Accepted Manuscript

Predicting of ultrafiltration performances by advanced data analysis

B. Teychene, A. Touffet, J. Baron, B. Welte, M. Joyeux, H. Gallard

PII: S0043-1354(17)30939-9
DOI: 10.1016/j.watres.2017.11.023
Reference: WR 13353

To appear in: Water Research

Received Date: 9 June 2017

Revised Date: 19 September 2017
Accepted Date: 8 November 2017

Please cite this article as: Teychene, B., Touffet, A., Baron, J., Welte, B., Joyeux, M., Gallard, H.,
Predicting of ultrafiltration performances by advanced data analysis, Water Research (2017), doi:
10.1016/j.watres.2017.11.023.

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to
our customers we are providing this early version of the manuscript. The manuscript will undergo
copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please
note that during the production process errors may be discovered which could affect the content, and all
legal disclaimers that apply to the journal pertain.
 ACCEPTED MANUSCRIPT

PT
RI
U SC
AN
M
D
TE
EP
C
AC
 ACCEPTED MANUSCRIPT
Highlights :

- Cluster analysis help to discriminate filtration performances of five different water resources

sampled during one year.

- Principal component analysis confirms the good correlation between filtration performances and

organic matter indicators.

PT
- Prediction of filtration performances was obtained based on fluorescence analysis and carbon

content.

RI
U SC
AN
M
D
TE
C EP
AC
 ACCEPTED MANUSCRIPT
Predicting of Ultrafiltration Performances by Advanced Data

Analysis.

B.Teychene1+*, A.Touffet1+, J. Baron2, B. Welte2, M. Joyeux2, H.Gallard1

1.Institut de Chimie des Milieux et des Matériaux de Poitiers (UMR CNRS 7285) École Nationale

Supérieure d'Ingénieurs de Poitiers (ENSIP), Université de Poitiers. 1 rue Marcel Doré, Bâtiment 1,

PT
86022 POITIERS CEDEX, France. Benoit.teychene@univ-poitiers.fr

RI
2.Direction de la Recherche et du Développement de la Qualité de l'Eau. Eau de Paris. 33 avenue

Jean Jaurès, 94200 IVRY SUR SEINE, France

SC
*corresponding author: benoit.teychene@univ-poitiers.fr ; Tel : + 33 (0)549453846

+These authors contributed equally to this work

U
Keywords: Ultrafiltration, karst area, Principal component analysis, Fluorescence Excitation
AN
Emission matrix, Cluster analysis, ARIMA model.

Highlights :
M

- Cluster analysis help to discriminate filtration performances of five different water resources
D

sampled during one year.

TE

- Principal component analysis confirms the good correlation between filtration performances and

organic matter indicators.

EP

- Prediction of filtration performances was obtained based on fluorescence analysis and carbon

content.
C
AC
 ACCEPTED MANUSCRIPT
Abstract

In order to optimize drinking water production operation, membrane users can use several analytical

tools that help membrane fouling prediction and alleviate fouling by a proper feed water resource

selection. However, during strong fouling event, membrane decision-makers still face short-term

deadline to decide between different options (e.g. optimization of pretreatment or change in feed

PT
water quality). Hence, statistical approach might help to better select the most relevant analytical

parameter related to fouling potential of a specific resource in order to speed-up decision taking. In

RI
this study, the physical and chemical properties and the filtration performances (at lab-scale) of five

SC
ground water resources, selected as potential resources of a large drinking production site of Paris

(France), was evaluated through one year. Principal component analysis emphasizes the strong link

U
between waters’ organic matrix and fouling propensity. Cluster analysis of filtration performances
AN
allowed classifying the water samples into three groups exhibiting strong, low and intermediate

fouling. Finally, multiple linear regressions performed on all collected data indicated that strong
M

fouling events were related to a combined increase of carbon content and protein like-substances

while intermediate fouling might only be anticipated by an increase of fluorescence signal

D

associated to protein like-substances. This study demonstrates that advanced data analysis might be
TE

a powerful tool to better manage water resources selection used for drinking water production and
EP

to forecast filtration performances in a context of water quality degradation.

C
AC
 ACCEPTED MANUSCRIPT
1. Introduction

Aquifers in karst regions are at least a partial source of drinking water supply to almost a quarter of

the world's population [Ford and Williams, 2007]. These water resources are characterized by

strong and unpredictable variation of water quality with strong increases of turbidity and dissolved

organic carbon (DOC) during rain events (Pronk et al. 2006). Furthermore, climate simulations

PT
project a strong increase in temperature and a decrease of precipitation in many karst regions in the

world over the next decades, which will likely increase the stress on karst water resource in terms of

RI
both quantity and quality (Hartmann et al., 2014). Membrane technologies including UF and

SC
combination PAC/UF have been proposed as reliable, flexible and cost-effective treatments for karst

water resources (Pianta et al., 2000).

U
Despite widespread application of ultrafiltration (UF) for drinking water production, membrane
AN
fouling is one of the main limitations of membrane installations (Nakatsuka et al. (1996), Pianta et

al. (1998), Gao et al. (2011)). Several studies were dedicated to membrane fouling understanding
M

and results demonstrated a strong impact of process management (i.e. backwash and chemical

cleaning) and water quality on filtration performances (Barbot et al. (2009), Touffet et al. (2015),
D

Huang et al. (2014)). Among all parameters, natural organic matter (NOM) was found to be the
TE

most problematic regarding fouling propensity (Lee et al. (2006), Hong et al. (1997), Amy (2008)).
EP

More precisely, protein- and polysaccharides-like compounds were found to play a crucial role in

membrane fouling (Zheng et al. (2009) & (2017)). However, due to widely varying
C

physicochemical properties the prediction and understanding of fouling is difficult. Consequently,

AC

membrane users aim at reducing as much as possible NOM concentration in membrane feed water

by using pre-treatment operations (coagulation-flocculation, adsorption, prefiltration) or by

appropriate selection of water resources (Guigui et al. (2002), Campinas et al. (2010), Peiris et al.

(2011), Mao et al. (2013) ). Hence, water quality and NOM content have to be monitored using

several global (organic carbon, pH, turbidity...) or more specific analytical tools

(liquid chromatography with on-line organic carbon detection, Fluorescence Excitation Emission
 ACCEPTED MANUSCRIPT
Matrix (FEEM)). Real-time data monitoring (RTDM) of global analytical parameters (i.e.: TOC,

UV@254nm,...) can be used in order to predict and to reduce the impact of NOM on membrane

fouling (Vera et al. (2017), Chen et al. (2006), Hamachi et al. (1999), Curcio et al. (2006), Wetterau

et al. (1996), Shamsuddin et al. (2015)). Thus, a large amount of data needs to be continuously

processed using advanced data analysis in order to better adapt membrane pre-treatment or filtration

PT
process. The use of principal component analysis (PCA) could help identifying most relevant

parameters regarding membrane fouling for a given water resource (Peiris et al. (2010a), Peiris et al.

RI
(2010b), Peiris et al. (2012), Chen et al. (2014), Cruz et al. (2017)) or used to monitor fouling with

SC
innovative analytical technique (Virtanen et al. 2017). The important role of colloidal compounds

and protein-like substances (identified by FEEM) in UF fouling was confirmed by PCA (Peiris et

U
al. (2010a)). More recently, a framework of UF fouling quantification and classification using
AN
advanced data analysis tools (PCA, cluster analysis, PARAFAC, etc…) was developed based on

data obtained at pilot scale during two years (Chen et al. (2017)). The results confirmed the
M

important role of protein-like substances and biopolymers on both reversible and irreversible

fouling, and emphasized that colloidal/particulate matter can alleviate the extent of irreversible
D

fouling during filtration of surface water (Chen et al. (2017)). Recently, Liu et al. (2017) used PCA
TE

to correlate UF performances regarding different natural waters and confirmed that biopolymers
EP

fulvic-acid-like substances were linked to hydraulic reversible fouling. Therefore, applying

advanced data analysis might help membrane users in their choice of feed water resource in order to
C

alleviate membrane fouling and to improve produced water quality.

AC

The present study aims to use advanced data analysis tools to correlate data obtained from several

physicochemical analyses to fouling potential of five different groundwater resources in a karst area

and potentially used as drinking water resources for Paris (France). Water resources were sampled,

analyzed and filtered every month in 2012. All generated data (more than 1000) were processed

through PCA, cluster analysis and Autoregressive integrated moving average (ARIMA) model in

order to develop a filtration performances prediction model based on physical and chemical
 ACCEPTED MANUSCRIPT
parameters of the tested resources.

2. Materials and methods

2.1 Sampled waters and analytical tools

PT
The five ground water resources, situated in the Dreux district about 75 km SW of Paris

(France), are called Erigny, Foisys, Graviers, Nouvet and Rivière. They are owned by the water

RI
company, Eau de Paris, and are used as water resource of the Saint Cloud drinking water production

plant. The St Cloud drinking water treatment plant can produce up to 100 000 m3 per day using a

SC
process combining powdered activated carbon and ultrafiltration (360 hollow fiber membrane

U
modules made of cellulose acetate, Aquasource, France).
AN
As shown on figure 1, the groundwater resources are located close to each other in a karst area

region with high soil permeability. The flood mapping risk revealed a high risk of sediment flooding
M

and the occurrence of an overlying aquifer. The water quality monitoring performed by Eau de Paris

showed that water quality remained stable in dry period but turbidity and fecal indicator bacteria
D

increased significantly during strong rain events (data not shown). Also as indicated on figure 1,
TE

three wastewater treatment plants (WWTP) are close to the five water resource locations (approx.
EP

2 km).

Each resource was sampled every month at the same day, during one year between November
C

2011 and December 2012 (except during August 2012). Once received all waters were prefiltered at
AC

200 µm on a steel sieve, stored at 4°C in dark before analysis and filtration for a maximum duration

of 10 days.

All water samples were characterized in terms of pH (LPH 330 T TACUSSEL), conductivity

(Cyberscan 510, EUtech), anion concentrations (Cl-, NO3-, SO42-) using ion chromatography

(Dionex ICS 3000), cations (Ca2+, K+, Na+, Mg2+) by ICP-OES (Inductively Coupled Plasma with

Optical Emission Spectrometry, Perkin Elmer Optima 4300 DW), alkalinity by acid titration
 ACCEPTED MANUSCRIPT
according to NF EN ISO 9963 method, turbidity by using a WTW Lab Turb® 550 IR, TOC by

using a Shimadzu TOC-Vcsh, UV254 absorbance (SAFAS Double Energy System) and fluorescence

excitation-emission matrix (FEEM) using a Fluoromax-2 ISA (Jobin Yvon-Spex). The FEEM range

was set at 200-400 nm and 300-550 nm for excitation and emission wavelengths, respectively.

These analytical tools were selected because they are already used on site by membranes' operators.

PT
2.2 Description of the bench scale filtration unit and protocol

RI
SC
Every dead-end filtration tests were performed using a new homemade hollow fiber mini-module

(surface area of 29 cm2) made of 4 pristine fibers (25 cm long) closed at one end (Touffet (2014)).

U
The fibers were the same as used in the St Cloud drinking water treatment plant. They are made of
AN
cellulose acetate (Aquasource, France) with a 100 kDa MWCO and a pure water flux of

260 ± 40 L h-1 m-² bar-1. Every home made mini-modules exhibiting pure water flux higher than
M

15% of the average value were discarded in order to maximize filtration results reproducibility.

Filtration tests were performed using an automated filtration unit with the same design as used by
D

Touffet et al. (2015). The maximum transmembrane pressure allowed on the filtration unit was set
TE

at 1.1 bars, which is lower than the burst pressure of 1.5 bars given by the manufacturer For the
EP

evaluation of filtration performances, the suspension was filtered at 20°C for 24 hours

corresponding to 70 filtration cycles. Each cycle consisted of 20 min of inside-out dead-end

C

filtration at a constant flow of 96 L h-1 m-2 (4 ml min-1) followed by a backwash (BW) step. BW
AC

steps were processed using chlorinated permeate at 4 mgCl2 L-1 for 1 min at 92 L h1 m-2 (8 ml min-
1
). These filtration conditions were chosen and adapted to reflect the operating conditions used on

site (i.e. filtration flux of 100 L h-1 m-2 interspersed with chlorinated permeate backwash).

The filtration performances were determined in terms of irreversible fouling rate (IFR) (Eq 1) and

critical filtered volume (CFV) defined here as the filtered volume value when the fouling becomes

irreversible (Transmembrane pressure (TMP) after BW higher than 10% of the initial TMP)
 ACCEPTED MANUSCRIPT
(Touffet et al. (2015), Chen et al. (2017), Bessière et al. (2005)).

IFR=
1
(
µ 20°C J dt
d (Pio Po )
) Equation 1

PT
with IFR the irreversible fouling rate (m-1 h-1), Pio and Po are the TMP (Pa) at the beginning of the

i-th filtration cycle and the initial TMP (virgin membrane), respectively. µ(20°C) is the permeate water

RI
viscosity at 20°C (Pa s) and J the filtration flux (m3 m-2 h-1).

SC
Therefore, those two parameters were used to characterize the fouling propensity during

filtration tests. It is important to note that these two parameters are not independent. If no

U
irreversible fouling occurred, the IFR value was equal to zero and a maximum value of CFV was
AN
obtained. In such case, the maximum value of the CFV was the maximum filtered volume i.e.

around 1900 L m-2 over 24-hours filtration run. In contrast a strong fouling was represented by a
M

high IFR and small CFV values. Among all filtration tests the maximum observed IFR values were

about 55 m-1 h-1 and the corresponding minimal CFV value was equal to 100 L m-2. The calculated
D

standard variations of CFV and IFR, obtained from a repetition test performed on Nouvet (n=2) and
TE

Foisys (n=2) during January 2012, were equal to 20% and 17%, respectively.
EP

2.3 Statistical analysis approach

C
AC

First, FEEM data were computed using parallel factor analysis (usually called PARAFAC) thank

to the algorithm developed by the PROTEE laboratory (University of Toulon, France) on the N-way

Matlab 7.0 toolbox (Mathworks, Natick, USA). This statistical tool allows to construct a three-

dimensional matrices on all collected fluorescence data (Matrix dimension: Excitation × Emission ×

Number of samples) (Alberts and Tackacks, (2004)).

Filtration results were classified using cluster analysis, performed on all IFR and CFV values
 ACCEPTED MANUSCRIPT
collected over the year, using the Ward's method. Subsequently, all analytical results (64 different

samples and 16 analytical parameters) were correlated to filtration tests data (two parameters : CFV

and IFR) using Principal Component Analysis (PCA) performed on a 1152 data matrix (64 × 18).

Finally, multiple linear regressions (MLR) were used to develop prediction models of filtration

performances of all studied resources from their specific physical and-chemical characteristics.

PT
MLR was performed using the forward/backward method based on the Bayesian information

criterion (BIC). Also, the autoregressive integrated moving average model (ARIMA(p,d,q)) was

RI
applied to forecast Nouvet and Foisys resources filtration performances. The applied parameters to

SC
the ARIMA model were : the order p equals to 2, d the degree of differencing equals to 1, and q the

order of the moving-average model defined to 0. All statistical tools used in this study (PCA,

U
Cluster analysis, MLR, ARIMA) were computed using R Commander Software version 1.8-1 with
AN
FactomineR and time series packages (Husson et al. (2008)).
M
D
TE
C EP
AC
 ACCEPTED MANUSCRIPT
3. Results and discussion

3.1 Evolution of water quality.

The analytical results (i.e. pH, TOC, turbidity, major ion concentrations and fluorescence

data) obtained for each water resource and every month between November 2011 and December

2012 are given in supporting information (Table S1). TOC values shown in Figure 2 ranged from

PT
0.5 to 3.6 mg L-1 for all investigated resources throughout the year. Globally, higher TOC values

were observed for Rivière and Nouvet. Two large TOC increases were reported in January and

RI
October 2012 with TOC values ranging from 2.15 and 3.60 mg L-1. An increase in the precipitation

SC
level was usually followed by an increase of TOC as observed for October 2012 and from February

to June 2012. Significant snowfall in late December 2011 followed by progressive snow melting

U
during January 2012 would explain the discrepancy between precipitation event and TOC increase
AN
at that period. The UV absorbance at 254 nm followed the same trend as TOC with a strong linear

correlation between both parameters (Pearson’s r = 0.952, n = 65, Table 1).

M

Figure S1 shows that, regardless the resource or the sampling period, conductivity values

were relatively similar and ranged from 368 to 489 µS cm-1 (T = 20°C). The lowest values were
D

obtained during summer 2012 (June & July) due to dilution after high precipitation events at the
TE

same period. As said before, the five studied resources are situated in a karst area and consequently
EP

are influenced by rainfall events. As with conductivity values, concentrations of inorganic species

and pH slightly varied over the year (Table S1). The pH exhibited an average value of 7.4 ± 0.1 for
C

all waters with min and max values of 7.2 and 7.9, respectively. The hardness was moderately hard
AC

with average calcium concentration of 84 mg L-1 for Rivière and Nouvet and 90 mg L-1 for Erigny,

Foisys and Graviers. Alkalinity varied from 104 mg.L-1 as CaCO3 for Nouvet to 188 mg.L-1 for

Erigny.

Average turbidity was 0.77 NTU and maximum values of 2.2 NTU were reported in January and

October 2012 for Nouvet and Graviers (Table S1). These results further demonstrate that

investigated resources are impacted by rainfall; important rain events, particularly during winter and
 ACCEPTED MANUSCRIPT
fall, led to an increase of allochtonous organic matter and turbidity due to soil lixiviation. During

spring and summer, the impact of rain events was lower because vegetation reduced soil lixiviation.

In order to deeply investigate the NOM properties throughout the year, all water samples were

analyzed using FEEM and computed with PARAFAC. As depicted in figure 3 and Table S1, three

components were extracted from the FEEM analysis with 99.8% statically convergence and a

PT
CONCORDIA test over 50%. These three components were previously identified by different

published works (Stedmon and Markager (2003), Yao et al. (2011), Ishii et al. (2012)). Component

RI
C1 refers to the fluorophore centered at excitation and emission wavelengths of 250 nm and 470 nm,

SC
respectively (Fig. 3-left). Component C2 corresponds to the fluorophore centered at excitation and

emission wavelengths of 250 nm and 400 nm, respectively (Fig. 3-center). Components C1 and C2

U
were attributed to humic and fulvic-like substances, respectively (Yang et al. (2015)). The third
AN
component C3 was represented by a fluorophore centered at excitation and emission wavelengths of

225 and 325 nm, respectively (Fig.3-right). This fluorophore is usually attributed to protein-like
M

substances and exhibits a strong correlation with biochemical oxygen demand (Baker and Genty

(1999)). The presence of this fluorophore in natural waters was demonstrated to be site-dependent
D

and usually associated to organic matter with microbial origin (Yang et al. (2015), Hudson et al.
TE

(2008), Her et al. (2003), Baker (2001), Filloux et al. (2012); Sorensen et al. (2016)). Consequently,
EP

the component C3 might be due to organic matter from infiltration of secondary effluents discharged

by the WWTPs located in the area, leaching of agricultural surface and/or biofilm associated with
C

the sediments of the resources (Simon et al., (2001), Pronk et al., (2006)).
AC

The variations of fluorophores C1 and C2 over the year of sampling exhibited similar variation as

observed for the TOC, UV254 and turbidity (Figure 4-a&b, Table S1). A high Pearson’s coefficient r,

over 0.9, was found for both fluorophores (C1 and C2) against TOC and UV254 parameters

(Table 1). Indeed, the highest values were observed during January and October 2012 for high

rainfall. Globally, the highest fluorescence signals were observed for Nouvet and Rivière.

In contrast, the C3 fluorophore showed only slight variation throughout the year and did not seem to
 ACCEPTED MANUSCRIPT
be directly impacted by rainfall (Figure 4-c). According to table 1, no clear correlation could be

drawn between C3 fluorophore and all other investigated parameters. Thus, the origin of NOM was

mainly allochtonous and component C3 accounted probably for a minor part of the organic carbon.

To conclude, the physical and chemical analysis of the five resources over year 2012 revealed that

parameters related to NOM (UV254, TOC, fluorophores C1 & C2) and particulate content (turbidity)

PT
were strongly impacted by precipitation as it is usually observed for karst area (Wittmann et al.

(2002), Pronk et al. (2006), Pianta et al., (2000)).

RI
SC
3.2 Filtration performances

As an example the results of the 24h-filtration tests obtained with the five waters sampled in May

U
2012 are reported on Figure 5-a. The variation of the transmembrane pressure showed typical
AN
variations for constant flux filtration (Peldszus et al. (2011)). As depicted, the five waters can be

classified according to their fouling ability for this month as follow : Nouvet > Rivière ≥ Erigny >
M

Foisys > Graviers, with corresponding CFV values of 168, 245, 364, 638, 1001 L m2 and IFR

values of 42.7 ; 39.7 ; 29.3 ; 14.7 ; 5.4 m-1 h-1. The reported TOC values for this month were very
D

close, the highest value being for Graviers (1.74 mg L-1) and the lowest for Foisys (1.54 mg L-1)
TE

with no clear link with the filtration performances. Indeed according to Table 1, IFR and CFV
EP

exhibited a Pearson correlation coefficient (obtained throughout the investigated year) against TOC

lower than 0.8.

C

Filtration performances of the five resources over the year and against rainfall are summarized on
AC

Figure 5-b. The highest IFR value of 55 m-1 h-1 (with a corresponding CFV equals to 196 L m-2)

was obtained for Rivière in January 2012. High monthly rainfall during fall or winter induced a

strong increase in IFR values. For example, high rainfall during October 2012 induced high IFR

values from 32 to 49 m-1 h-1 for all investigated waters. For the sake of comparison, an IFR value of

21.2 m-1 h-1 was estimated during the filtration of a surface water (DOC :5.4 mg L-1) on PVDF

hollow fiber membranes in Peiris et al. (2013). Globally, the lowest filtration performances (high
 ACCEPTED MANUSCRIPT
IFR and low CFV) were obtained for Nouvet and Rivière. Interestingly, the best filtration

performances (low IFR and high CFV) were observed during June & July 2012 even though

rainfalls were exceptionally high at this period of the year.

To conclude, the modification of organic and inorganic matrices due to rainfall events in karst area

influence the filtration performances especially during winter season when soil lixiviation is high.

PT
Moreover, the results emphasis the difficulty to link data obtained during water monitoring to

filtration performances.

RI
SC
3.3 Statistical analysis

In order to simplify data handling and to identify the most relevant parameters that could accurately

U
predict filtration performances, intensive data analysis was performed on all collected data. Firstly,
AN
all IFR and CFV values determined during the year were classified using Hierarchical Ascendant

Classification (HAC) analysis and the Ward’s method (Figure 6).

M

Clustering dendrogram (Fig.6) and the total inertia (Figure S2) show that filtration results could be

divided into three groups (i.e: most of the inertia loss was obtained for three classes). The first
D

group consisting of 13 water samples, named low fouling, regroups IFR values equal to 0 m-1 h-1
TE

and high CFV values equal to 1740 L m-2. This first group consist of filtration performances
EP

obtained during late summer 2012, Graviers and Foisys occurred 4 and 5 times in this group,

respectively. The second group, called high fouling, regroups 19 waters samples with high IFR and
C

low CFV values (with values from 24 to 55 m-1 h-1 and 130 to 364 L m-2, respectively). This second
AC

group consists mainly on filtration performances of Rivière and Nouvet waters both occurring 6

times. In addition, these high IFR values were observed during winters 2011 & 2012 and early

spring 2012. The last group consisting of 32 samples, called intermediate fouling, regroups all other

IFR and CFV values ranged from 1 to 21 m-1 h-1 and 217 to 1330 L m-2, respectively. This third

cluster regroups filtration performances observed during late spring and summer 2012. In this

cluster Erigny occurred 7 times while all other resources appeared 6 times.
 ACCEPTED MANUSCRIPT
To conclude, the HAC confirms that winter and early spring are the most problematic period

regarding ultrafiltration performances. Also, this classification help to discriminate Nouvet and

Rivière as potential resources for the drinking water process.

The principal component analysis (PCA) results obtained from all collected data (analytical and

filtration results) are plotted on figure 7. More than 57 % of the total system variance were

PT
explained by only two components allowing a 2D representation of PCA and the total system’s

variance could be explained by 10 components (Figure S3). The first component Dim1 (x-axis on

RI
fig.7) is strongly correlated to the organic matter parameters (TOC, UV254, C1, C2, C3) and

SC
turbidity of the investigated water samples. In contrast, the second component Dim2 (y-axis on

fig.7) is correlated to the inorganic matrix parameters (metal species, NO3-, etc.) of the water

U
samples. The conductivity is not well correlated to the inorganic matrix probably due to its really
AN
slight variation over the investigated year.

The PCA in Figure 7 confirms that the IFR is strongly correlated to the organic matrix of the water
M

samples. As expected, the CFV values are inversely correlated to the IFR values. The IFR is mainly

correlated with the TOC followed by the UV254 and turbidity (Table 1). The Pearson's correlation
D

coefficients for the fluorescence signals were similar whatever the considered fluorophore (ranged
TE

from 0.682 to 0.657). Table 1 demonstrates that no clear link exists between filtration parameters
EP

and one specific analytical result. Consequently, this emphasises the complexity of fouling

phenomena which consists in a combination of several factors. Globally, the CFV parameter is less
C

correlated to the analytical results than the IFR.

AC

Based on this statistical analysis, the irreversible fouling of all investigated water samples over the

year is strongly related to their organic matter matrix. These results confirm that irreversible fouling

phenomena is related to the organic matter content as reported in literature (Guigui et al. (2002),

Chen et al. (2014), Sorensen et al. (2015), Ayache et al. (2013), Pagano et al. (2014), Liu et al.

(2017)). This observation suggests that, model prediction of fouling phenomena should include

several physical and chemical parameters in order to better describe fouling event throughout long
 ACCEPTED MANUSCRIPT
time period.

In order to complete this statistical analysis, several multiple linear regressions (MLR) were done

on IFR against all investigated parameters related to the organic matrix. The best model was found

using the forward/backward method based on the Bayesian information criterion (BIC). Results

reported in table 2 show that IFR values throughout the year are well modeled by considering TOC

PT
and component C3 of the fluorescence signal. The coefficients are equal to 10.960 and 4.876 for the

TOC and C3 parameters, respectively. This prediction model shows an adjusted R2 value of 0.7825.

RI
The Student’s t-test shows that TOC’s coefficient is highly significant (Pr(>|t|) : 6.71 10-8, Table 2).

SC
In contrast, the Student’s t-test performed on the C3 coefficient returns a high value equals to 0.876,

indicating that this parameter is not enough significant to describe alone the IFR parameter.

U
In order to better describe fouling phenomena related to seasonal variation, the same methodology
AN
was applied on the two last clusters obtained from fig.6: intermediate and high fouling

corresponding to late spring/early summer and winter/early spring, respectively. The best model for
M

the intermediate fouling cluster was obtained with only the C3 parameter (Table 2). The Student’s t-

test reveals that C3’s coefficient is highly significant to describe intermediate fouling propensity
D

observed during late spring/early summer.

TE

High fouling phenomena is also well modeled by the TOC and C3 parameters with coefficients
EP

equal to 7.095 and 152.820, respectively. The Student’s t-test shows that these two coefficients are

highly significant to describe high fouling phenomena. Also this model returns a relatively high
C

adjusted R2 value equals to 0.944.

AC

To conclude, all data collected from the year of investigation confirms that the waters’ filtration

performances are strongly related to the organic matrix of the investigated karst springs.

Specifically, the MLR suggests that an increase in the protein-like substances decrease filtration

performances and combined increases of TOC and protein-like substances led to a strong fouling

event. This point is of great importance in the understanding and the validation of conclusion drawn

by many researches in literature. One another hand, only common analytical parameters were used
 ACCEPTED MANUSCRIPT
in this study. Hence, the statistical analysis could only explained fouling by the common analytical

parameters employed here (i.e TOC, fluorescence...).

Moreover, investigation of the seasonal change of the collected data might conduct to filtration

performances forecast for the coming months. As an example, the ARIMA (2,1,0) model was used

to predict the filtration performances of Foisys and Nouvet resources for the next coming 6 months

PT
(Fig. 8). Therefore, the IFR forecast for Foisys (Fig.8-a) showed high fouling events with a strong

variation of IFR during the next 6 months. In contrast, the IFR forecast for Nouvet (Fig.8-b)

RI
predicts less variation over the next 6 months with relatively high IFR values corresponding to the

SC
cluster called “high fouling”. This forecast suggests that Nouvet should not be used for the next

coming months while Foisys might have its quality strongly changing. However, as shown in

U
Figure 8, the IFR predictions exhibit high confidence interval due to the low number of data used
AN
here to process the ARIMA model (13 IFR values for each resource). Therefore more data should be

collected to reduce the confidence interval of the forecasting.

M

Finally, such approach employed for a long period can be used to predict the effect of water quality

change on filtration performances due to increase of anthropic activities or climate change. Indeed,
D

a global increase of 30-80% DOC has been observed in surface water in Northern Europe and North
TE

America since 1980s and would be linked to temperature increase and to extreme weather events
EP

leading strong soil lixiviation and increased bacterial activity (Pagano et al. (2014)) . Therefore

based on these increases and the fouling prediction model developed here, it appears that in the next
C

10 years the average IFR value of all investigated resources would increase of about 85 % (based on
AC

DOC increase of 30 % in ten years). Consequently, most of these groundwater resources would be

problematic regarding the ultrafiltration process.

4. Conclusion

The main objective of the present work was to link filtration performances to physical and chemical

properties of five underground resources situated into a karst area near Paris, France. All data were
 ACCEPTED MANUSCRIPT
computed using advanced data analysis in order to identify the most relevant parameters which

could predict membrane fouling. Whatever the investigated resources, results demonstrated a strong

relation between precipitation levels and the waters’ organic matrix. Specifically, the TOC and

fluorescence signal related to fulvic and humic like-substances were strongly correlated and

influenced by rainfall event. Interestingly, the investigated resources exhibited a fluorescence signal

PT
related to protein like-substances with no clear link with meteorological data.

Cluster analysis performed on filtration performances results allowed to classify the five resources

RI
into three groups related to their specific fouling potential (low, high and intermediate). In addition,

SC
results revealed that best filtration performances were obtained during summer. Nouvet and Rivière

were found to be the most problematic resources for ultrafiltration. PCA and MLR confirmed the

U
strong link of filtration performances and the waters’ organic carbon contents and protein like-
AN
substances fluorescence signal. Specifically, it was shown that a coupled increase of TOC and

protein fluorescence signal induced strong degradation of filtration performances with really high
M

irreversible fouling (irreversible to chlorinated permeate backwash).

To conclude, this study confirms that fluorescence analysis coupled to advanced data processing is a
D

versatile tool to membrane fouling understanding. Such methodology should be employed by

TE

membranes users to speed up decision making in order to prevent strong fouling event by a better
EP

feed water resource selection. Also, in the context of water quality degradation linked to global

warming and intensive land use, such methodology offers a unique opportunity for filtration
C

performances forecast.
AC

References

Alberts J.J., Tackacs M., 2004. Comparison of the natural fluorescence distribution among size
fractions of terrestrial fulvic and humic acids and aquatic natural organic matter. Org. Chem. 35,
1141-1149.
Amy G., 2008. Fundamental understanding of organic matter fouling of membranes. Desal. 231, 44-
51.
Ayache C., Pidou M., Croué J.P., Labanowski J., Poussade Y., Tazi-Pain A., Keller J., Gernjak W.,
 ACCEPTED MANUSCRIPT
2013. Impact of effluent organic matter on low pressure membrane fouling in tertiary treatment.
Water. Res. 47, 2633-2642.
Barbot E., Carretier E., Wyart Y., Marrot B., Moulin P., 2009. Transportable membrane process to
produce drinking water. Desal., 248, 58-63.
Baker A., 2001. Fluorescence excitation–emission matrix characterization of some sewage-
impacted rivers. Environ. Sci. Technol., 35 (5), 948–953.
Baker A., Genty D., 1999. Fluorescence wavelength and intensity variations of cave waters. J. Hydr.

PT
217, 19-34.
Bessiere Y., Guigui C., Remize P.J., Cabassud C., 2005. Low fouling conditions in dead-end

RI
filtration: Evidence for a critical filtered volume and interpretation using critical osmotic pressure.
J. Membr. Sci. 264, 37-47.

SC
Campinas M., Rosas M.J., 2010.Assessing PAC contribution to the NOM fouling control in
PAC/UF systems. Water Res., 44, 1636-1644.

U
Chen H., Kim A.S., 2006. Prediction of permeate flux decline in crossflow membrane filtration of
colloidal suspension: a radial basis function neural network approach. Desal., 192, 415-428.
AN
Chen F., Peldzus S., Peiris R.H., Ruhl A.S., Mehrez R., Jekel M., Legee R.L., Huck P.M., 2014.
Pilot scale investigation of drinking water ultrafiltration membrane fouling rates using advanced
M

data analysis techniques. Water Res. 48, 508-518.

Cruz M. C., Romero L. C., Vicente M. S., Rajal V. B., 2017. Statistical approaches to understanding
D

the impact of matrix composition on the disinfection of water by ultrafiltration, Chemical

Engineering Journal, 316, 305-314.
TE

Curcio S., Calabrò V., Lorio G., 2006. Reduction and control of flux decline in cross-flow
membrane processes modeled by artificial neural networks. J. Membr. Sci., 286, 125-132.
EP

Filloux E., Labanowski J., Croué J.P., 2012. Understanding the fouling of UF/MF hollow fibers of
biologically treated wastewaters using advanced EfOM characterization and statistical tools. Bio.
C

Tech. 118, 460-468.

Ford D. C., Williams P. W., 2007. Karst Hydrogeology and Geomorphology, Wiley, Chichester.
AC

Gao W., Liang H., Ma J., Han M., Chen Z.L., Han Z.S., Li G.B., 2011. Membrane fouling control in
ultrafiltration technology for drinking water production: A review. Desal., 272 1-8.
Guigui C., Rouch J.C., Durand-Bourlier L., Bonnelye V., Aptel P., 2002. Impact of coagulation
conditions on the in-line coagulation/UF process for drinking water production. Desal. 147, 95-100.
Hamachi M., Cabassud M., Davin A., Peuchot M.M., 1999. Dynamic modelling of crossflow
microfiltration of bentonite suspension using recurrent neural networks. Chemical Engineering and
Processing: Process Intensification 38, 203-210.
Hartmann A., Golscheider N., Wagener T., Lange J., Weiler M., 2014. Karst water resources in a
 ACCEPTED MANUSCRIPT
changing world: Review of hydrological modeling approaches. Rev. Geophys. 52, 3, 218-242.
Her N., Amy G., McKnight D., Shon J., Yoon Y., 2003. Characterization of DOM as a function of
MW by fluorescence EEM and HPLC-SEC using UVA, DOC and fluorescence detection. Water.
Res. 37, 4295-4303.
Hong S., Elimelech M., 1997. Chemical and physical aspects of organic matter (NOM) fouling of
nanofiltration membranes. J. Membr. Sci., 132, 159-181.
Huang W., Chu H., Dong B., Liu J., 2014. Evaluation of different algogenic organic matters on the

PT
fouling of microfiltration membranes. Desal., 344, 329–338.
Hudson N., Baker A., Ward D., Reynolds D.M., Brundson C., Carliell-Marquet C., Browning S.,

RI
2008. Can fluorescence spectroscopy be used as a surrogate for the biochemical Oxygen Demand
(BOD) test in water quality assessment ? An example from South West England. Sci. Total. Env.

SC
391, 149-158.
Husson F., Josse J., Lê S., 2008. FactoMineR: An R Package for Multivariate Analysis. J. Stat. Soft.

U
25, 1-18.
Ishii S.K.L., Boyer T.H., 2012. Behavior of reoccurring PARAFAC Components in Fluorescent
AN
dissolved organic matter in natural and engineering systems : A critical review. Env. Sci. & Tech.
46, 2006-2017.
M

Lee N., Amy G., Croué J.P., 2006. Low pressure membrane (MF-UF) fouling associated with
allochtonous versus natural organic matter. Water Res., 40, 2357-2368.
D

Liu, J., Wang, Z., Cao, B., Zhao, D., Chu, H., 2017. Ultrafiltration of distinct natural waters:
TE

Correlation of fouling resistances with water constituents, Desalin. water treat. 68, 40-48.

Mao R., Wang Y., Zhao Y., Gao B., Dong M., 2013. Impact of various coagulation technologies on
EP

membrane fouling in coagulation/ultrafiltration process, Chemical Engineering Journal, 225, 387-

393.
Nakatsuka S., Nakate I., Miyano T., 1996. Drinking water treatment by using ultrafiltration hollow
C

fiber membranes. Desal., 106, 55-61.

AC

Pagano T., Bida M., Kenny J.E., 2014. Trends in Levels of Allochthonous Dissolved Organic
Carbon in Natural Water: A Review of Potential Mechanisms under a Changing Climate. Water 6,
2862-2897.
Peiris R.H., Budman H., Legge R.L. , Moresoli C., 2011. Assessing irreversible fouling behavior of
membrane foulants in the ultrafiltration of natural water using principal component analysis of
fluorescence excitation-emission matrices Water Sci. Technol. Water Supply, 11 (2), 179–185
Peiris R.H., Hallé C., Budman H., Moresoli C., Peldszus S., Huck P.M., Legge R.L., 2010a.
Identifying fouling events in a membrane-based drinking water treatment process using principal
 ACCEPTED MANUSCRIPT
component analysis of fluorescence excitation-emission matrices Water Res., 44 (1), 185–194
Peiris R.H., Budman H., Moresoli C., Legge R.L., 2010b. Understanding fouling behavior of
ultrafiltration membrane processes and natural water using principal component analysis of
fluorescence excitation-emission matrices. J. Memb. Sci., 357 (1–2), 62–72
Peiris R.H., Ignagni N., Budman H., Moresoli C., Legge R.L., 2012. Characterizing natural
colloidal/particulate-protein interactions using fluorescence-based techniques and principal
component analysis. Talanta, 99, 457–463.

PT
Peldszus S., Hallé C., Peiris R.H., Hamouda M., Jin X., Legge R.L., Budman H., Moresoli C.,
Huck P.M., 2011. Reversible and irreversible low-pressure membrane foulants in drinking water

RI
treatment: Identification by principal component analysis of fluorescence EEM and mitigation by
biofiltration pretreatment. Water Res. 45, 5161-5170.

SC
Pianta R., Boller M., Janex M.L, Chappaz A., Birou B., Ponce R., Walther J.L, Micro- and
ultrafiltration of karstic spring water, 1998. Desal., 117(1–3),61-71.

U
Pianta R., Boller M., Urfer D., Chappaz A., Gmünder a., Costs of conventional vs. membrane
treatment for karstic spring water, 2000. Desal.,131(1–3), 245-255.
AN
Pronk M., Goldscheider N., Zopfi J., 2006. Dynamics and interaction of organic carbon, turbidity
and bacteria in a karst aquifer system. Hydrogeol. J. 14 (4), 473-484.
M

Shamsuddin N., Das D.B., Starov V.M., 2015. Filtration of natural organic matter using
ultrafiltration membranes for drinking water purposes: Circular cross-flow compared with stirred
D

dead end flow, Chemical Engineering Journal, 276, 331-339

Simon K.S., Gibert J., Petitot P., Laurent R., 2001. Spatial and temporal patterns of bacterial density
TE

and metabolic activity in a karst aquifer. Arch Hydrobiol. 151:67–82.

Sorensen J.P., Lapworth D.J., Marchant B.P., Nkhuwa D.C. , Pedley S., Stuart M.E., Bell R.A. ,
EP

Chirwa M., Kabika J., Liemisa M., Chibesa M., 2015.. In-situ tryptophan-like fluorescence: a real-
time indicator of faecal contamination in drinking water supplies. Water Res., 81, 38–46
C

Sorensen J.P.R., Sadhu A., Sampath G., Sugden S., Dutta Gupta S., Lapworth D.J., Marchant B.P.,
Pedley S., 2016. Are sanitation interventions a threat to drinking water supplies in rural India? An
AC

application of tryptophan-like fluorescence. Water Res., 88, 923–932

Stedmon C.A. , Markager S. , 2003. Behaviour of the optical properties of coloured dissolved
organic matter under conservative mixing. Estuarine, coastal and Shelf Science 57, 973-979.
Touffet A., Baron J., Welte B., Joyeux M., Teychene B., Gallard H., 2015. Impact of pretreatment
conditions and chemical ageing on ultrafiltration membrane performances. Diagnostic of a
coagulation/adsorption/filtration process. J. Membr. Sci. 489, 284-291
Touffet A., Study of the performances of Ultrafiltration units : Impact of water quality, water
pretreatment and chemical cleaning. (Etude des performances d’unités d’ultrafiltration : Impact de
 ACCEPTED MANUSCRIPT
la qualité de la ressource, des prétraitements et des lavages chimiques). Thèse université de
Poitiers, 2014.
Vera M., Martín-Alonso J., Mesa J., Granados M., Beltran J.L., Casas S., Gibert O., Cortina J.L.,
2017. Monitoring UF membrane performance treating surface-groundwater blends: Limitations of
FEEM-PARAFAC on the assessment of the organic matter role. Chem. Eng. J. 317, 961-971.
Virtanen, T., Reinikainen, S.-P., Kögler, M., Viitala, T., Kallioinen, M., 2017. Real-time fouling

monitoring with Raman spectroscopy, J. Membr. Sci. 525, 312-319

PT
Wetterau G.E., Clark M.M., Anselme C., 1996. A dynamic model for predicting fouling effects

RI
during the ultrafiltration of a groundwater. J. Membr. Sci. 109, 185-204.
Wittmann E., Tazi-Pain A., Chanussot T., Patterson A., Niay R., Gaïd A., Canto F., 2002.

SC
Clarification of a highly turbid karstic water by microfiltration, Desal., 145 (1), 309-313.
Yao X., Zhang Y., Zhu G. , Qin B., Feng L., Cai L., Gao G., 2011. Resolving the variability of
CDOM fluorescence to differentiate the sources and fate of DOM in lake Taihu and its tributaries.

U
Chemosphere 82, 145-155.
AN
Yang L., Hur J., Zhuang W., 2015. Occurrence and behaviors of fluorescence EEM-PARAFAC
components in drinking water and wastewater treatment systems and their applications: a review
M

Environ. Sci. Pollut. Res., 1–11.

Zheng X., Ernst M., Jekel M., 2009. Identification and quantification of major organic foulants in
treated domestic wastewater affecting filterability in dead-end ultrafiltration. Water Res., 43 (1),
D

238-244.
TE

Zheng X., Zietzschmann F., Plume S., Paar H., Ernst M., Wang Z., Jekel M., 2017. Understanding
and control of biopolymer fouling in ultrafiltration of different water types. Water., 9, 298.
C EP
AC
 ACCEPTED MANUSCRIPT

Figure 1: Location of the 5 groundwater resources near the Dreux district (100 km East from Paris,

48°42'56.9"N 0°58'08.8"E). Color map corresponds to the sediment flood risk. Maps obtained from BRGM

website (Infoterre) and the French Ministry of Ecology, Sustainable Development and Energy

[http://assainissement.developpement-durable.gouv.fr/ & http://infoterre.brgm.fr/].

Figure 2. TOC variations of the 5 sources throughout the year of study (Average: 1.5 mg L-1, Standard deviation:

PT
0.86 mg L-1). Monthly rainfall (red line, data obtained from Météo France, Paris).

RI
Figure 3: Picture of the three components obtained from PARAFAC analysis of all FEEM. Left : C1 – Humic-

like substances. Center : C2 – Fulvic-like substances. Right : C3-Protein-like substances. *

SC
Figure 4 : Fluorescence signal variation over the year of investigation of all water resources. (a) C1 component

– Humic-like substances. (b) C2 component – Fulvic-like substances. (c) C3 component – Protein-like

U
substances. Monthly rainfall (red line, data obtained from Météo France Paris).
AN
Figure 5: (a) TMP variation during constant flux filtration of water resources sampled during May 2012 - Line

represents the average TMP values. (b) IFR (upper axis) and CFV (lower axis) values over the year obtained on
M

filtration tests performed on the five resources (Left). Precipitation level (Right).

Figure 6: Cluster analysis using HAC of all investigated water samples based on filtration performances (IFR
D

and CFV). Sample nomenclature (resources-month-year).

TE

Figure 7: Principal component analysis performed on all collected data over the year of investigation. Black

arrow represents the projection of physicochemical results. Dotted blue arrows represent the projection of the
EP

filtration performances parameters.

Figure 8 : IFR (13 data points for each resource) and ARIMA(2,1,0) forecast for Foisys (a) and Nouvet (b) over
C

the year of investigation and the six coming months.

AC
 ACCEPTED MANUSCRIPT
Table 1 : Pearson correlation matrix obtained on the organic matrix of water samples.

C1 C2 C3
TOC UV254 Turbidity IFR CFV
(Humic) (Fulvic) (Protein like)

TOC 1

UV254 0.952 1

PT
Turbidity 0.785 0.799 1

C1 0.904 0.938 0.773 1

RI
C2 0.944 0.937 0.767 0.957 1

SC
C3 0.559 0.526 0.476 0.426 0.460 1

IFR 0.771 0.754 0.694 0.682 0.676 0.657 1

U
VFC -0.564 -0.563 -0.567 -0.480 -0.508 -0.652 -0.787 1
AN
Table 2: Multiple linear regression models obtained from forward/backward stepwise method
M

and BIC criterion.

D

Coefficients
Adjusted
TE

Cluster Model STD.

R2 Parameters t value Pr(>|t|)
Error*

TOC 1.788 6.130 6.71 10-8

EP

Global IFR data IFR = 10.960 TOC + 4.876 C3 0.7825

C3 31.113 0.157 0.876

Intermediate fouling IFR = 106.436 C3 0.7862 C3 9.926 10.72 8.78 10-12

C

TOC 2.2 3.225 0.00469

AC

High fouling IFR = 7.095 TOC + 152.820 C3 0.944

C3 46.958 3.254 0.00440

*Standard error
 ACCEPTED MANUSCRIPT

PT
RI
U SC
AN
M
D
TE
EP
C
AC
 ACCEPTED MANUSCRIPT

PT
RI
U SC
AN
M
D
TE
EP
C
AC
 ACCEPTED MANUSCRIPT

PT
RI
U SC
AN
M
D
TE
EP
C
AC
 ACCEPTED MANUSCRIPT

PT
RI
U SC
AN
M
D
TE
EP
C
AC
 ACCEPTED MANUSCRIPT

PT
RI
U SC
AN
M
D
TE
EP
C
AC
 ACCEPTED MANUSCRIPT

PT
RI
U SC
AN
M
D
TE
EP
C
AC
 ACCEPTED MANUSCRIPT

PT
RI
U SC
AN
M
D
TE
EP
C
AC
 ACCEPTED MANUSCRIPT

PT
RI
U SC
AN
M
D
TE
EP
C
AC