Artículo Académico / Academic Paper
Recibido: 11-01-2023, Aprobado tras revisión: 14-06-2023
Forma sugerida de citación: Izquierdo, C.; Pezantes, B.; Ayala, E. (2022). “Prediction of the Optimal Dosage of Poly Aluminum
Chloride for Coagulation in Drinking Water Treatment using Artificial Neural Networks”.
Revista Técnica “energía”. No. 20, Issue I, Pp. 93-99
ISSN On-line: 2602-8492 - ISSN Impreso: 1390-5074
Doi: https://doi.org/10.37116/revistaenergia.v20.n1.2023.562
© 2023 Operador Nacional de Electricidad, CENACE
Prediction of the Optimal Dosage of Poly Aluminum Chloride for Coagulation
in Drinking Water Treatment using Artificial Neural Networks
Predicción de la Dosis Óptima de Policloruro de Aluminio para el Proceso de
Coagulación en el Tratamiento de Agua Potable mediante Redes Neuronales
Artificiales
C. Izquierdo1 B. Pezantes1 E. Ayala1
1Grupo de investigación de inteligencia artificial y tecnologías de asistencia – GI IATA, Universidad Politécnica
Salesiana, Cuenca, Ecuador
E-mail: cizquierdov@est.ups.edu.ec; bpezantes@est.ups.edu.ec; eayala@ups.edu.ec
Abstract
The addition of chemicals in drinking water
treatment is usually a manual procedure
performed by highly trained and experienced
persons. To solve this problem, this study is based
on the analysis of data collected from a raw water
source located in Ecuador. Then, using the
information on the physical-chemical parameters
of the raw water such as pH, turbidity and color,
the definition of the doses of Poly Aluminum
Chloride (PAC), and the input and output
variables of the dosage process are identified.
Consequently, the implementation of an
intelligent control system based on Artificial
Neural Networks (ANN) is proposed in order to
reduce the dependence on experienced people.
These experiments start with data collection and
analysis in order to establish the variables
involved in the process. The proposed neural
model has three hidden layers, and it uses adaptive
gradient algorithms. An analysis of the results was
performed using Mean Absolute Percentage Error
(MAPE) and Root Mean Square Error (RMSE).
The PAC predictive model in the training phase
gives a MAPE value of 0.0425 for the not adjusted
values and 0.0262 for the adjusted numerical
values. However, in the test phase the neural
model achieves a MAPE of 0.057 for the not
adjusted PAC values and 0.041 for the adjusted
values. This alternative provides an efficient
solution to solve dosing problems in drinking
water treatment plants (DWTP), with reliable
results according to RMSE and MAPE metrics.
Resumen
La adición de las sustancias químicas en el tratamiento
del agua potable comúnmente es un procedimiento
manual realizado por personal altamente capacitado y
experimentado. Esta resulta una tarea crítica debido a
que requiere cierto nivel de experiencia para una
correcta dosificación. Como posible solución, este
estudio se basa en el análisis de datos recolectados de
una fuente de agua cruda ubicada en Ecuador.
Utilizando la información de los parámetros
fisicoquímicos del agua cruda, como el pH, turbidez y
color, se identifican las dosis de Policloruro de
Aluminio (PAC), y las variables de entrada y salida del
proceso. En consecuencia, se propone la
implementación de un sistema de control inteligente
basado en Redes Neuronales Artificiales (RNA) con la
finalidad de reducir la dependencia del personal
experimentado. Para ello, se parte con la recolección y
análisis de datos y así establecer las variables
involucradas en el proceso. El modelo neuronal
propuesto dispone de tres capas ocultas y utiliza
algoritmos de gradiente adaptativo. El análisis de los
resultados se realizó mediante el error porcentual
absoluto medio (MAPE) y el error cuadrático medio
(RMSE). El modelo predictivo de PAC en etapa de
entrenamiento indica un valor MAPE de 0,0425 para
los valores no ajustados y de 0,0262 para los valores
numéricos ajustados. Sin embargo, en la etapa de
prueba el modelo neuronal alcanza un MAPE de 0,057
para los valores de PAC no ajustados y de 0,041 para
los ajustados. Esta alternativa brinda una solución
eficiente a la hora de resolver problemas de
dosificación en las plantas de tratamiento de agua
potable (PTAP), teniendo resultados confiables según
las métricas RMSE y MAPE.
Index terms−− Drinking water, Dosing, DWTP,
Coagulant chemicals, Artificial neural networks,
Control system.
Palabras clave−− Agua potable, Dosificación, PTAP,
Químicos coagulantes, Redes neuronales artificiales,
Sistema de control.
93
Edición No. 20, Issue I, Julio 2023
1. INTRODUCTION
Currently there are a lot of industrial processes that
demand a specific control method for its proper
operation. Within this group of procedures requiring
modernization and implementation of control systems,
are processes related to water treatment, specifically
drinking water treatment, which involves transforming
raw water from natural sources into drinking water within
the parameters established under specific standards for
human consumption. A conventional drinking water
treatment process consists of sequential steps and the
most important one, is coagulation since it ensures that
the dosed quantities of coagulating chemicals are in
accordance with the properties of the raw water such as
color, turbidity, pH and alkalinity [1]. Obtaining the
doses of these chemicals reacts to a non-linear response
done by experts. It is performed by jar testing, which is
not adaptive to changes in real time and needs a
considerable amount of time for its execution [1]. This is
an issue since there is an immediate and constant
dependence on qualified and experienced operators.
To ensure good quality of treated water, operators
must adjust the amounts of coagulant chemicals at certain
time intervals or climatic conditions where the water has
parameters outside the usual range. Excessive amounts of
coagulant chemicals correspond to increased treatment
costs and public health problems. An under dosage
corresponds to a failure in the flocculation of the water
and increases the frequency of maintenance of DWTP
increasing the cost of production. Moreover, an
implementation of algorithms based on artificial
intelligence and machine learning, which have been
investigated and implemented in treatment plants around
the world, is proposed.
For instance, [2] uses the potential provided by
artificial neural networks, supporting vector machines,
and gene expression programming to approximate the
model of trihalomethane formation generated by chlorine
water disinfection processes. They obtained as a result
three models that capture the complex nonlinear behavior
of the collected data. They also indicated excellent
predictive and generalization capability. Furthermore, it
demonstrated that these types of models, which
commonly need a large amount of data, apply to a smaller
amount of data. In another research, [3] artificial neural
networks (ANN) are utilized to model the PAC dose.
This method responds well when obtaining the
appropriate dose in real-time when a storm brings high
turbidity in raw water. In fact, they defined the input
variables using Pearson’s correlation and validated their
model using the mean square error obtained.
In addition, [4] developed a model where the type of
coagulant to be used is set by decision trees and the dose
was estimated by ANN, allowing to calculate from the
raw water parameters (pH, turbidity, and temperature),
the amount and type of coagulant to be used (PAC, PASS
and PSO-M). Moreover, [5] developed a model to predict
turbidity and color of treated water at the outlet of the
Rossdale WTP located in Edmonton, Alberta in Canada.
In 2009, [6] determined that the coagulant dose cannot be
settled under traditional mathematical models, because it
depends on several factors. Stating that the prediction of
coagulant by neural network provides high accuracy and
faster convergence speed and can be used to predict in
real time online.
These types of neural models are seen as standard
estimators of nonlinear relationships and their predictive
and generalization capabilities let them have successful
applications in different fields of knowledge [7].
On account of the above-mentioned research, the
objective of this study is to build a neural model that
adequately adapts to the relationship between raw water
quality and the doses of chemicals needed for treatment.
Initially, the data obtained involves a dosing history over
a period of 14 months. The correlation between raw water
quality and coagulant dosage was found. We will have to
find a middle ground in the learning of our model in
which we are not underfitting and not overfitting. For
this problem, the input data set for training should be
subdivided into two: one for training and one for the test
that the model will not know beforehand. This division is
usually made of 80% for training and 20%. The Test set
should have diverse samples and enough samples to be
able to check the results once the model has been trained.
We proceeded with the training and validation process of
the neural model by using the adaptive gradient algorithm
and the analysis of the results using MAPE and RMSE.
The results for the training set were an RMSE value
below 2.82 and for the MAPE, a value of less than 0.045.
On the other hand, for the test, set a RMSE value below
3.3 and 0.06 for the MAPE. It has been observed that the
RMSE metric does not predict whether the estimation
model is ideal or not. On the contrary, MAPE offers a
better way to determine the accuracy of the model. The
higher model accuracy is achieved when the value of
MAPE is lower. Proving that the system had the ability
to get information from dosing background and be able
to estimate PAC doses for different raw water qualities.
The following paper is organized as follows: an
overview of the water treatment process, determination
of the variables, data analysis, construction and training
of the neural model, analysis of results and conclusions.
2. METHODOLOGY
2.1. Determination of Process Variables
The conventional drinking water treatment process
consists of 6 stages where the predominant process is
dosing. This defines the success of the following stages.
The process diagram of the drinking water treatment
process is shown in Fig. 1.
94
Izquierdo et al. / Predicción de Dosis Óptima de Policloruro de Aluminio para el Tratamiento de Agua Potable.
Figure 1: Conventional drinking water treatment process
The measurement of organoleptic and physical-
chemical parameters is performed on the raw water.
These measured values help in acquiring the amount of
chemicals to be dosed for water coagulation. The
amounts of chemicals can be tested by a jar test in the
DWTP laboratory before dosing. Based upon the dosing
process carried out at the DWTP, from the studies
previously mentioned and the experience, the minimum
and necessary amount of input variables involved in the
process was defined. The input variables of the control
system are pH, turbidity, and color of the raw water.
These were selected through expert knowledge and were
determined to be the minimum necessary to estimate the
doses of chemicals considering the limited
instrumentation existing in the DWTP. These variables
can be quantified by means of sensors located at the inlet
of the DWTP. The output variables are directly related to
the amount of chemicals to be dosed in parts per million
(ppm). Therefore, the control system in a generic way
was shaped by the mentioned process variables and is
structured in Fig. 2.
Figure 2: Control diagram for dosage
2.2. Data Collection and Analysis
The data obtained correspond to a water source that
supplies one of the DWTP in Ecuador. This database
contains the input parameters of the control system,
which are the quantifiable characteristics of the water that
took one year and two months of data collection. It started
on October 1, 2017 and ended on December 31, 2018
with a total of 438 data points available. The raw water
parameters of the raw water source are shown in Fig. 3 -
Fig. 5.
Figure 3: Raw water pH
Figure 4: Raw water turbidity.
Figure 5: Raw water color
Fig. 6 shows the PAC dosages. The graph shows the
dosage in normalized values. A linear scale between 0
and 1 was used to normalize the data using the following
equation.  
 (1)
Where D is the PAC dose, Max is the maximum PAC
dose value, and Min is the minimum PAC dose value.
95
Edición No. 20, Issue I, Julio 2023
Figure 6: PAC Dosage
Pearson's correlation was used to determine the
statistical relationship between the process variables.
This is shown in Table 1.
Table 1: Pearson's Correlation Between each Input and Output
Pearson's correlation coefficient (r)
PAC
Color
Turbidity
PAC
1
-0.63012
-0.70006
Color
1
0.89924
Turbidit
y
1
pH
The relationship between the PAC output and the
Color, Turbidity and pH inputs is a negative association,
i.e. as the raw water parameters increase their value for
the amount of PAC doses decreases. It can also be said
that this correlation between PAC doses and the
parameters: Color and Turbidity is higher compared to
the association between PAC and water pH. The color
and turbidity of the water have a relatively high positive
correlation, that means that as the color of the raw water
increases its value so does the turbidity. Additionally,
color and turbidity are positively related to pH but their
relationship value is low. Finally, the coefficient of
determination between the variables was obtained, which
is illustrated in the following table:
Table 2: Determination Between each Input and Output
Pearson's correlation coefficient (r)
PAC
Color
Turbidity
PAC
1
0.39706
0.49009
Color
1
0.80864
Turbidity
1
pH
It is proved that the indicator of determination
between the PAC, the variable to be predicted, and the
Color, Turbidity and pH inputs is found to be less than
50% in each case. As a result, the model is a poor fit to
its data. Consequently, the model belongs to a non-linear
system and demands an intelligent control system that is
suitable for the data.
2.3. Structure of the Neural Model
An Artificial Neural Network (ANN) represents a
computer model based on the application of theoretical
neurophysiology that replicates the way in which the
human nervous system communicates and propagates. In
[8] McCulloch & Pitts developed the first computer
model that captures this work.
A Multilayer Perceptron (MLP) consists of a network
architecture composed of one or several hidden layers, an
interconnected system that examines information in a
parallel but non-linear way, giving the ability to solve
non-linear problems. The hidden layer makes a
connection between each input and each output of the
neural network, forming a fully connected or "dense"
model. The information that reaches the input layer
generates an activation pattern that in turn is an input
signal applied to the neurons of the hidden layer [9]. If
there is more than one hidden layer, the output signal of
the first hidden layer is the input of the next hidden layer
and so on until the output layer is accomplished. The
matrix-expressed implementation of this algorithm called
forward propagation is shown as follows:
󰇟󰇠󰇟󰇠󰇟󰇠
 (2)
󰇟󰇠󰇛󰇜
Being the number of layers of the MLP, is the
transfer activation function, 󰇟󰇠 is the current synaptic
weights, 󰇟󰇠 represents the output of the previous
layer,
is the bias vector and 󰇟󰇠 represents the output
of the current layer.
An MLP is regularly trained with stochastic gradient
descent methods [10], a technique in which the ANN
parameters are updated at each iteration and the error is
propagated backwards, updating the synaptic weights
and decreasing the error in prediction.
In the present case study, the inputs of the ANN are
the physical-chemical characteristics of the raw water:
Turbidity, Color and pH. The output is the amount in
parts per million of PAC.
Recordings of data corresponding to 438 dosages of
the chemical agent PAC, with their respective input
parameters, were used. The outcomes suggest that 80%
of the data should be used for the training set, and the
remaining 20% for the test set. Through the training
process of the neural network, the best hyperparameters
suit the model and allow the best accuracy in the
prediction of the output were determined. Which are
shown in the following table:
96
Izquierdo et al. / Predicción de Dosis Óptima de Policloruro de Aluminio para el Tratamiento de Agua Potable.
Table 3: Hyperparameters of the neural model
Dense Sequential Deep ANN
Layer
Number of
neurons
Number of
parameters
Activation
function
Input
3
-
-
Hidden 1
50
20
200
Sigmoid
Hidden 2
200
10200
Sigmoid
Hidden 3
50
10050
Sigmoid
Output
1
51
Linear
Total parameters: 20501
Trainable parameters: 20501
Untrainable parameters: 0
Learning rate: 0.1
Number of training epochs: 3000
Optimizer: Adaptative Gradient Algorithm (Adagrad)
Root Mean Square Error (RMSE): 2.54
The design of the deep ANN for PAC dosing is
illustrated in the Fig. 7.
Figure 7: ANN Architecture
3. RESULTS AND DISCUSSIONS
Once the artificial neural model has been trained, a
validation was performed using real vs. predicted data.
For this, the data from the training set was initially
employed, later the data for the test set was used, which
represented new data for the neural model. The trained
model has the ability to predict the PAC dose with the
test data and with new data in real time.
It should be noted that the neural model predicts
decimal values due to the activation functions used in its
neurons, which may be found in the database. The PAC
dosage values are integer values (ppm). Hence, the next
step has been used to approximate each datum to its
immediate superior in order to get the predicted doses
correctly. This does not affect the output results because
the dosages are made through doses with specific steps
(50, 60, 70). Therefore this approach helps us to stay in
the practical range of dosages. However, for their
representation, both results were considered as an
illustration. The performance and precision of the neural
model with respect to the real data of the dosages were
through the Mean Absolute Percentage Error (MAPE)
and the Root Mean Square Error (RMSE), metrics
expressed by the following equations:
󰇛󰇜


Where is the measured value, is the predicted
value, and
represents the number of samples. The
results obtained in the training and testing phase are
shown in the Table 4.
Table 4: Accuracy of the neuronal model
RMSE
Training set
(Not adjusted values)
2.549
Training set
(Adjusted values)
2.816
Test set
(Not adjusted values)
3.141
Test set
(Adjusted values)
3.254
Fig. 8 and 9 show the performance of the Deep ANN
through the comparative graphs of the real data vs. the
data predicted by the neural network in the training set.
In the same way, the performance of the Deep ANN with
the test set is shown through the Fig. 10 and 11.
Figure 8: Training set (Not adjusted values)
97
Edición No. 20, Issue I, Julio 2023
Figure 9: Training set (Adjusted values)
Figure 10: Test set (Not adjusted values)
Figure 11: Test set (Adjusted values)
The main problem with the RMSE metric is that it
does not predict whether the estimation model is good or
not. MAPE offers a better way to determine model
accuracy, while the value of MAPE is lower, then the
model is more accurate.
As can be seen in Table 4, the PAC predictive model
in the training phase achieves a MAPE value of 0.0425
for PAC values and 0.0262 for adjusted values. This
demonstrates that the predictions delivered by the neural
model are quite accurate. However, in the testing phase,
the neural model reaches a MAPE of 0.057 for PAC
values and 0.041 for adjusted, which still displays the
reliability and accuracy of the system.
4. CONCLUSIONS
In conclusion, this deep ANN gives an efficient
solution to solve dosing problems that may occur in any
DWTP. Within the DWTP, pH Regulator, PAC and
flocculant are dosed, due to the focus of this project, a
model with 3 inputs (pH, Turbidity, Color) - 1 output
(PAC) was studied. Only the PAC will be used as output
because it represents the minimum model that can be
developed. To verify that the dosage is correct, pH,
turbidity and color are measured at the outlet of the WTP
(treated water) and check that the parameters are within
the norm. The system automatically defines the
appropriate dosage regardless of the parameters that are
presented without the need to call or have the immediate
help of a specialist. The various implementations of these
systems have been studied within the state of the art in
DWTPs in different parts of the world, and it has been
possible to verify the success of this alternative in
comparison to methods that do not use artificial
intelligence.
This system can have a good reception in companies
dedicated to the treatment of drinking water leading to
answer the need of a possible booming market, which
coincides with the implementation of automatic and
intelligent systems in our country.
According to the metrics used, in the evaluation of the
unadjusted training set, an RMSE of 2.549 and MAPE of
0.0425 were obtained. For the adjusted training data, an
RMSE of 2.816 and a MAPE of 0.0262 were also
acquired. As demonstrated in both cases of the training
test, the MAPE is below 0.045, which reveals that the
model is quite accurate.
Furthermore, for the unadjusted test set an RMSE
value of 3.141 and a MAPE of 0.057 was obtained. For
the adjusted data set an RMSE of 3.254 and MAPE of
0.0401 were obtained. As can be seen in the testing phase,
for both cases, MAPE of less than 0.06 is examined,
which still shows the efficiency of the ANN.
According to the aforementioned values it is
concluded that the Deep ANN model is correctly adapted
to the nonlinear behavior describing the chemical dosing
processes from the parameters of the raw water discharge
and has the ability to gain knowledge from a dosing
history. It is mentioned that in the proposed treatment
process, only the PAC doses are predicted, the doses of
pH regulator and flocculant are calculated in the field
with the knowledge of the expert without affecting the
operation of the DWTP.
A comparison of the different neural models for the
prediction of coagulant chemicals can be proposed as a
98
Izquierdo et al. / Predicción de Dosis Óptima de Policloruro de Aluminio para el Tratamiento de Agua Potable.
future study. In addition, another system can be
implemented where the outputs are all the coagulating
chemicals. In this case only one output (PAC) is used as
a starting point for future projects.
REFERENCES
[1] J. A. Rojas, Potabilización del agua, Mexico, DF:
ALFAOMEGA GRUPO EDITOR, 1999.
[2] K. P. Singh and S. Gupta, Artificial intelligence based
modeling for predicting the disinfection by-products
in water, Chemometrics and Intelligent Laboratory
Systems, vol. 114, pp. 122-131, 2012.
[3] D. W. Guan and L. L. Shang, Effects of data
normalization and inherent-factor on decision of
optimal coagulant dosage in water treatment by
artificial neural network, Expert Systems with
Applications, vol. 37, nº7, pp. 4974-4983, 2010.
[4] H. Bae, S. Kim and Y. Kim, Decision algorithm based
on data mining for coagulant type and dosage in
water treatment systems, Water Science and
Technology, vol. 53, pp. 321-329, 2006.
[5] Q. J. Zhang and S. J. Stanley, Real-time water
treatment process control with artificial neural
networks., Journal of Environmental Engineering,
vol. 125, pp. 153-160, 1999.
[6] Z. Song, Y. Zhao, X. Song and C. Liu, Research on
prediction model of optimal coagulant dosage in
water purifying plant based on neural network, 2009
ISECS International Colloquium on Computing,
Communication, Control, and Management, vol. 4,
pp. 258-261, 2009.
[7] H. Yoon, S.-C. Jun, Y. Hyun, G.-O. Bae and Kang-
Kun Lee, A comparative study of artificial neural
networks and support vector machines for predicting
groundwater levels in a coastal aquifer, Journal of
Hydrology, vol. 396, pp. 128-138, 2011.
[8] W. McCulloch S y W. Pitts, “A logical calculus of the
ideas immanent in nervous activity”, The bulletin of
mathematical biophysics, vol. 5, n.º4, pp. 115–133,
1943.
[9] A. Romero Herrera, “Aproximacion de funciones con
redes neuronales y algoritmos evolutivos”, Master’s
thesis, Universidad Autonoma del Estado de
Mexico.
[10] D. E. Rumelhart, G. E. Hinton y J. L. McClelland,
”A general framework for parallel distributed
processing”, Parallel distributed processing:
Explorations in the microstructure of cognition, vol.
1, n.º 45-76, p. 26, 1986.
Cristopher Izquierdo Verdugo.-
He was born in Cuenca, Ecuador in
1998. He received his Mechatronic
Engineer degree in Automation of
Industrial Processes from the
Salesian Polytechnic University in
2022; He did a student exchange
through the ERASMUS+ program
within the field of Artificial Intelligence at the University
of Pisa, Italy in 2020; and his field of research is related
to artificial intelligence and assistive technologies.
Braulio Pezantes Domínguez.-
He was born in Cuenca, Ecuador in
1998. He received his degree in
Mechatronic Engineer in 2022 and
currently, he is pursuing his Master
studies in Electronics and
Automation, both at the the
Salesian Polytechnic University of
Ecuador. His field of research is related to artificial
intelligence, industrial automation and assistive
technologies.
Edy Ayala Cruz.- He was born in
Cuenca in 1987. He received his
degree in Electronic Technologist
in 2009 and Electronic Engineer in
2011, both from the Salesian
Polytechnic University of
Ecuador;
and his Master of Engineering
Science (Electrical and Electronic)
degree from Swinburne University of Technology in
Australia, 2015. Currently, he is pursuing his Doctorate
studies in Engineering at the University of Ferrara. His
field of research is related to control systems in
engineering and renewable energies.
99