., Pozdnukhov, A., Timonin, V., Maignan, M., Demyanov, V., Ayars, J.E., Nakayama, F.S. (Eds.), Microirrigation for Crop Production. Design,
Canu, S., 2004. Environmental data mining and modeling based on machine Operation and Management. Elsevier, Amsterdam, pp. 329–356.
learning algorithms and geostatistics. Environ. Model. Softw. 19, 845–855. Vapnik, V., 1998. Statistical Learning Theory. Wiley-Interscience, New York.
Kennedy, J., Eberhart, R., 1995. Particle swarm optimization. Proceedings of the Wasserman, L., 2003. All of Statistics: A Concise Course in Statistical Inference.
Fourth IEEE International Conference on Neural Networks, vol. 4. IEEE Service Springer, New York.
Center, Perth, Australia, pp. 1942–1948. Wen-Yong, W., Yan, H., Hong-Lu, L., Shi-Yang, Y., Yong, N., 2015. Reclaimed water
Hansen, T., Wang, C.J., 2005. Support vector based battery state of charge estimator. filtration efficiency and drip irrigation emitter performance with different
J. Power Sources 141, 351–358. combinations of sand and disc filters. Irrigation Drainage 64 (3), 362–369.
Hastie, T., Tibshirani, R., Friedman, J., 2003. The Elements of Statistical Learning. Wu, Q., 2009. The forecasting model based on wavelet m – support vector machine.
Springer-Verlag, New York. Expert Syst. Appl. 36 (4), 7604–7610.
Haghverdi, A., Öztürk, H.S., Cornelis, W.M., 2014. Revisiting the pseudo continuous Xu, Y., Ma, C., Liu, Q., Xi, B., Qian, G., Zhang, D., Huo, S., 2015. Method to predict key
pedotransfer function concept: impact of data quality and data mining method. factors affecting lake eutrophication – a new approach based on Support Vector
Geoderma 226–227, 31–38. Regression model. Int. Biodeter. Biodegr. 102, 308–315.
Lamorski, K., Pachepsky, Y., Slawinski, C., Walczak, R.T., 2008. Using support vector Yang, X.-S., Cui, Z., Xiao, R., Gandomi, A.H., Karamanoglu, M., 2013. Swarm
machines to develop pedotransfer functions for water retention of soils in Intelligence and Bio-Inspired Computation: Theory and Applications. Elsevier,
Poland. Soil Sci. Soc. Am. J. 72 (5), 1243–1247. London.
Li, X., Lord, D., Zhang, Y., Xie, Y., 2008. Predicting motor vehicle crashes using Zeng, J., Qiao, W., 2013. Short-term solar power prediction using a support vector
Support Vector Machine models. Accident Anal. Prev. 40, 1611–1618. machine. Renew. Energy 52, 118–127.
Martí, P., Shiri, J., Duran-Ros, M., Arbat, G., Ramírez de Cartagena, F., Puig-Bargués, J.,
2013. Artificial neural networks vs. gene expression programming for