ARTIFICIAL NEURAL NETWORKS

MODELLING FOR AL-RUSTUMIYA

WASTWATER TREATMENT PLANT IN

BAGHDAD

Dalia H. aldahy*

Civil Engineering Department, College of Engineering, Al-Nahrain University,

Baghdad, Iraq

st.dalia.hussein@ced.nahrainuniv.edu.iq

Mohammed A. Ibrahim

Civil Engineering Department, College of Engineering, Al-Nahrain University,

Baghdad, Iraq

Reception: 03/12/2022 Acceptance: 27/01/2023 Publication: 13/02/2023

Suggested citation:

H. A., Dalia and A. I., Mohammed (2023). Artiﬁcial Neural Networks

Modelling For AL-Rustumiya Wastwater Treatment Plant in Baghdad . 3C

Empresa. Investigación y pensamiento crítico, 12(1), 257-271. https://doi.org/

10.17993/3cemp.2023.120151.257-271

https://doi.org/10.17993/3cemp.2023.120151.257-271

257

3C Empresa. Investigación y pensamiento crítico. ISSN: 2254-3376

Ed. 51 Iss.12 N.1 January - March, 2023

ABSTRACT

In the present research, Artificial Neural Networks (ANNs) were developed for

modelling the performance of Al-Rustamiya wastewater treatment plant, Baghdad,

Iraq. There were created two models and the outputs were the removal efficiency of

BOD and COD parameters. Four main input parameters were selected for modelling,

namely Total suspended solids (TSS), Total dissolved solids (TDS), chloride ion (Cl-),

and pH. Influent and effluent concentrations of the parameters were collected from

Mayoralty of Baghdad for the period from 2011 to 2021. The results of the modelling

were in terms of mean square error (MSE) and correlation coefficient (R). The results

indicated that the ANNs models were accurately able to predict the removal of the

BOD, and COD, and the optimum topology of the ANNs is obtained at 13 neurons in

the hidden layer for both with 3.09 MSE, 0.96 and 4.28 MSE, 0.96 R for BOD and

COD respectively.

KEYWORDS

ANNs, BOD, COD, modelling, wastewater

PAPER INDEX

ABSTRACT

KEYWORDS

INTRODUCTION

WORK METHOD

1. Description of the Study Area

2. Data Collection

3. ANNs modelling in MATLAB

RESULTS AND DISCUSSION

1. Removal Efficiency of BOD and COD

2. ANNs

2.1. ANNs for BOD modelling

2.2. ANNs for COD modelling

CONCLUSION

ACKNOWLEDGEMENTS

CONFLICT-OF-INTEREST STATEMENT

REFERENCES

https://doi.org/10.17993/3cemp.2023.120151.257-271

3C Empresa. Investigación y pensamiento crítico. ISSN: 2254-3376

Ed. 51 Iss.12 N.1 January - March, 2023

258

INTRODUCTION

The change in the life styles due to urbanization and industrial practices has

resulted in increasing production of wastewater that adversely impacts on human lives

[1]. Water can be contaminated by human utilization in any mixture of mechanical,

residential, businesses, storm water, surface overflow, and inflow of sewers [1, 2].

Wastewater treatment plants (WWTPs) are considered as significant infrastructure for

a society. With the realization of the importance of such plants, the achievement of

integrated and efficient units demands adequate operational and maintenance

practices. Wastewater treatment plants mainly consist of three process, primary,

secondary, and tertiary treatments. The combinations of these processes depends on

the characteristics requirements of effluent [3]. Typically, the reduction degree in the

biological oxygen demand (BOD) as well as chemical oxygen demand (COD)

represent the basic indicator of the effectiveness of the plant [4, 5].

Moreover, modelling of WWTPs represents a difficult task as the treatment involves

complex processes. The physical, biological, and chemical stages of the treatment

plants provide non-linear performance which is complicated to presented in linear

models. Thus, providing an efficient monitoring technique can be accomplished by the

development of non-linear model to predict the performance of the treatment plant

under previous observed water characteristics. Artificial neural networks (ANNs)

represent computerized non-linear models for simulating the decision-making and

functions of the brain of humans. It is being used for many wastewater quality issues.

It has also been properly used in the modelling of the WWTPs for predicting

wastewater characteristics, controlling stages of treatments, and providing estimation

of effluent characteristics [6-9].

ANNs are usually used for predicting the parameters of water quality. It solves an

issue through the development of a memory with the ability to relate large input data

with a set of outputs [10]. A significant feature of the ANNs is its ability to handle

considerable and complicated systems with various related parameters [11]. ANNs

forms the basis of deep learning where the modelling algorithms are inspired by the

brain structure of humans. After taking in the data, the ANNs train themselves for the

recognition of data patterns and providing outputs. The model consists of grouped

artificial neurons representing the core units of the modelling process [12]. The model

represents an alternative method to conventional water quality models through the

provision of advanced predictions and forecasts [13, 22].

In Baghdad, the capital city of Iraq, Al-Rustumiya WWTP is one of the main sewage

water treatment facilities in the country. The plant recently has been expanded (3rd

expansion) with the construction of nearby new plant for increasing capacity

purposes. It releases the treated waters in Diyala River and then into the Tigris River

[14, 15]. This study aims to at develop an ANNs model in MATLAB for investigating

correlation between pairs of parameters to predict the performance of the 3rd

expansion plant in terms of the removal efficiency of BOD, and COD. This work can

assist in facilitating assessment or process control of effluent quality.

https://doi.org/10.17993/3cemp.2023.120151.257-271

259

3C Empresa. Investigación y pensamiento crítico. ISSN: 2254-3376

Ed. 51 Iss.12 N.1 January - March, 2023

WORK METHOD

1. DESCRIPTION OF THE STUDY AREA

Baghdad is the capital city of Iraq with an area approximately equals 800 square

miles and population of 7.145 million according to the water and sewage sector report

[16]. Three significant central WWTPs were built in Baghdad., namely Al-Karkh

wastewater treatment plant, Southern Al-Rustumiya WWTP numbered 0, 1, and 2,

and Northern Al-Rustumiya WWTP (3rd expansion). The effluent from the old plant

being discharged in Tigris River while for the 3rd expansion, it is being discharged in

Diyala River (also known as Sirwan River). The 3rd expansion of the plant began in

1984. Approximately third of the population in the city depends on the Al-Rustumiya

WWTP [17]. The plant lies on the south-east part of Baghdad, Iraq, on Diyala River

with longitudinal coordinate of 44˚32΄05˝E and latitudinal coordinate of 33˚17΄15˝N.

The plant was mainly designed for the treatment of domestic wastewater which serves

a population of 1500000 [17]. The plant consists of conventional activated sludge for

biologically treating carbon compounds with average wastewater influent capacity

equals 300MLD [17]. Figure 1 showed the 3rd expansion of the plant.

Figure 1. Al-Rustumiya WWTP3rd expansion in Baghdad, Iraq [14].

https://doi.org/10.17993/3cemp.2023.120151.257-271

3C Empresa. Investigación y pensamiento crítico. ISSN: 2254-3376

Ed. 51 Iss.12 N.1 January - March, 2023

260

2. DATA COLLECTION

The required data were obtained from Al-Rustamiya WWTP administration Office

Mayoralty of Baghdad, over the period of ten years from January 2011 and December

2021 on a monthly basis. Two locations were selected: at entrance of the plant, and

after secondary treatment. The selected physiochemical parameters included the

biochemical oxygen demand (BOD5), chemical oxygen demand (COD), chloride (Cl-)

and Total dissolved solids(TDS), and total suspended solids (TSS). The Removal

efficiency of BOD, and COD were determined using Equation (1):

Removal efficiency (%) = (1)

3. ANNS MODELLING IN MATLAB

The ANNs were created in MATLAB. This software allows creation, usage, export,

and input of neural networks. Two models were created, with 4 inputs and 1 output.

The modelling procedure and equations were based on Tümer and Edebali [18],

Ammari [19] and Alsulaili and Refaie [20].

The ANNs architecture was defined by its number of layers and their neurons.

Feedforward multi-Layered perception ANNs consist of various artificial neurons

known as nodes, or processing elements (PEs). These are normally arranged in three

layers, input, hidden or intermediate, and output layers. As was indicated in Figure 2

that for each processing element, the input from a layer (xi) was multiplied by an

adjustable connection weight (wij). The summation was performed for the weighted

inputs with the addition of a threshold value (θj). The resulted combined input was

then transferred to activation function (f(ij)) for generating the output (yj). This output

was then used as input for the next layer. The activation function represents the

nonlinear mapping tool in the network before delivering the output to the next layer.

Figure 2. Typical structure and operation of ANNs [21].

The summary of the process was indicated in Equations (1 & 2) as follows:

Summation f(IJ) = (2)

Transfer (yj)= f(Ij) (3)

inputvalue −outputvalue

inputvalue

x

100

∑WijXi+θj

https://doi.org/10.17993/3cemp.2023.120151.257-271

261

3C Empresa. Investigación y pensamiento crítico. ISSN: 2254-3376

Ed. 51 Iss.12 N.1 January - March, 2023

Where,

f(IJ) is the level of activation at j node,

Wij is the weight of the connection between the nodes i and j,

xi is the i node input and equals 0, 1,2, …,n,

θj is the threshold for j node,

j is the j node output, and

f(ij) is the activation function.

Backpropagation algorithm is normally used for excellent results and easy

application. The ANNs data propagation began at the input layer at which the

information was provided. In order for the weights to be assigned, the network utilized

the presented information and learning rule for the production of input and output

maps with negligible errors. This is called training or learning which was divided into

supervised and unsupervised learnings. Feedforward ANNs usually works with

supervised learning. In this, model inputs and desired outputs were provided to the

network. Errors were determined through the network by comparing the desired and

actual ANNs produced outputs. These errors were then utilized for adjusting the

weights given to the connections between the input and the outputs for reducing the

errors between the actual output and the desired ones. Thus, the network learns for

the presented data for adjusting the weights and capturing the relationship between

the input and the output without the need of any previous knowledge about such

relationship. Thus it regulates the bias and weights of the Multi-Layered perception

ANNs. In order to evaluate the efficiency of the treatment plant, the backpropagation

algorithm was improved by the incorporation of Levenberg-Marquadrt algorithm. This

algorithm is a local optimization algorithm with a gradient basis. The advantage of its

utilization over normal backpropagation is the better stability, advanced performance,

and faster and advanced training and convergence properties.

Several forms are available for the activation function. The most commonly known

were used in this research which was hyperbolic tangent transfer and logistic sigmoid

functions. The functions are presented in Equations (4 & 5):

Hyperbolic tangent transfer function (3)

logistic sigmoid function (4)

)()(

)( jj

jee

If −

−

)(

−

https://doi.org/10.17993/3cemp.2023.120151.257-271

3C Empresa. Investigación y pensamiento crítico. ISSN: 2254-3376

Ed. 51 Iss.12 N.1 January - March, 2023

262

RESULTS AND DISCUSSION

1. REMOVAL EFFICIENCY OF BOD AND COD

Figures 3 presented the average yearly removal efficiency of BOD and COD from

Al-Rustumiya WWTP. The figure indicated that the removal efficiencies were variable

for both water parameters, mostly more than 80% and having approximately the same

trend. The greatest removal efficiency occurred in 2020 with about 93% for BOD and

95% for COD. Generally, the removal efficiency of the plant increased from 2015

onwards. The lowest BOD and COD removals occurred in 2012. This could be

attributed to the improper aeration in the aeration basin, or the measurement of high

concentration of settling microbial mass in the secondary clarifier.

Figure 3. Removal efficiency of BOD and COD from Al-Rustumiya WWTP.

2. ANNS

The results of the ANNs modelling were evaluated in terms of Mean Square Error

(MSE) and correlation coefficient (R) The functions are presented in Equations (6&7):

Correlation coefficient (R) = (6)

Mean Square Error (MSE) = (7)

where X= observed yt , = mean of X, Y= predicted yt, = mean of Y, and n=

number of observations.

Removal Efficiency

Removal Efficiency, %

100

Year

2011

2012

2013

2014

2015

2016

2017

2018

2019

2020

2021

BOD COD

∑n

i=1

(X−¯

X)(Y−¯

∑n

i=1 (

X−¯

)

(

Y−¯

)



∑n

i=1

(X−Y)

https://doi.org/10.17993/3cemp.2023.120151.257-271

263

3C Empresa. Investigación y pensamiento crítico. ISSN: 2254-3376

Ed. 51 Iss.12 N.1 January - March, 2023

These are the most common criteria for the evaluation of a model’s performance.

The MSE provides the difference between the outputs and the targets. The R

coefficient reflects how well the created model fits with used outputs. An R equaling 1

means a very close relationship while an R equaling 0 means a random relationship.

Models were carried out for the removal efficiency of both BOD and COD. They

represent the main parameters for evaluating organic pollution. They provide

measurements of organic matter and oxygen demands.

2.1. ANNS FOR BOD MODELLING

The modelling results in terms of the MSE with the number of hidden layers are

shown in Figure 4. and Table 1. The results indicated that the MSE was significantly

decreased with the increase in the number of hidden neurons from 2 to 9 and the

minimum MSE result is obtained at 13 hidden neurons. Afterward, training was

stopped when reached 90 epochs for the Levenberg–Marquardt algorithm as shown

in Figure 5. This is because of the difference between the training and validation

errors increases. A plot of Levenberg–Marquardt algorithm regression for training,

validation, and testing with R is shown in Figure 6. This revealed that the R values

were 0.98, 0.94, 0.88, and 0.96 for training, validation, testing, and for all data

respectively. Thus, the optimum topology of the ANNs is obtained at 13 neurons in the

hidden layer with 3.09 MSE and 0.96 R. The architectural model of the optimum

topology is shown in Figure 7. This is 4:13:1, indicating the input layer with the used

four parameters, thirteen neurons at the hidden layers, and the output layer in terms

of the removal efficiency of the BOD.

Figure 4. Mean Square Error (MSE) with different numbers of neurons at the hidden layers

for ANNs modelling of BOD output.

https://doi.org/10.17993/3cemp.2023.120151.257-271

3C Empresa. Investigación y pensamiento crítico. ISSN: 2254-3376

Ed. 51 Iss.12 N.1 January - March, 2023

264

Table 1. ANNs details for BOD modelling.

Figure 5. Training, validation, and test mean square errors for the Levenberg–Marquardt

algorithm for BOD removal.

BOD

Model No. No. of neurons at hidden

layers

MSE Correlation Coefficient (R)

1 2 15.72 0.80

2 3 14.13 0.83

3 4 13.59 0.83

4 5 12.83 0.84

5 6 9.75 0.88

6 7 6.00 0.93

7 8 4.87 9.43

8 9 3.42 0.96

9 10 3.96 0.95

10 11 3.41 0.96

11 12 4.00 0.95

12 13 3.09 0.96

13 14 4.43 0.95

14 15 3.21 0.96

https://doi.org/10.17993/3cemp.2023.120151.257-271

265

3C Empresa. Investigación y pensamiento crítico. ISSN: 2254-3376

Ed. 51 Iss.12 N.1 January - March, 2023

Figure 6. Training, validation and testing regression for the Levenberg–Marquardt

algorithm for BOD removal.

Figure 7. The architecture of the ANN model for the prediction of BOD removal.

2.2. ANNS FOR COD MODELLING

The modelling results in terms of the MSE with the number of hidden layers are

shown in Figure 8. and Table 2. The results indicated that the MSE was significantly

decreased with the increase in the number of hidden neurons from 2 to 11 and the

minimum MSE result is obtained at 13 hidden neurons. Afterward, training was

stopped when reached 78 epochs for the Levenberg–Marquardt algorithm as shown

in Figure 9. This is because of the difference between the training and validation

https://doi.org/10.17993/3cemp.2023.120151.257-271

3C Empresa. Investigación y pensamiento crítico. ISSN: 2254-3376

Ed. 51 Iss.12 N.1 January - March, 2023

266

errors increases. A plot of Levenberg–Marquardt algorithm regression for training,

validation, and testing with R is shown in Figure 10. This revealed that the R values

were 0.96, 0.94, 0.96, and 0.96 for training, validation, testing, and for all data

respectively. Thus, the optimum topology of the ANNs is obtained at 13 neurons in the

hidden layer with 4.28 MSE and 0.96 R. The architectural model of the optimum

topology is shown in Figure 11. This is 4:13:1, indicating the input layer with the used

four parameters, thirteen neurons at the hidden layers, and the output layer in terms

of the removal efficiency of the COD.

Figure 8. Mean Square Error (MSE) with different numbers of neurons at the hidden layers

for ANNs modelling of COD output.

Table 2. ANNs details for COD modelling.

COD

Model No. No. of neurons at

hidden layers

MSE Correlation Coefficient (R)

1 2 19.34 0.79

2 3 16.79 0.82

3 4 14.74 0.84

4 5 12.03 0.87

5 6 11.58 0.88

6 7 8.81 0.91

7 8 8.31 0.90

8 9 7.68 0.91

9 10 6.36 0.93

10 11 6.85 0.93

11 12 4.61 0.95

https://doi.org/10.17993/3cemp.2023.120151.257-271

267

3C Empresa. Investigación y pensamiento crítico. ISSN: 2254-3376

Ed. 51 Iss.12 N.1 January - March, 2023

Figure 9. Training, validation, and test mean square errors for the Levenberg–Marquardt

algorithm for COD removal.

Figure 10. Training, validation and testing regression for the Levenberg–Marquardt

algorithm for COD removal.

12 13 4.28 0.95

13 14 4.50 0.95

14 15 4.86 0.95

https://doi.org/10.17993/3cemp.2023.120151.257-271

3C Empresa. Investigación y pensamiento crítico. ISSN: 2254-3376

Ed. 51 Iss.12 N.1 January - March, 2023

268

Figure 11. The architecture of the ANN model for the prediction of COD removal.

CONCLUSION

The results of the ANNs modelling for the prediction of BOD, and COD provided

several advantages over the traditional calculation methods. The models were

accurately able to determine the removal efficiency of the BOD, and COD of Al-

Rustumiya WWTP by the use of raw dataset. The developed models can be

considered as simple, fast, and most accurate determination tools. This proved that

the developed MLP network trained with backpropagation incorporated with

Levenberg–Marquardt algorithm was adequate in predicting the performance of Al-

Rustamiya WWTP. For BOD and COD, the best results were obtained with 13

neurons at 3.09 MSE and 0.96 R for BOD and 4.28 MSE and 0.95 R for COD.

ACKNOWLEDGEMENTS

This research was carried out in Civil Engineering at Al-Nahrain University,

Baghdad, Iraq. The assistance and support of Al-Rustamiya WWTP Administration

Office Mayoralty of Baghdad are gratefully acknowledged.

CONFLICT-OF-INTEREST STATEMENT

The authors declare no conflict of interest for this research

REFERENCES

(1) Fahad A., Mohamed R. M. S., Bakar R., and Al-Sahari M. (2019). Wastewater

and its Treatment Techniques: An Ample Review. Indian Journal of Science

and Technology, 12(25).

https://doi.org/10.17993/3cemp.2023.120151.257-271

269

3C Empresa. Investigación y pensamiento crítico. ISSN: 2254-3376

Ed. 51 Iss.12 N.1 January - March, 2023

(2) Kılıç Z. (2021).

Water Pollution: Causes, Negative Effects and Prevention

Methods. IZU Journal of the Institute of Science and Technology, 3(2).

(3) Al-Attar R. T., Faris H. Al-Ani, and Al-Khafaji M. S. (2022).

SWOT-Based

Assessment of the Maintenance Management of the Wastewater Treatment

Plants in Iraq. Engineering and Technology Journal, 40(5), 677-694.

(4)

G. Sin, S. W. H. Van Hulle, D. J. W. De Pauw, A. van Griensven and P. A.

Vanrolleghem. (2005).

A critical comparison of systematic calibration

protocols for activated sludge models: A SWOT analysis. Water Research,

39(12), 2459–2474.

(5) A. A. Abbas, Y. T. Yousif and H. H. Almutter. (2022).

Evaluation of Al-Thagher

Wastewater Treatment Plant. Periodica Polytechnica Civil Engineering

, 66(1),

112–126.

(6) Han H. G., Qiao J.F., and Chen Q.L. (2012).

Model predictive control of

dissolved oxygen concentration based on a self-organizing RBF neural

network, Control. Eng. Pract., 20(4), 465–476.

(7) Olyaie E., Banejad H., Chau K.W., and Melesse A.M. (2015)

A comparison of

various artificial intelligence approaches performance for estimating

suspended sediment load of river systems: A case study in United States.

Environ. Monit. Assess., 187(4),189.

(8) Seo Y., Kim S., and Singh V.P. (2018).

Comparison of different heuristic and

decomposition techniques for river stage modelling.

Environ. Monit.

Assess., 190(7), 392.

(9)

Shamshirband S., Jafari Nodoushan E., Adolf J.E., Abdul Manaf A., Mosavi A.,

and Chau K.W. (2019). Ensemble models with uncertainty analysis for multi-

day ahead forecasting of chlorophyll a concentration in coastal waters.

Eng. Appl. Comp. Fluid., 13(1), 91–101.

(10)

Ahmed A. A. M., Hossain M. I., Rahman M. T. and Chowdhury M. A. I. (2013).

Application of artificial neural network models for predicting dissolved

oxygen concentration for Surma River, Bangladesh.

Journal of Applied

Technology in Environmental Sanitation, 3(3), 135–140.

(11) Nourani V., Kisi O., and Komasi M. (2011).

Two hybrid Artificial Intelligence

approaches for modeling rainfall runoff process. Journal of Hydrology

, 402,

41-59.

(12) Mohammed N. Y. (2016).

Assessment and Modelling of the Performance of

Al Wahda Water Treatment Plant. Master Thesis: University of Baghdad.

(13) Abrahart R., Kneale P. E. and See L. M. (2004).

Neural networks for

hydrological modelling. CRC Press.

(14) Ali ZH., Abd. Salman K., and Abd-Alateef K.A. (2001).

Evaluating the efficiency

of the filter in AlRustamiya wastewater treatment plant. Almustansriyah J.

12, 200-212.

(15) Rasheed Kh. A. (2016).

Study the Efficiency of two waste Treatment Plants

in the Al-Karkh and Al-Rusafa of Baghdad Region in 2015.

J. Int.

Environmental Application & Science, 11(2), 176-179.

(16) Dunia Frontier Consultants. (2013).

Water and Sewage Sectors in Iraq: Sector

Report. A report for JCCME prepared by Dunia.

https://doi.org/10.17993/3cemp.2023.120151.257-271

3C Empresa. Investigación y pensamiento crítico. ISSN: 2254-3376

Ed. 51 Iss.12 N.1 January - March, 2023

270

(17) Amanat Baghdad. (2021).

Baghdad Water Supply and Sewerage

Improvement Project. https://amanatbaghdad.gov.iq/page.php?

link=World%20Bank%20ads&lang=en

(18) Tümer A. E. and Edebali S. (2015).

An Artificial Neural Network Model for

Wastewater Treatment Plant of Konya.

International Journal of Intelligent

Systems and Applications in Engineering, 3(4), 131-135.

(19) Ammari A. (2016). MATLAB Code of Artificial Neural Networks Estimation.

ResearchGate.

(20) Alsulaili A. and Refaie A. (2021).

Artificial neural network modeling approach

for the prediction of five-day biological oxygen demand and wastewater

treatment plant performance. Water Supply, 21(5), 1861-1877.

(21) Shahin M. A. (2003).

Use of artificial neural networks for predicting

settlement of shallow foundation on conhesionless soils.

The university of

Adelaide school of civil and environmental engineering, 6-37.

(22)

Ibrahim M. A., Mohammed-Ridha M. J., Hussein H. A., and Faisal A. A.H. (2020).

Artificial Neural Network Modeling of the Water Quality Index for The

Euphrates River in Iraq. Iraqi Journal of Agricultural Sciences,

51(6),

1572-1580.

https://doi.org/10.17993/3cemp.2023.120151.257-271

271

3C Empresa. Investigación y pensamiento crítico. ISSN: 2254-3376

Ed. 51 Iss.12 N.1 January - March, 2023