Escolar Documentos
Profissional Documentos
Cultura Documentos
a r t i c l e i n f o a b s t r a c t
Keywords: Engineered floodplain filtration (EFF) system is a versatile low-cost water treatment process wherein
Sorption water contaminants are removed, mainly by adsorption and–or degraded by microorganisms, as the infil-
Triclosan trating water moves from the wastewater treatment plants to any natural water stream. An artificial neu-
Ibuprofen ral network (ANN) based on multilayer perceptrons with back propagation algorithm was used to
Activated carbon
approximate and interpret the complex input/output relationship, essentially to understand the break-
EFF system
Artificial neural networks
through times in EFF system. Triclosan and ibuprofen were selected as the two model pollutants in this
study owing to their environmental significance. The input parameters to the ANN model were inlet con-
centration (ppm) and flow rate (m/d), and the output parameters were six concentration-time pairs (C, t).
These C, t pairs were the times in the breakthrough profile, when 1%, 5%, 25%, 50%, 75% and 95% of the
pollutant was present at the outlet of the system. The 10 set of experimental data points, 5 for each pol-
lutant, statistically investigated in the continuous column studies using the full-factorial design, were
divided into training (8 8) and testing (2 8) set. The most dependable condition for the network
was selected by a trial and error approach and by estimating the determination coefficient (R2) value
(>0.99) achieved during prediction of the testing set. The proposed ANN model for EFF system operation
could be used as a potential alternative for knowledge-based models through proper training and testing
of variables.
Ó 2011 Elsevier Ltd. All rights reserved.
1. Introduction ment systems can easily increase the costs several-fold (Ternes
et al., 2003). The partially eliminated as well as the residual quan-
In the last several decades, the production of pharmaceuticals tities of PPCPs emerging from secondary treatment systems usually
and personal care products (PPCPs) has increased rapidly due to tend to accumulate in the aquatic environment (Hernando, Mez-
the steadfast consumption by both humans and animals. A signif- cua, Fernández-Alba, & Barceló, 2006). In this context, for waters
icant amount of these PPCPs discharged into the aquatic environ- already treated to primary and secondary levels, land treatment
ment have been reportedly detected in wastewater and surface systems could be considered as a promising low-cost, eco-friendly
water at ppb to ppm levels (Kim, Cho, Kim, Vanderford, & Snyder, alternative. One such land treatment system is the floodplain filtra-
2007; Pedrouzo, Reverté, Borrull, Pocurull, & Marcé, 2007; Vander- tion (FPF) system (Fig. 1). FPF is a low-cost water treatment tech-
ford, Pearson, Rexing, & Snyder, 2003). Furthermore, PPCPs have nology that has shown to be efficient in eliminating organics and
shown to pose severe risks and side-effects to the aquatic ecosys- nutrients from secondarily treated wastewater (Kunjikutty,
tem owing to their inherent biological effects (Clara et al., 2005; 2006). Simultaneously, natural attenuation of emerging pollutants
Kasprzyk-Hordern, Dinsdale, & Guwy, 2008; Nakada et al., 2007; including PPCPs discharged in the secondarily treated wastewater
Sanderson et al., 2004). Collectively, pollution of the aquatic eco- effluents cannot also be ruled out during their infiltration and sub-
system by PPCPs is becoming an emerging unmanageable environ- sequent emergence in the rivers. The degree of PPCP elimination in
mental problem worldwide. FPF depends strongly on the proportion of clay, which has the abil-
The conventional treatment practices in wastewater treatment ity to sorb the PPCPs in their matrix, present in the FPF and this va-
plants by and large do not effectively eliminate and/or degrade lue is typically 62% (Jeong et al., 2003; Kunjikutty, Prasher, Patel,
all the PPCPs. Ozonolysis and advanced oxidation techniques have Barrington, & Kim, 2007). The removal of PPCPs from secondarily
shown to remove higher amount of PPCPs, but these tertiary treat- treated wastewater can be enhanced if the floodplains available
nearby wastewater treatment facilities are supplemented with
⇑ Corresponding author. Tel.: +82 52 259 1050; fax: +82 52 221 0152. adequate proportions of a suitable sorbent. This retro-fitted engi-
E-mail address: parkhs@ulsan.ac.kr (H.-S. Park). neered system, hereafter referred as engineered floodplain filtra-
0957-4174/$ - see front matter Ó 2011 Elsevier Ltd. All rights reserved.
doi:10.1016/j.eswa.2011.12.009
S.K. Behera et al. / Expert Systems with Applications 39 (2012) 6052–6060 6053
Table 1
Structures and properties of triclosan and ibuprofen.a
Name Molecular structure MW Log Kow pKa Water solubility (ppm) Application
Triclosan 289.5 4.76 8.14 10 Anti-septic
a
Adapted from Trenholm, Vanderford, Holady, Rexing, and Snyder (2006).
6054 S.K. Behera et al. / Expert Systems with Applications 39 (2012) 6052–6060
Fig. 2. Experimental set-up of the EFF column. (Adapted from Behera (2010)).
distilled water from the bottom to prevent the formation of air bub- 1
f ðxÞ ¼ ð1Þ
bles and to reduce the risk of channeling and then left to equilibrate 1 þ ex
for 2 d. Experiments were performed at ambient room tempera- This produces output in the range of 0–1 and introduces non-line-
tures (20 °C) and the entire unit used for column experiments, arity into the network, which gives the power to capture non-linear
including the tanks, was protected against exposure to light. relationships. The back propagation network is the most prevalent
supervised ANN learning model (Rumelhart, Hinton, & Williams,
2.4. Chemical analysis 1986). It uses the gradient descent algorithm to correct the weights
between interconnected neurons (Maier & Dandy, 2001). During the
An HPLC system (UltimateÒ 3000) (Dionex, Sunnyville, CA, USA) learning process of the network, the algorithm computes the error
equipped with AcclaimÒ C-18 (4.6 mm 250 mm, 5 lm) and Dis- between the predicted and specified target values at the output
covery RP-Amide C-16 (4.6 mm 150 mm, 5 lm) columns were layer. The error function at the output layer can be defined by
used to measure the triclosan and ibuprofen concentrations,
1X
respectively. The UV–vis detection wavelengths were 220 and E¼ ðOd Op Þ2 ð2Þ
230 nm, respectively for triclosan and ibuprofen. The mobile phase 2
used for elution was deionized water (30%) and acetonitrile (70%) where E is the global error function, Od and Op are the desired out-
for triclosan and 25 mM KH2PO4 at pH 3.0 (40%) and acetonitrile put and predicted output by the network, respectively.
(60%) for ibuprofen, delivered at 1 mL/min through the columns.
A sample injection volume of 100 lL was used. The column tem- 2.5.1. Data pre-processing and division
peratures were maintained at 25 °C and 30 °C for triclosan and ibu- The data points from all the EFF systems, operated at initial con-
profen, respectively. centrations of 0.1, 0.55 and 1 ppm were normalized and scaled (Eq.
(3)) to the range of 0–1, so as to suit the transfer function in the
2.5. ANN modeling hidden (sigmoid) and output layer (linear). More comprehensible
and relevant information on the different types of transfer func-
ANNs consists of a system of simple interconnected processing tion, algorithms and internal parameter used for developing ANN
element called neurons. This gives the ability to model any non- models are given elsewhere (Haykin, 1994; Hoskins & Himmelblau,
linear process through a set of unidirectional weighted connec- 1988)
S.K. Behera et al. / Expert Systems with Applications 39 (2012) 6052–6060 6055
b ¼ X X min
X ð3Þ
2.5.4. Selecting the best model architecture
X max X min The performance of the training and test data set were evalu-
ated in terms of the correlation coefficient values (R2), which sub-
where X b is the normalized value, Xmin and Xmax are the minimum
stantially determines the closeness of prediction between the
and maximum values of X, respectively. This processed data was la- desired and predicted output from the network. This is given by:
ter divided into two sets; training and testing. 80% (8) of the data P P
points were used for training the network, while 20% (2) were used
P X Y
2 XY N
for testing the developed network. R ¼ sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
P P ð4Þ
P 2 ð X Þ2 P 2 ð Y Þ2
X N Y N
2.5.2. Model-inputs and outputs
The input vectors to the network were the easily monitored where X is the experimental value, Y is the model predicted value,
parameters namely; input concentration (X1) and flow rate (X2). and N is the total number of observations. Furthermore, for the test-
The output of the ANN model was the times (day) corresponding ing data set, the total mean squared error (RMSE) was computed
to the six concentration levels. The six concentration levels consid- according to the formulae given elsewhere (Elias, Berastegi, Arias,
ered are those corresponding to the non-dimensional parameter & Barona, 2006).
(C/Co) values of 1%, 5%, 25%, 50%, 75% and 95%. These outputs are
denoted as T1, T5, T25, T50, T75 and T95, respectively. Concentra- 2.5.5. Software used
tions below C = 0.01Co and the concentrations of C = Co are practi- ANN based predictive modeling was carried out using the
cally difficult to detect. The six-point method proposed here is shareware version of the neural network and multivariable statis-
assumed to be sufficient to sketch an approximate, yet a close en- tical modeling software NNMODEL (version 1.4, neural fusion, NY).
ough breakthrough curve. Besides, breakpoints at 1% and 5% are The same software was also used to carry out the sensitivity
commonly used in water and wastewater treatment applications. analysis.
Table 4 fixed-bed system, packed with a tree bark as adsorbent, ANN mod-
Best values of network parameters used for training the network. el has been developed with bed depth, influent dye concentration
Training parameters Value and volume of wastewater as the input parameters. The break-
Training count (iterations) 20,000 through curves of adsorption were also exhibited by this model,
Number of neurons in input layer 2 which was found to describe the system well. It was observed that
Number of neurons in hidden layer 4 with increase of bed height and decrease of influent dye concentra-
Number of neurons in output layer 6 tion, the breakthrough time was delayed. In that study, the authors
Learning rate 0.9
Momentum term 0.8
showed that neural networks can be used as an effective technique
Error tolerance 0.00001 to model, estimate state variables and predict the absorption
Training algorithm Gradient descent algorithm behavior with satisfactory results (Balci, Keskinkan, & Avci,
Number of training data set 8 2010). Similarly, Oguz and Ersoy (2010) conducted continuous
Number of test data set 2
flow experiments in fixed-bed adsorption columns to remove
Cu2+ ions from aqueous solution using sunflower shells as adsor-
bent. They investigated the effects of flow rate, inlet concentration,
3.2.2. Predictive capability of the developed model pH, bed depth and particle size on the breakthrough characteristics
From Figs. 5 and 6, it can be seen that some data points show of the adsorption system and subsequently developed an ANN
deviancy between the measured and predicted C/Co profiles. This model to predict the Cu2+ concentration removed from aqueous
is probably due to the fact that different conditions exhibited dif- solution as a function of empirical parameters. The relationship be-
ferent onset times for reaching breakthrough. The delay in achiev- tween the predicted and observed data of the model yielded in a
ing breakthrough, i.e., at low flow rates and concentrations would correlation coefficient of 0.986. On the other hand, ANN models
have caused an impact in the neural networks learning/generaliza- have also been proposed to predict the performance of a continu-
tion pattern while predicting the performance parameter. The cor- ously operating immobilized-cell biofilter treating NH3 vapors.
responding connection weights and bias terms for the developed The input parameters to the ANN model were inlet concentration,
network are presented in Table 6. As observed in Fig. 5, the neural loading rate, flow rate and pressure drop, and the output parame-
network predicted breakthrough curves are very close to the corre- ters were removal efficiency and elimination capacity. The results
sponding exact ones. Although the developed neural network is showed that a multilayer network (4–4–2) with a BPNN was able
specific, more general networks using wider ranges and variety to predict biofilter performance effectively with R2 values more
of input parameters can be developed using the procedure adopted than 0.98 (Rene, Kim, & Park, 2009).
in this work.
The application of the BPNN to predict the filtrate water quality 3.2.3. Sensitivity analysis of the developed model
parameters; temperature, turbidity, heterotrophic bacteria, and Sensitivity analysis is a useful technique to assess the relative
coliform removal have been studied using riverbank filtration facil- contribution of the input variables to the performance of a neural
ities, wherein optimization of the ANN models’ structure was per- network by testing the neural network when each input variable
formed on a trial and error basis. It was observed that the BPNN is unavailable. This is performed by varying the value of one
model prediction result produced excellent agreement with the parameter over a certain range and by observing the associated ef-
measured data at a very high correlation coefficient (>0.99) in all fect on the desired outcome. For the problem studied in this work,
the cases (Sahoo et al., 2005). Experimental breakthrough curves the two input parameters were checked for their sensitivity to the
of pesticide in surface water were assessed in fixed-beds and mod- breakthrough curve. The four-hidden nodes network was used for
eled using neural network approaches. Fifteen potentially influen- this test. However, to avoid any bias, a representative set was se-
tial parameters were identified in that study for the dynamic lected by taking the average of each input parameter from the
adsorption of pesticides from surface water. For all breakthrough training and testing databases. The two parameters in the average
curves, computations were performed for specific zones of the set are Co = 0.55 ppm and Q = 3 m/d. The value of each of these two
breakthrough curves; breakthrough zone (C/C0 < 0.10), saturation parameters is perturbed about its mean and the output break-
zone (C/C0 > 0.9), and the global breakthrough curve (0 < C/ through curve is determined using the predictive network. It is
C0 < 1). Neural networks provide quite satisfactory predictions for noteworthy to mention here that all perturbations were made
pesticide dynamic adsorption onto activated carbon with R2 larger within the limits of the applicability of each input parameter.
than 0.981 (Faur, Cougnaud, Dreyfus, & Le Cloirec, 2008). ANN- The average absolute sensitivity matrix (Ski,abs) can be defined as
based models were also developed to predict the performance of follows:
moving bed biofilm reactor for the treatment of a toxic and highly Pp ðpÞ
refractory aromatic amine compound, aniline. For experiments p¼1 ski
conducted over a period of 3 months, the reactor performance effi- Ski;abs ¼ ð5Þ
p
ciency was predicted at different retention times of 8, 24, 48 and
ðpÞ
72 h, respectively with influent COD ranging from 100 to where Ski is the sensitivity of the trained output and p is the num-
4000 ppm. The proposed ANN model demonstrated the ability of ber of training patterns presented to the network.
BPNN to predict the performance of moving bed biofilm reactor The higher average absolute sensitivity values (Table 7) for flow
with sufficient accuracy (Delnavaz, Ayati, & Ganjidoust, 2010). rate rather than concentration suggest that flow change has a sig-
For determining adsorption capability of two textile dyes in a nificant and greater influence on the C/Co values than the initial
Table 5
Coefficient of regression (R2) values obtained during training and testing.
Triclosan Ibuprofen
Run-1 Run-2 Run-3 Run-4 Run-1 Run-2 Run-3 Run-4
Training 0.999 0.989 0.994 0.999 0.999 0.984 0.993 0.999
Testing 0.967 0.993
6058 S.K. Behera et al. / Expert Systems with Applications 39 (2012) 6052–6060
Fig. 5. Training data for triclosan (a) and ibuprofen (b) (Run 1–4/p: model predicted; Run 1–4/e: experimentally observed).
70 Table 6
Ibuprofen - CP Hidden layer connection weights and bias term after network training.
60 Triclosan - CP Variable HID001 HID002 HID003 HID004
ANN Predicted
Table 7
concentrations of both triclosan and ibuprofen. In general, the pre- Average absolute sensitivity values for the outputs.
dictive ability of the proposed model using the concepts of artificial
intelligence and the back propagation algorithm was high at the Variable T1 T5 T25 T50 T75 T95
95% confidence interval and meaningfully significant, as ascer- Q 0.7586 0.7934 0.7368 0.7214 0.7186 0.7585
tained from the correlation coefficient values between the mea- C 0.2414 0.2066 0.2632 0.2785 0.2814 0.2415
sured and predicted outputs in the training and test data. Note. Q: flow rate (m/d), C: concentration (ppm).
the column can be easily predicted. Consequently, the flow can be Acknowledgements
diverted to a fresh column while this previously used column can
be regenerated separately. Furthermore, the column lysimeter data This research work was supported by the research funds from
can also be used to predict the performance of EFF system in real- University of Ulsan in South Korea. S.K.B. thankfully acknowledges
time applications. For example, Nema, Chelani, Ojha, Kumar, and the Brain Korea 21 Post Doctoral fellowship from Ministry of Edu-
Khanna (2004) have shown that the ANN predicted results of their cation, Science and Technology through Environmental Engineer-
column lysimeter experiment was in perfect agreement with ob- ing Program at University of Ulsan. E.R.R. thanks the Ministerio
served field data from their pilot soil aquifer treatment system de Ciencia e Innovación, Spain, for his ‘Juan de La Cierva’
and indicated the utility of column lysimeter studies in evaluating fellowship.
soil aquifer treatment systems. However, since the study period
was only about 3 months resulting in limited data collection, fur-
ther studies along similar lines over longer periods would add a
References
further degree of confidence in the translation of column lysimeter
results to the actual full-scale EFF system. Anew, in order to extrap- Ahmad, A. A., & Hameed, B. H. (2010). Fixed-bed adsorption of reactive azo dye onto
olate the results obtained in this modeling work from lab-scale granular activated carbon prepared from waste. Journal of Hazardous Materials,
experiments to practical situations, it is always possible to intro- 175, 298–303.
Balci, B., Keskinkan, O., & Avci, M. (2010). Use of BDST and an ANN model for
duce new inputs and new values to the already existing database prediction of dye adsorption efficiency of Eucalyptus camaldulensis barks in
and the network can be trained whenever required. Apart from this fixed-bed system. Expert Systems with Applications, 38, 949–956.
flexibility in updating the model as and when required, ANNs also Basheer, I. A., & Najjar, Y. M. (1996). Predicting dynamic response of adsorption
columns with neural nets. Journal of Computing in Civil Engineering, 10(1),
provide sufficient information to the plant operator about all pos- 31–39.
sible interactions between predictor variables (T1–T95) and can Behera, S. K. (2010). Removal of triclosan and ibuprofen by engineered floodplain
also signal/alarm the operator if unusual deviations in values are filtration system. Unpublished doctoral dissertation, University of Ulsan, South
Korea.
noticed. Clara, M., Strenn, B., Gans, O., Martinez, E., Kreuzinger, N., & Kroiss, H. (2005).
Removal of selected pharmaceuticals, fragrances and endocrine disrupting
compounds in a membrane bioreactor and conventional wastewater treatment
4. Conclusions plants. Water Research, 39, 4797–4807.
Delnavaz, M., Ayati, B., & Ganjidoust, H. (2010). Prediction of moving bed biofilm
reactor (MBBR) performance for the treatment of aniline using artificial neural
The following conclusions were made based on the perfor- networks (ANN). Journal of Hazardous Materials, 179, 769–775.
mance modeling of the EFF systems using neural networks: Elias, A., Berastegi, I., Arias, R., & Barona, A. (2006). Neural networks as a tool for
control and management of a biological reactor for treating hydrogen sulphide.
Bioprocess and Biosystems Engineering, 29, 129–136.
1. The results from this ANN-based data-driven modeling strategy Faur, C., Cougnaud, A., Dreyfus, G., & Le Cloirec, P. (2008). Modeling the
shows that the EFF system performance, in terms of C/Co values, breakthrough of activated carbon filters by pesticides in surface waters
with static and recurrent neural networks. Chemical Engineering Journal, 145,
can be predicted with high confidence level (95%) using easily 7–15.
measurable operational parameters, namely, inlet concentra- Haykin, S. (1994). Neural networks: A comprehensive foundation. New York:
tion and flow rate. The model was adequately trained with Macmillan College Co.
Hernando, M. D., Mezcua, M., Fernández-Alba, A. R., & Barceló, D. (2006).
the laboratory-scale EFF column data and tested with a separate Environmental risk assessment of pharmaceutical residues in wastewater
data set. effluents, surface waters and sediments. Talanta, 69, 334–342.
2. A three-layered MLP was sufficient to describe the break- Hoskins, J. C., & Himmelblau, D. M. (1988). Artificial neural network models of
knowledge representation in chemical engineering. Computers and Chemical
through using connection weights and bias terms (thresholds).
Engineering, 12, 881–890.
The suitable network architecture of the model was determined Jeong, B. R., Chung, J. B., Kim, S. H., Lee, Y. D., Cho, H. J., & Baek, N. J. (2003).
through a vigorous trial and error approach. The computed Rhizosphere enhances removal of organic matter and nitrogen from river
determination coefficient (R2) values for the test data set show water in floodplain filtration. Korean Journal of Soil Science and Fertilizer, 36,
8–15.
high correlation between the predicted and measured perfor- Kasprzyk-Hordern, B., Dinsdale, R. M., & Guwy, A. J. (2008). The occurrence of
mance values. The best topology was found to be a simple pharmaceuticals, personal care products, endocrine disruptors and illicit drugs
three-layered network, 2–4–6, achieved at the following set- in surface water in South Wales, UK. Water Research, 42, 3498–3518.
Kim, S. D., Cho, J., Kim, I. S., Vanderford, B. J., & Snyder, S. A. (2007). Occurrence and
tings of internal network parameters: learning rate (0.9), removal of pharmaceuticals and endocrine disruptors in South Korean surface,
momentum (0.8) and a training count (epoch size) of 20,000. drinking, and waste waters. Water Research, 41, 1013–1021.
3. From a practical application view point, if the pollutants and Knaebel, K., Ruthven, D. M., Humphrey, J. L., & Carr, R. (1999). Adsorption
technologies. In P. Radecki, J. C. Crittenden, D. R. Shonnard, & J. L. Bulloch
their concentration are known in the inlet stream, they can be (Eds.), Emerging separation and separative reaction technologies for process waste
added to the network and can be trained easily. Consequently, reduction (pp. 33–129). New York: AIChE.
the time for removing 1%, 5%, 25%, 50%, 75%, and 95% of the tar- Kunjikutty, S. P. (2006). Floodplain filtration for treating municipal wastewaters. Ph.D.
thesis, Department of Bioresource Engineering, McGill University, Montreal,
get pollutants in the EFF by adsorption can be evaluated, which Canada.
indicates the versatility of the model. Kunjikutty, S. P., Prasher, S. O., Patel, R. M., Barrington, S. F., & Kim, S. H. (2007).
4. The robustness of the ANN was further explored with a sensitiv- Simulation of nitrogen transport in soil under municipal wastewater
application using LEACHN. Journal of the American Water Resources Association,
ity analysis of the input parameters, which showed that flow
43, 1097–1107.
rate was a more critical factor affecting EFF system perfor- Lua, A. C., & Jia, Q. (2009). Adsorption of phenol by oil-palm-shell activated carbons
mance, than the inlet concentrations of triclosan and ibuprofen. in a fixed bed. Chemical Engineering Journal, 150, 455–461.
5. The models developed in this study can be extended to biotic Maier, H. R., & Dandy, G. C. (1998). The effect of internal parameters and geometry
on the performance of back-propagation neural networks: an empirical study.
conditions by incorporating some easily measurable state Environmental Modeling Software, 13, 193–209.
parameters such as oxygen utilization rate and specific activity Maier, H. R., & Dandy, G. C. (2001). Neural network based modeling of
of the microorganisms to predict the breakthrough in a long environmental variables: A systematic approach. Mathematical and Computer
Modeling, 33, 669–682.
term assessment. Montgomery, D. C. (1991). Design and analysis of experiments (3rd ed.). New York:
6. Although ANN models are not a substitute for a mechanistic Wiley.
model, they were found to be promising for real-time use, in Nakada, N., Shinohara, H., Murata, A., Kiri, K., Managaki, S., Sato, N., et al. (2007).
Removal of selected pharmaceuticals and personal care products (PPCPs) and
predicting the breakthrough characteristics at different process endocrine-disrupting chemicals (EDCs) during sand filtration and ozonation at a
conditions. municipal sewage treatment plant. Water Research, 41, 4373–4382.
6060 S.K. Behera et al. / Expert Systems with Applications 39 (2012) 6052–6060
Nema, P., Chelani, A., Ojha, C. S. P., Kumar, A., & Khanna, P. (2004). Utility of column Sahoo, G. B., Ray, C., Wang, J. Z., Hubbs, S. A., Song, R., Jasperse, J., et al. (2005). Use of
lysimeter for design of soil aquifer treatment system for wastewater renovation artificial neural networks to evaluate the effectiveness of riverbank filtration.
using artificial neural networks. Journal of Environmental Engineering – ASCE, Water Research, 39, 2505–2516.
130(12), 1534–1542. Sanderson, H., Johnson, D. J., Reitsma, T., Brain, R. A., Wilson, C. J., & Solomon, K. R.
Oguz, E., & Ersoy, M. (2010). Removal of Cu2+ from aqueous solution by adsorption (2004). Ranking and prioritization of environmental risks of pharmaceuticals in
in a fixed bed column and neural network modeling. Chemical Engineering surface waters. Regulatory Toxicology and Pharmacology, 39, 158–183.
Journal, 164, 56–62. Singh, S., Srivastava, V. C., & Mall, I. D. (2009). Fixed-bed study for adsorptive
Pedrouzo, M., Reverté, S., Borrull, F., Pocurull, E., & Marcé, R. M. (2007). removal of furfural by activated carbon. Colloids and Surfaces A: Physicochemical
Pharmaceutical determination in surface and wastewaters using high- and Engineering Aspects, 332, 50–56.
performance liquid chromatography–(electrospray)–mass spectrometry. Ternes, T. A., Stüber, J., Herrmann, N., McDowell, D., Ried, A., Kampmann, M., et al.
Journal of Separation Science, 30, 297–303. (2003). Ozonation: a tool for removal of pharmaceuticals, contrast media and
Rene, E. R., Kim, J.-H., & Park, H.-S. (2009). Immobilized cell biofilter: Results of musk fragrances from wastewater. Water Research, 37, 1976–1982.
performance and neural modeling strategies for NH3 vapor removal from waste Trenholm, R. A., Vanderford, B. J., Holady, J. C., Rexing, D. J., & Snyder, S. A. (2006).
gases. Aerosol and Air Quality Research, 9, 379–384. Broad range analysis of endocrine disruptors and pharmaceuticals using gas
Rumelhart, D. E., Hinton, G. E., & Williams, R. J. (1986). Learning internal chromatography and liquid chromatography tandem mass spectrometry.
representations by error propagation. In D. E. Rumelhart, J. L. McClelland, & Chemosphere, 65, 1990–1998.
the PDP Research Group (Eds.). Paralled distributed processing. Explorations in the Vanderford, B. J., Pearson, R. A., Rexing, D. J., & Snyder, S. A. (2003). Analysis of
microstructure of cognition. Foundations (Vol. 1, pp. 318–362). Cambridge, MA: endocrine disruptors, pharmaceuticals, and personal care products in water
The MIT Press. using liquid chromatography/tandem mass spectrometry. Analytical Chemistry,
Ruthven, D. M. (1984). Principles of adsorption and adsorption processes. New York: 75, 6265–6274.
John Wiley.