Applied Soft Computing 11 (2011) 3238–3246

Contents lists available at ScienceDirect

Applied Soft Computing

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

Prediction of effluent quality of a paper mill wastewater treatment using an

adaptive network-based fuzzy inference system
Jinquan Wan b , Mingzhi Huang a,c,∗ , Yongwen Ma a , Wenjie Guo a , Yan Wang a ,
Huiping Zhang c , Weijiang Li a , Xiaofei Sun a
a
College of Environmental Science and Engineering, South China University of Technology, Guangzhou 510640, People’s Republic of China
b
State Key Laboratory of Pulp and Paper Engineering, South China University of Technology, Guangzhou 510640, People’s Republic of China
c
School of Chemistry and Chemical Engineering, South China University of Technology, Guangzhou 510640, People’s Republic of China

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

Article history: Advanced neuro-fuzzy modeling, namely an adaptive network-based fuzzy inference system (ANFIS), was
Received 2 February 2010 employed to develop models for the prediction of suspended solids (SS) and chemical oxygen demand
Received in revised form 4 November 2010 (COD) removal of a full-scale wastewater treatment plant treating process wastewaters from a paper mill.
Accepted 13 December 2010
In order to improve the network performance, fuzzy subtractive clustering was used to identify model’s
Available online 23 December 2010
architecture and optimize fuzzy rule, meanwhile principal component analysis (PCA) was applied to
reduce the input variable dimensionality. Input variables were reduced from six to four for COD and SS
Keywords:
models, by considering PCA results and linear correlation matrices among input and output variables.
Adaptive network-based fuzzy inference
system
The results indicate that reasonable forecasting and control performances have been achieved through
Wastewater treatment the developed system. The minimum mean absolute percentage errors of 1.003% and 0.5161% for CODeff
Prediction and SSeff could be achieved using ANFIS. The maximum correlation coefficient values for CODeff and SSeff
Principal component analysis were 0.9912 and 0.9882, respectively. The minimum mean square errors of 1.2883 and 0.0342, and the
minimum RMSEs of 1.135 and 0.1849 for CODeff and SSeff could also be achieved.
Crown Copyright © 2010 Published by Elsevier B.V. All rights reserved.

1. Introduction for improving heuristic understanding of the system under study

[1,2].
A human operator with good field operating experience is gener- Intelligence control, such as neural networks and fuzzy systems,
ally able to manage and operate a biological wastewater treatment have been proved to be able to model nonlinear systems and suc-
plant satisfactorily using apparently few, target-driven operational cessfully applied for modeling various non-linear processes [3].
observations. To increase safety and improve operational per- The most significant advantage of intelligence control is that no
formance of wastewater treatment processes, it is important to precise mathematical model is needed, which can well approach
develop computer operational decision support systems that are any nonlinear continuous function and overcome the shortcomings
able to play a similar role to the expert in daily operation. Such intel- of traditional control that over depend on accurate mathematical
ligent computing systems are normally designed to exhibit expert model. Recently, the use of neural networks is also gaining popu-
level reasoning within a narrowly defined domain. The heuristic larity in modeling biological wastewater treatment processes [4].
knowledge or rules of thumb used in an intelligent system is gener- Neural networks can map a set of input patterns onto a correspond-
ally implicit, non-algorithmic and subjective. It is a general practice ing set of output patterns after learning a series of past process
that these rules are acquired through domain experience and used data from a given system. Moreover, a neural network model has
distinctive ability of learning non-linear functional relationships
without requirement of the structural knowledge of the process to
be modeled [5]. A novel approach on the basis of the ANN model that
Abbreviations: ANFIS, adaptive network-based fuzzy inference system; ANN,
was designed to provide better predictions of nitrogen contents in
artificial neural network; FLC, fuzzy logic control; WWTP, wastewater treatment
plant; SS, suspended solids; COD, chemical oxygen demand; BOD, biochemical oxy- treated effluents was reported by Chen et al. [6]. Total suspended
gen demand; PCA, principal component analysis; R, correlation coefficient; MSE, solid (TSS) is an indication of plant performance. A simple predic-
mean square error; RMSE, root mean square error; MAPE, mean absolute percentage tion models based on neural network for TSS was demonstrated
error; MF, membership function; DO, dissolved oxygen; Eq., equation. by Belanche et al. [7]. Serodes et al. [8] developed a decision sup-
∗ Corresponding author at: College of Environmental Science and Engineering,
port tool, Chlorocast, based on neural networks for modeling the
South China University of Technology, Guangzhou 510640, People’s Republic of
China. Tel.: +86 137 60624592; fax: +86 020 87114970. chlorination process in the final disinfection phase of a water treat-
E-mail address: [email protected] (M. Huang). ment system. They illustrated the power of the developed tool by

1568-4946/$ – see front matter. Crown Copyright © 2010 Published by Elsevier B.V. All rights reserved.
doi:10.1016/j.asoc.2010.12.026
 J. Wan et al. / Applied Soft Computing 11 (2011) 3238–3246 3239

Fig. 1. Process of wastewater treatment in the paper mill.

applying it to forecast the residual chlorine in the drinking water online process monitoring over the last few decades [17]. The pur-
tank and distribution system of the city of Sainte-Foy. Holubar pose of PCA is to identify linear correlations between correlated
et al. [9] applied neural networks based on the feed forward neu- variables aiming at data dimensionality reduction. The method
ral network trained with the back propagation algorithm to model generates a new set of variables, called principal components.
and control methane production in anaerobic digesters. The model Each principal component is a linear combination of the origi-
was trained using data generated from four anaerobic continuous nal variables. All the principal components are orthogonal to each
stirred tank reactors operating at steady state. other, so there is no redundant information. The principal com-
Although ANN can predict the effluent from WWTPs suc- ponents as a whole form an orthogonal basis for the space of the
cessfully, traditional neural network schemes still have several data [18]. In recent years, the PCA approach has been successfully
limitations which are resulted from possibility of getting trapped applied to wastewater treatment process monitoring and control
in local minimum, and the choice of model architecture. If the [19–22].
predicting performance can be further promoted, better operation Taking the wastewater treatment project for the DongGuan
strategy can be formed. To overcome these limitations of traditional Papermaking as a case, the main objective of this work was to eval-
ANNs, and to increase their reliability, many new training algo- uate the ANFIS modeling as a valid input–output model to predict
rithms have been proposed such as adaptive neuro-fuzzy inference the treatment performance of a paper mill wastewater treatment.
system (ANFIS) [10]. ANFIS combines fuzzy logic control (FLC) with In order to improve the network performance, fuzzy subtractive
artificial neural network (ANN) and realizes fuzzy logic by fuzzy clustering and principal component analysis (PCA) were used to
neural network. Meanwhile, the controller can get hold of fuzzy identify model’s architecture and optimize fuzzy rule. Meanwhile
rules and optimize its subjection function online by self-learning for comparison, ANN was also employed to predict the effluent in
ability of the neural network. Application of fuzzy neural network this study.
in wastewater treatment, it can acquire better effect.
Recently, active researches have been carried out in fuzzy- 2. Materials and methods
neural control [11–13]. Tay and Zhang [14] integrated fuzzy
systems and neural networks in modeling the complex process of 2.1. Wastewater treatment plant
anaerobic biological treatment of wastewater. They illustrated the
power of the technique in two case studies of upflow anaerobic A paper mill wastewater treatment plant (Fig. 1), located at the
sludge blanket and anaerobic fluidized bed reactor. The fuzzy- Dongguan city Guangdong province, was used as a demonstration
neural model simulated the system performance well and provided site for assessing the application of this hybrid fuzzy controller.
satisfactory prediction results based on observed past information, In this WWTP, the treatment processes were comprised of bar
although a disadvantage of the model was its high dependence on rack, equalization tank, primary settling tank, high efficient reactor
the quality of the training data. Steyer et al. [15] used the fuzzy (researched and developed by South China University of Technol-
logic and the artificial neural networks to on-line examine and ogy), hydrolysis and anaerobic tanks, aerated submerged biofilm,
analyze the question which appeared processing the 120 L grape secondary settling tank. The flow rate was 30,000 cubic meters
wine wastewater in the anaerobic digestion fluidized bed reactor. per day (CMD). The annual wastewater discharge amount from
Primary data like pH, the temperature, the backflow quantity, the the mill was 1.812 × 107 tons: They include 23696.56 tons of COD;
flow of the influent. According to the fuzzy logic which may dis- 3695.48 tons of BOD; 6.04 tons of volatile phenol.
tinguish, the characteristic vector was divided into the appointed The monitoring and control system is based on probes from
category. Then the process condition was classified by the artificial HACH® , cards and interfaces from Advantech® . The plant is
neural networks. Chen and Chang [16] integrated fuzzy systems equipped with DO-temperature (D53)) and pH (DRD1P5) probes,
and neural networks in modeling the complex process of aeration and COD (CODmax ) and NH4 + (Amtax compact) on-line monitoring
in a submerged biofilm wastewater treatment process. instrument. The signals, filtered in a transmitter, are captured by
As a multivariate statistical data analysis technique, principal a data acquisition card (ADAM4017, Advantech, China). The con-
component analysis (PCA) has become increasingly popular for trol is conducted using a power relay output board (ADAM4024,
3240 J. Wan et al. / Applied Soft Computing 11 (2011) 3238–3246

which represents the firing strength of each rule. The firing strength
means the degree to which the antecedent part of the rule is satis-
fied. 
Layer 3: The nodes in this layer are also fixed nodes labeled ,
indicating that they play a normalization role in the network. The
outputs of this layer can be represented as
wij
o3ij = wij = , i, j = 1, 2 (4)
w11 + w12 + w21 + w 22
which are called normalized firing strengths.
Layer 4: Each node in this layer is an adaptive node, whose out-
put is simply the product of the normalized firing strength and a
Fig. 2. ANFIS structure for a two-input Sugeno model with four rules. first-order polynomial (for a first-order Sugeno model). Thus, the
outputs of this layer are given by Eq. (5).

Advantech, China) which allowed an optimal equipment function- o4ij = wij fij = wij (pij x + qji y + rij ), i, j = 1, 2 (5)
ing. The software consisted of user-friendly interfaces and was able
Parameters in this layer are referred to as consequent parame-
to repeat over time a previously defined operation cycle by control-
ters.
ling pumps, mixing device and air supply. All chemical analytical
methods used in this study were according to standard method.  Layer 5: The single node in this layer is a fixed node labeled
, which computes the overall output as the summation of all
incoming signals, i.e.,
2.2. Adaptive neuro-fuzzy inference system (ANFIS)

2 
2

2 
2

ANFIS is a multilayer feed-forward network which uses neural z = O15 = Wij fij = Wij (pij x + qij y + rij )
network learning algorithms and fuzzy reasoning to map inputs i=1 j=1 i=1 j=1
into an output. It is a fuzzy inference system (FIS) implemented  
in the framework of adaptive neural networks. Fig. 2 shows the = (Wij x)pij + (Wij y)qij + (Wij )rij (6)
architecture of a typical ANFIS with two inputs, two rules and one
output for the first-order Sugeno fuzzy model, where each input is which is a linear combination of the consequent parameters when
assumed to have two associated membership functions (MFs). the values of the premise parameters are fixed. It can be observed
For a first-order Sugeno fuzzy model [23], a typical rule set with that the ANFIS architecture has two adaptive layers: Layers 1 and 4.
four fuzzy if–then rules can be expressed as Layer 1 has modifiable parameters {ai , bi , ci } and {aj , bj , cj } related
to the input MFs. Layer 4 has modifiable parameters {pij , qij , rij }
Rule 1: If x is A1 and y is B1 then f1 = p1 x + q1 y + r1 pertaining to the first-order polynomial. The task of the learning
Rule 2: If x is A2 and y is B2 then f2 = p2 x + q2 y + r2 algorithm for this ANFIS architecture is to tune all the modifiable
where A1 , A2 , B1 and B2 are the MFs for the inputs x and y, respec- parameters to make the ANFIS output match the training data.
tively, pij , qij and rij (i, j = 1, 2) are consequent parameters [10,24]. Learning or adjusting these modifiable parameters is a two-step
process, which is known as the hybrid learning algorithm. In the
As can be seen from Fig. 2, the architecture of a typical ANFIS forward pass of the hybrid learning algorithm, the premise param-
consists of five layers, which perform different actions in the ANFIS eters are hold fixed, node outputs go forward until Layer 4 and the
and are detailed below. consequent parameters are identified by the least squares method.
Layer 1: All the nodes in this layer are adaptive nodes. They gen- In the backward pass, the consequent parameters are held fixed, the
erate membership grades of the inputs. The outputs of this layer error signals propagate backward and the premise parameters are
are given by updated by the gradient descent method. The detailed algorithm
and mathematical background of the hybrid learning algorithm can
o1Ai = uAi (x) i = 1, 2 be found in Jang [10].
(1)
o1Bj = uBj (x) j = 1, 2
2.3. The index
where x and y are crisp inputs, and Ai and Bj are fuzzy sets such as
low, medium, high characterized by appropriate MFs, which could Mean square error (MSE), root mean square normalized error
be triangular, trapezoidal, Gaussian function or other shapes. In this (RMSE), mean absolute percentage error (MAPE) and correlation
study, the generalized bell-shaped MFs (Eq. (2)) defined below are coefficient (R) are used as a performance index to evaluate the
utilized prediction capability of ANFIS and ANN trained by each data set.
1 The MSE performance index was defined as
uAi (x) = i = 1, 2
1
2bi n
1 + ((x − ci )/ai )
(2) MSE = (ŷ − y)
2
(7)
1 n
uBj (x) = j = 1, 2
1 + ((x − cj )/aj ))2bj i=1

The RMSE performance index was defined as

where {ai , bi , ci } and {aj , bj , cj } are the parameters of the MFs,

governing the bell-shaped functions. Parameters in this layer are n 2
i=1
(ŷ − y)
referred to as premise parameters or antecedent parameters. RMSE = (8)
Layer 2: The nodes in this layer are fixed nodes labeled 2, indi- n
cating that they perform as a simple multiplier. The outputs of this where y is the measured values, ŷ the corresponding predicted val-
layer are represented as ues and n is the number of samples. It could be noted that the
validation error of the abnormal batch operations was much higher
o2ij = wij = uAi (x)uBj (y), i, j = 1, 2 (3)
than the training error for both the anaerobic and aerobic model,
 J. Wan et al. / Applied Soft Computing 11 (2011) 3238–3246 3241

Fig. 4. Variations of wastewater parameters measured for the ANFIS and ANN mod-
els.

tain adequate and correct information on the system. On the other

hand, it is common for a raw database to contain some redundant
and conflicting data. Thus, sometimes it is necessary for the raw
training database to be pretreated to remove redundancies and
resolve conflicts in the data. Since the ANFIS is usually started with
a prototype fuzzy system, a fuzzy system generator is needed. The
software MATLAB (Matworks Inc.) provides this function. Tay and
Zhang [25] used MATLAB for model programming and they proved
that the MATLAB language is suitable for the programming of the
Fig. 3. Architecture of the neural fuzzy model for anoxic/oxic process.
model.

whereas the validation error of the normal batch operations were 4. Results and discussion
comparable to the training error.
Mean absolute percentage error (MAPE): 4.1. Data collection and preprocessing
 
1   At − Ft 
n
MAPE =
n
 A
 × 100 (9) The data from 15th of March 2006 to 21st of December 2006
t were obtained from the plant and used to develop two separate
i=1
N N ANFIS models, as shown in Fig. 4. They were collected from the
where Ā = N1 A and F̄ = N1
t=1 t
F are the average values of At
t=1 t
treatment plant to form daily composite samples for analysis and
and Ft over the training or testing dataset. The smaller RMSE and investigated every two to three days and their total numbers were
MAPE mean better performance. 150. Among the total numbers of data, the numbers for training
Correlation coefficient (R): and testing (predicting) were 120 and 30, respectively. Also there
n was a validation data (see Fig. 4), whose number is 20. In China,
(At − Ā)(Ft − F̄)
R=  t=1
(10) the effluent regulation limits of SSeff and CODeff were 30 mg/L and
n 2 N 2 100 mg/L, respectively. The effluent quality from this WWTP met
(A − Ā) ·
t=1 t
(F − F̄)
t=1 t the Effluent Standard of China.
N N The main objective of the data preprocessing is to determine the
where Ā = N1 A and F̄ = N1
t=1 t
F are the average values of At
t=1 t suitable locations for the data acquisition required for the model-
and Ft over the training or testing dataset.
ing activities. This is a standard procedure for the networks data
preparation. The main objective here is to ensure that the statisti-
3. Model architecture and model components cal distribution of the values for the net input and output is roughly
uniform. The data sets are often scale so that they always fall within
The schematic architecture of the neural fuzzy model is depicted a specified range or they are normalized so that they have zero
in Fig. 3. It consists of the five key components: inputs and out- mean and unitary variance. These data were normalized by
puts database and preprocessor, a fuzzy system generator, a fuzzy
inference system, and, an adaptive neural network representing s(i) − min(s)
S(i) = (11)
the fuzzy system. The fuzzy inference system and its associated max(s) − min(s)
adaptive network are a Sugeno fuzzy inference system and an adap-
tive network-based fuzzy inference system (ANFIS) [10]. Input and 4.2. Analysis of historical process data
output variables are selected or generated from the variables com-
monly used for system description. A database that contains system PCA was performed to clarify and evaluate the relation-
performance information is a prerequisite for model development. ships among model variables. The percentage of process variance
Generally, it is developed by collecting regularly monitored param- explained as a function of the number of principal components is
eters. The quality of the training database is critical for the model shown in Fig. 5(a) and Tables 1 and 2. As can be noted from this
to produce correct information about the system. In order for the figure, six PCs were extracted from the PCA. The transformation
model to describe the system accurately, the database should con- matrix of PC was defined as: pcs = [pcs1 pcs2 . . . pcs6] X. So the
3242 J. Wan et al. / Applied Soft Computing 11 (2011) 3238–3246

As showed in Table 1, over 81% of the variation within the data

could be explained by the former four PCs, so the importance of the
six variables for the former four PCs was based on the following
order:

PCs1: 6 > 1> 4 > 5 > 3 > 2;

PCs2: 2 > 5 > 3 > 4 > 6 > 1;
PCs3: 3 > 1 > 2 > 5 > 4 > 6;
PCs4: 5 > 2 > 4 > 6 > 1 > 3;

From this, it can be seen that 1, 2, 3 5 and 6 were the importance

variables.
In addition, over 71% of the variation within the data could be
explained by three PCs for COD and SS models. The projections of
the process vectors into the space of the first three loading vectors
are shown in Fig. 5(b) and (c), respectively. Based on the relative
position of the model input, a lumping of variables was observed,
i.e., within Q and SSinf (see Fig. 5(b)). Similarly, CODinf and SSinf
can be considered as a separate group (see Fig. 5(c)). Grouping of
the variables indicates strong correlations among these variables, in
other words, they similarly contribute to score patterns. Lumping of
the process variables (DO, CODinf , SSinf , pH, Q and T) is also useful to
understand the behavior of the processes and make decision on the
input variable matrix which might be representative. So from Fig. 5
and the coefficients of each Pcs, it can be seen that the importance
of the six variables for output was based on the following order:
DO > CODinf > SSinf > pH > Q > T.
Based on the PCA results and the correlation coefficients pre-
sented in Table 2, different input variable matrices were developed
for the two independent ANFIS models. The selection criteria of
these variables were the ability of the variable to represent its
group and the ease and practicality of the variable measurement
in the WWTP. For the COD and SS models, the variables (pH, T and
DO) having negative correlations with the output variable were
excluded from the input variable matrix. Concentrations of COD and
SS in the tanks strongly correlated with the output variables; there-
fore, they were included as input variables. Since DO is an important
Fig. 5. Percentage of process variance explained as a function of number of principal
variable for the plant control, it was included in the input matrix.
components and loading vectors for COD and SS models.
Overall, Input variables were reduced from six to four for COD and
SS ANFIS models, by considering PCA results and linear correlation
Table 1
The percentage of principle component variances. matrix. After reduction in input variables, two independent ANFIS
models were developed by using the selected input variables (DO,
Pcs 1 2 3 4 5 6
CODinf , SSinf , and Q).
Variance explained 31.48 22.77 15.94 11.68 10.53 7.6
percent (%) 4.3. Determination of an appropriate ANFIS and ANN model
Variance cumulative 31.48 54.25 70.19 81.87 92.40 100
percent (%)
4.3.1. Clustering of input and output variable values
In this study the training data were analyzed by fuzzy subtrac-
coefficients of each Pcs were got as follow: tive clustering function in Matlab software, and were divided into
16 clusters and 15 clusters for ANFISCOD model and ANDIFSSS
Pcs1 = [−0.5555 −0.0096 0.1350 −0.5372 0.1468 0.6025], model, respectively. Thus each cluster represented a rule, the rules
Pcs2 = [0.0967 −0.6661 −0.2728 −0.1831 −0.6488 0.1344], corresponding to Eq. (3) may be stated as:
Pcs3 = [−0.699 0.2699 −0.9452 0.0818 0.1421 0.0442],
Pcs4 = [−0.0126 0.6508 0.1092 −0.3293 −0.6560 −0.1595], Ri : If x is Ai and v is Bi and . . . and y is Fi and y2 is Hi , then z = f(x, v,
Pcs5 = [−0.7182 0.0165 −0.0412 0.6312 −0.2888 −0.0197], . . ., y), i = 1, 2, . . ., 16.
Pcs6 = [0.4014 0.2439 0.0178 0.4054 −0.1532 0.7689]. where A, B, . . ., G and H are the fuzzy sets in the antecedents and
z = f(x, v, . . ., y) is a crisp function in the consequent. The cluster cen-
Table 2
ters represent the initial value of premise parameter in function
Correlation coefficients between the measured input and output variables for each (3).
model.

Output Input variables

4.3.2. Development of the ANFIS and ANN model
variables In this paper, the ANFIS model was built by the function of getfis2
in the Matlb. Firstly the data was analyzed by fuzzy subtractive clus-
CODinf Q pH SSinf T DO
tering function in Matlab, the cluster centers were acquired. And
CODeff 0.79810.0995 −0.1447 0.2781 −0.1358 −0.4117 the cluster centers represent the initial value of premise parame-
SSeff 0.43120.0308 0.0084 0.6708 −0.0930 −0.3863
ters, so the initializing parameters were determined.
 J. Wan et al. / Applied Soft Computing 11 (2011) 3238–3246 3243

Table 3
Determination of the appropriate ANFIS and ANN models.

ANFIS ANN

Basic structure Basic structure

No. of total layers 5 No. of total layers 3
No. of layers excepting 3 No. of hidden layers 1
input and output layer
No. of nodes input 4 No. of neurons in input 4
layers layers
No. of nodes in output 1 No. of neurons in 1
layers output layers
CODeff CODeff
Shape of MFs Bell Speed of training 0.1
No. of training 200 No. of neurons in 10
hidden layers
No. of fuzzy rules 16 No. of training 2000
SSeff SSeff
Shape of MFs Bell Speed of training 0.1
No. of training 200 No. of neurons in 10
hidden layers
No. of fuzzy rules 15 No. of training 2000

The types and numbers of MFs in ANFIS including Gaussian, gen-

eralized bell-shaped, triangular and trapezoidal shaped functions,
and the parameters were tested to determine an appropriate ANFIS
model. The choosing criteria for selecting best final architecture
were based on the values of MAPE and R between the model out-
put values and observed values. After many trials in which back
propagation for the parameters were implemented, the final archi-
tectures of the ANFIS models are given in Table 3. With different
input variables, all ANFIS models had generalized bell shaped MFs
for each input variables gave the best result. The models were used
for predicting COD and SS, respectively. So it was satisfactory to
monitor the SS and COD dynamics for the paper mill wastewater
treatment process, which was optimized by trial and error during
the training process.
After determining the initial value of premise parameter and the
architecture of the predictive model, the network was trained by
the hybrid algorithm. Then the premise and consequent parameters
of the network were pruned. Membership functions of the variables
were drawn after the premise parameter was obtained. Fig. 6(a)
shows FIS Editor Screen in which the input and output variables
can be seen.
To compare with ANFIS, the appropriate ANN models were also
shown in Table 3. All ANN consisted of three independent layers:
input, hidden, and output layers. The hidden layer was comprised
of 10 operating neurons. The numbers of training were 2000.

4.4. Modeling results

Fig. 6. Relation figure of five primary GUIs and fuzzy inference systems.
After the model was trained, the inference was performed in
accordance with 16 fuzzy linguistic rules and 15 fuzzy linguistic
rules for ANFIS COD and ANFIS SS, respectively (see Fig. 6(b)). Those 4.4.1. Simulation of CODeff
rules were obtained after the network was trained. Some other All MAPE, R, MSE, and RMSE values for CODeff are also shown in
rules were also included heuristically in terms of comparing out- Table 4. When training, MAPE between the predicted and observed
put values in accordance with input values. In addition, defuzzified values of CODeff was 1.003% using ANFIS, but it was 1.7934% using
results and graphical outputs can be derived. Fig. 7(a)–(c) illustrates ANN. When validating, the MAPE was 2.815% using ANFIS, but it was
an example of Surface Viewer screen obtained from Fuzzy Logic 9.007. When predicting, the MAPE lay 3.2812% adopting ANFIS, but
Toolbox. Two- or three-dimensional graphic results of variables it was 5.7356% when using ANN. When training, R value was 0.9912
can be plotted and compared. Fig. 6(c) shows the results of applied using ANFIS, but it was 0.9758 using ANN. When validating, R value
rules and their corresponding outputs according to the mass center was 0.9459 using ANFIS, but it was 0.5371 using ANN. When pre-
of variables. Using the interface, defuzzified values for output vari- dicting, R value was 0.9093 using ANFIS, but it was 0.7719 using
ables can be derived by changing input values manually. Different ANN. MSE and RMSE values also showed that the predicting per-
output values can be obtained from the Rule Viewer according to formance of ANFIS prevailed. The MSE value of 1.2883 using ANFIS
the given input values. To get defuzzified output values for all the was lower than that of 3.5151 using ANN when model training.
real input values is not flexible using the interface. For that reason a When validating, the MSE value of 4.8321 using ANFIS was lower
program is written using Matlab codes to drive defuzzified output than that of 57.053 using ANN. When predicting, the MSE value of
results in accordance with real input values. 8.4034 using ANFIS was also lower than that of 31.639 using ANN.
3244 J. Wan et al. / Applied Soft Computing 11 (2011) 3238–3246

Table 4
Predicting performance using ANFIS and ANN.

ANFIS (PCA) ANFIS ANN (PCA) ANN

CODeff
MAPE (%) Train 1.003 0.026 1.7934 1.2016
Validate 2.8145 9.3703 8.0072 7.9522
Predict 3.2812 10.04 5.7356 11.877
R Train 0.9912 0.9999 0.9758 0.9889
Validate 0.9459 0.4517 0.5371 0.5074
Predict 0.9093 0.4487 0.7719 0.3809
MSE Train 1.2883 0.0012 3.5151 1.6196
Validate 4.8231 64.411 57.053 53.475
Predict 8.4034 71.929 31.639 112.81
RMESE Train 1.135 0.0347 1.8749 1.2726
Validate 2.1962 8.0256 7.5534 7.3127
Predict 2.8989 8.4811 5.6249 10.621
SSeff
MAPE (%) Train 0.5161 0.0135 1.2616 0.6063
Validate 1.0458 2.6906 2.2147 4.0616
Predict 1.9726 4.6121 3.6844 5.2797
R Train 0.9882 0.9999 0.9413 0.9882
Validate 0.9292 0.5010 0.6669 0.3661
Predict 0.9023 0.5951 0.7725 0.5315
MSE Train 0.0342 2.69e−5 0.1655 0.0338
Validate 0.0927 0.6221 0.3812 1.4269
Predict 0.3207 1.4369 0.9878 2.1139
RMESE Train 0.1849 0.0052 0.4608 0.1841
Validate 0.3045 0.7887 0.6174 1.1945
Predict 0.5663 1.1987 0.9939 1.4539

ANFIS, but it was 3.6844% when using ANN. When training, R value
was 0.9882 using ANFIS, but it was 0.9413 using ANN. When val-
idating, R value was 0.9292 using ANFIS, but it was 0.6669 using
ANN. When predicting, R value was 0.9023 adopting ANFIS, but it
was 0.7725 using ANN. MSE and RMSE values also showed that
the predicting performance of ANFIS prevailed. The MSE values of
0.0342 using ANFIS wad lower than that of 0.1655 using ANN when
model training. When validating, the MSE value of 0.0927 using
ANFIS was lower than that of 0.3812 using ANN. When predicting,

Fig. 7. 3D response surface graph for CODeff and SSeff.

When training and validating, the RMSE value of 1.135 and 2.1962
using ANFIS was lower than that of 1.8749 and 7.5534 using ANN.
The RMSE value of 2.8989 using ANFIS was also lower than that
of 5.6249 using ANN when predicting. Fig. 8(a) and (b) shows the
training and predicting results using ANFIS and ANN, respectively.

4.4.2. Simulation of SSeff

Fig. 9(a) and (b) depict the prediction results of SSeff using ANFIS
and ANN respectively. All MAPE, R, MSE, and RMSE values are
shown in Table 4. The 1st to 120th values were used for model train-
ing, 121st to 150th values were used to evaluate the fitness, and
the 151st to 170th were used to validating the model. As shown in
Table 3, when training, MAPEs between the predicted and observed
values of SSeff was 0.5161% using ANFIS, but it was 1.2616% using
ANN. When validating, the MAPE was 1.0458% using ANFIS, but it
was 2.2147%. When predicting, the MAPEs lay 1.9726% adopting Fig. 8. Prediction results of CODeff. (a) ANFIS and (b) ANN.
 J. Wan et al. / Applied Soft Computing 11 (2011) 3238–3246 3245

According to the results, ANFIS could predict the effluent vari-

ation. The minimum MAPEs of 1.003% and 0.5161% for CODeff and
SSeff could be achieved using ANFIS. The maximum R values for
CODeff and SSeff were 0.9912 and 0.9882, respectively. The mini-
mum MSEs of 1.2883 and 0.0342, and the minimum RMSEs of 1.135
and 0.1849 for CODeff and SSeff could also be achieved. The neural
network models provided good estimates for the BOD and SS data
sets, which cover a range of data for training and testing purposes.
The results overall indicated that the simulated effluent COD and
SS concentrations well fit measured concentrations, which was also
supported by the relatively low RMSE, MAPE and MSE and very high
R values. Such very good prediction performances of ANFIS mod-
els for both the two effluent parameters are particularly important
considering the high level of complexity in the paper mill wastew-
ater treatment process, the large quantity of variable information
spread in the dataset and the wide concentration ranges. Thus,
the ANFIS modeling approach may provide an alternative generic
framework for the modeling of other treatment processes. Further-
more, ANFIS modeling approach may have application potential
for performance prediction and control of treatment processes in
treatment plants.

Acknowledgements

This research has been supported by Guangdong Provin-

cial Department of Science (2008A080800003) and Technology
Research Project (2003A30406) and the Science and Technology
Foundation of Guangzhou city (2004Z3-D027). The authors are
Fig. 9. Prediction results of SSeff. (a) ANFIS and (b) ANN. thankful to the anonymous reviewers for their insightful comments
and suggestions.
the MSE values of 0.3207 using ANFIS was also lower than that of
0.9878 using ANN. When training and validating, the RMSE values References
of 0.1849 and 0.3045 using ANFIS was lower than that of 0.4608
using ANN. The RMSE values of 0.5663 and 0.6174 using ANFIS was [1] J.Q. Wan, M.Z. Huang, Y.W. Ma, Studies on predicting the effluent treatment
process with fuzzy neural network, Transations of China Pulp and Paper 23
also lower than that of 0.9939 using ANN when predicting.
(2008) 96–99.
After comparing with BP network, the error is less and the out- [2] M.Z. Huang, J.Q. Wan, Y.W. Ma, Y. Wang, Simulation of a paper mill wastewater
come of ANFIS model is better (see Table 4). Table 4 summarizes treatment using a fuzzy neural network, Expert Systems with Applications 36
MAPE, R, MSE and RMSE values of the training and validation data (2009) 5064–5070.
[3] N. Bhat, T.J. McAvoy, Use of neural nets for dynamic modeling and control
sets. It is very clear from Table 3 that the ANFIS has smaller RMSE of chemical process systems, Computers and Chemical Engineering 14 (1990)
(or MSE), MAPE and higher R values for both the training and testing 573–583.
datasets than the ANN model. In other words, the ANFIS achieves [4] S.H. Hong, M.W. Lee, D.S. Lee, J.M. Park, Monitoring of sequencing batch reac-
tor for nitrogen and phosphorus removal using neural networks, Biochemical
better performances than the ANN model. In this study, the min- Engineering Journal 35 (2007) 365–370.
imum MAPEs of 1.003% and 0.5161% for CODeff and SSeff could [5] I.N. da Silva, R.A. Flauzino, An approach based on neural networks for estimation
be achieved using ANFIS. The maximum R values for CODeff and and generalization of crossflow filtration processes, Applied Soft Computing 8
(2008) 590–598.
SSeff were 0.9912 and 0.9882, respectively. The minimum MSEs of [6] J.C. Chen, N.B. Chang, W.K. Shieh, Assessing wastewater reclamation potential
1.2883 and 0.0342, and the minimum RMSEs of 1.135 and 0.1849 for by neural network model, Engineering Applications of Artificial Intelligence 16
CODeff and SSeff could also be achieved. ANFIS’s architecture con- (2003) 149–157.
[7] L. Belanche, J.J. Valdes, J. Comas, I.R. Roda, Prediction of the bulking phe-
sists of both ANN and fuzzy logic including linguistic express of MFs
nomenon in wastewater treatment plants, Artificial Intelligence in Engineering
and if–then rules, so it can overcome the limitations of traditional 14 (2000) 307–317.
neural network including possibility of getting trapped in local min- [8] J.B. Serodes, M.J. Rodriguez, A. Ponton, Chlorcast: a methodology for devel-
oping decision-making tools for chlorine disinfection control, Environmental
imum and the choice of model architecture, and to increase the
Modelling and Software 16 (2001) 53–62.
predicting performance. Therefore, ANFIS is a good choice for mod- [9] P. Holubar, L. Zani, M. Hager, W. Froschl, Z. Radak, R. Braun, Advanced control-
eling the paper mill wastewater treatment performance. Moreover, ling of anaerobic digestion by means of hierarchical neural networks, Water
ANN is a black box in nature and its relationships between inputs Research 36 (2002) 2582–2588.
[10] S.R. Jang, ANFIS Adaptive-network-based fuzzy inference systems, IEEE Trans-
and outputs are not easy to be interpreted, while ANFIS is transpar- actions on Systems Man and Cybernetics 23 (1993) 665–685.
ent and its if–then rules are very easy to understand and interpret. [11] M.Z. Huang, J.Q. Wan, Y.W. Ma, Y. Wang, W.J. Li, X.F. Sun, Control rules of aera-
tion in a submerged biofilm wastewater treatment process using fuzzy neural
networks, Expert Systems with Applications 36 (2009) 10428–10437.
5. Conclusions [12] W.C. Chen, N.B. Chang, W.K. Shieh, Advanced hybrid fuzzy-neural controller for
industrial wastewater treatment, Journal of Environmental Engineering 127
(2001) 1048–1059.
In this paper, two models based on ANFIS were developed to
[13] M. Bongards, Improving the efficiency of a wastewater treatment plant by fuzzy
predict the effluent concentrations of COD and SS for a paper mill control and neural networks, Water Science and Technology 3 (2001) 189–196.
wastewater treatment process in Dongguan (China). The developed [14] J.H. Tay, X. Zhang, A fast predicting neural fuzzy model for high rate anaerobic
models were trained and tested on daily sets of COD and SS mea- wastewater treatment systems, Water Research 34 (2000) 2849–2860.
[15] J.P. Steyer, D. Rolland, J.C. Bouvier, R. Moletta, Hybrid fuzzy neural network for
surements over a period of 10 months. The ANN was also adopted diagnosis application to the anaerobic treatment of wine distillery wastewater
for comparison. in a fluidized bed reactor, Water Science and Technology 36 (1997) 209–217.
3246 J. Wan et al. / Applied Soft Computing 11 (2011) 3238–3246

[16] J.C. Chen, N.B. Chang, Mining the fuzzy control rules of aeration in a submerged [21] D.J. Choi, H. Park, A hybrid artificial neural network as a software sensor for
biofilm wastewater treatment process, Engineering Applications of Artificial optimal control of a wastewater treatment process, Water Research 35 (2001)
Intelligence 20 (2007) 959–969. 3959–3967.
[17] J. Lennox, C. Rosen, Adaptive multiscale principal components analysis for [22] C.K. Yoo, P.A. Vanrolleghem, I.B. Lee, Nonlinear modeling and adaptive moni-
online monitoring of wastewater treatment, Water Science and Technology toring with fuzzy and multivariate statistical methods in biological wastewater
45 (2002) 227–235. treatment plants, Journal of Biotechnology 105 (2003) 135–163.
[18] I.T. Jolliffe, Principal Component Analysis, 2nd ed., Springer, New York, 2002. [23] T. Takagi, M. Sugeno, Fuzzy identification of systems and its applications to
[19] G. Civelekoglu, A. Perendeci, N.O. Yigit, M. Kitis, Modeling carbon and modeling and control, IEEE Transactions on Systems Man and Cybernetics 15
nitrogen removal in an industrial wastewater treatment plant using an (1985) 116–132.
adaptive network-based fuzzy inference system, Clean 35 (2007) 617– [24] Y.M. Wang, M. Taha, Elhag, an adaptive neuro-fuzzy inference system for bridge
625. risk assessment, Expert Systems with Applications 34 (2008) 3099–3106.
[20] C. Rosen, M. Larsson, U. Jeppsson, Z. Yuan, A framework for extreme-event [25] J.H. Tay, X. Zhang, Neural fuzzy modeling of anaerobic biological wastewater
control in wastewater treatment, Water Science and Technology 45 (2002) treatment systems, Journal of Environmental Engineering – ASCE 125 (1999)
299–308. 1149–1159.