Prediction of Air pollution in Al-Hmadi City using Artificial Neural Network (ANN)

Journal of Environmental Treatment Techniques

2020, Volume x, Issue x, Pages: 1390-1399

J. Environ. Treat. Tech.

ISSN: 2309-1185

Journal web link: http://www.jett.dormaj.com

https://doi.org/10.47277/JETT/8(4)1399

Prediction of Air pollution in Al-Hmadi

City using Artificial Neural Network

(

ANN)

Al-Rashed, A. *, Al-Mutairi, N. and Boureslli A.

Science department, public Authority of Applied Education and Training, Kuwait. Email: ahmedbufarsan@gmail.com

Department of Civil Engineering, Kuwait University, P.O.Box 5969, 13060, Safat Kuwait

Science department, public Authority of Applied Education and Training, Kuwait

Received: 11/04/2020

Accepted: 27/08/2020

Published: 20/12/2020

Abstract

Neural networks application (ANNs) has been used at recent years in several technology areas. Applying artificial

neural networks (ANN) has too many environmental engineering problems provide remarkable success. The objective of

this project is to predict and show the accuracy of using neural network in air quality prediction in Al Ahmadi city.

Several parameters such as carbon dioxide (CO ), ammonia, benzene, relative humidity, temperature, air velocity and air

direction are used in this study, for period of four years. The correlation coefficient was used to indicate the performance

of ANN. The networks provide a good prediction performance with R value 0.93308 for CO , 0.81805 and 0.923 for

Ammonia and Benzene, respectively.

Keywords: Neural Network, Forecast, Air pollution, Quality

Introduction¹

electrical power plants in Kuwait reﬁneries, gases emission

from exhausts of cars, are major sources of pollution, [4-6].

The public health has been impacted by air smog for long

Kuwait Environmental Public Authority did

study between March 2011 and February 2012 for

Ahmadi area on NO and SO concentrations concluded

that winter had higher NO and SO readings than in the

time. Natural and man-made actions cause Harmful chemicals

emission to the environment which have great influence on

health and environment. In the last few years burning fuel

accelerated and become the main effect on changing air quality.

Poor air quality in cities is associated with rapid rate of great

deliberation of vehicular consume fuels. Air pollution

increased regularly and in vigorous and spread in vast

geographical locations. Know a day's industry considered vital

reason causing damage for the environment and air quality

particularly. Carbon monoxide, sulfur dioxide, Nitrogen oxide,

ozone and other air pollutants with different chemical

properties are in great effect to the environment either in a short

or long term. Diseases such, respiratory cancer and heart attack

resulted from gases spread in the open air [1] and [2].

According to the World Health Organization statistics, more

than two million people die every year. The main defendant for

such high death numbers is simply different gases emitted from

factories with all aspect of industrial activates. Percentage wise

more than from car accidents. In conclusion, Human must

implement the essential arrangements to reduce the pollution

caused by their harmful waste. As addition air pollution has

aggressive effect in people wellbeing correspondingly other

sides of the environment like optical potentials, plant life,

animals, lands and water feature [3]. The main factors that

should take in consideration to determine the level of air

pollution are the physical and chemical properties of pollutants,

topography and weather conditions such as wind, temperature,

air turbulence and rainfall. The emissions of gases form

summer. According to the study the source of nitrogen

oxides was the motor vehicles and the source of sulfur

dioxide was different industrial sources. Artificial

neural networks has an ability in forecasting unclear

data and effective use of this approach in different

domains encourage of using ANN to forecast air feature

depend on an ancient data.one of the proposes of The

study is the usage of neural network over years to

predict air quality. Quality Forecasting - also referred to

sometimes 'Atmospheric Diffusion Modeling. Data- is

simulated to predict how air pollutants disband in the

atmosphere. The simulation provides result represent

concentration for each air pollutants in air of the chosen

region. Experts develop different models of prediction

in many institutions and research labs around the world

successfully use one or several of them to realize the

real-time forecast simulation and effectiveness. The

main purpose of many studies is to find air quality

prophets based in neural networks to handle the narrow

amount of data groups and be durable for dealing with

records with faults and noise. The use of neural

networks has expended over the years in several

technology areas. Implementation of Artificial neural

network to many problems in environmental domain

have achieved satisfied results. Advanced models were

Corresponding author: Al-Rashed, A., Science department, public Authority of Applied Education and Training, Kuwait. Email:

ahmedbufarsan@gmail.com

390

Journal of Environmental Treatment Techniques

2020, Volume x, Issue x, Pages: 1390-1399

introduced for prediction based on annual and daily

data. These simulations could forecast the quality of air

with conservative reliability.

by a number for designs and a quantity for product / input

neurons, a volume with sound measurements, a type of system,

a kind for stimulation reaction was applied in a secret coating

[

10]. One concealed layer was capable of estimating several

nonlinear functions as well as levels of ins and outs [11]. By

training, in a hidden layer we can get the number, many

networks and valuing the consistent faults of the test group data

in the unseen sheet of few neurons and yield great testing errors

because of under-fitting and mathematical nuances. As

conflicting, many hidden layer neurons chief little training

fault, however high testing error, because of through

convenient and great variance [12].

Al-Ahmadi city

Al-Ahmadi city is located in the southeastern of

Kuwait (Fig. 1) with an area of 5,120 km ,apopulation

of (394,000) heavily polluted city due to rapid increase

in vehicles and industry. Such as oil refiners, causing

sufficient effect in air quality there. Kuwait Institute for

Scientific Research involved in different studies and

projects to monitor and evaluate air quality in hospitals,

factories and oil -fields in the area and it relation to

urban growth, such growth increase the demand on

different resources (electricity and water). Dust, gases

and other manufacturing pollutants resulted in a sharp

increase in a cancer among al-Ahmadi residents (Table

.1.2 Stimulation Function

A function is needed in network to build non-linear

connection concerning input and output factor, to be able to

attach and relate the factors [12]. In research for [7] and [9],

articles related to poor air quality were discussed confirming

which the hyperbolic sigmoid activation function seems to be

faster and efficient than logistic sigmoid activation function to

represent the nonlinearity between the hidden layer neurons

, Figure 2).

[13]. Build on the hyperbolic tangent feature for hidden

neuronal layers for this research. In addition, the ' individuality

feature ' was used for their specific purpose results in the entry

and output layer cells [9].

.1.3 Factor in Learning

Training is carried out on the multilayer neural network.

Preparation requires a custom information series comprising

input and connected production. Back propagation neural

network seems to be highly useful for training in the multi-layer

NN [14]. In the back-propagation training, η and µ are utilize

to ‘fast’ or ‘slow down’ the interchange the mistakes [15]. Its

learning scheme for back-propagation offers an "assessment"

of the heavy space route calculated by a Gradient Descent

Technique [16]. Significant lowering of the 'η' rate effects in

small changes in the weight of the synapse from one repeat to

the next and reduces the speed of training. However, its rise in

the 'η' speed increases the speed of practice of a system because

of big differences throughout neuronal measurements and

therefore creates a system when unbalanced. In back-

propagation teaching algorithm, the word 'μ' was adapted in

preventing system disturbance. The range of values 'π' and 'μ'

ranges from 0 to 1 [17] and [13].

Figure 1: Al-Ahmadi location in Kuwait

ANN modeling approach

The ANN designs have many benefits over old-

style semi-empirical simulations, although without

theory, it is essential to identify the information cluster

7]. This displays data about processing and is capable

[

of enhancing the depiction of parameters of input and

output. Using match inputs, this picture can be used to

forecast results [8]. The dual-layer ANN predicted just

about every measurable feature between results and

.1.4 Original Weights of the Network

Before beginning the instruction, the weights of neural

networks and free parameters are crucial. The network's main

weight and prejudice values help the learning processes to

converge rapidly. Throughout the present research, both

network distortion variables are laid to random amounts

equally distributed with the distance −2.4/Fi to + 2.4/Fi, where

Fi has been complete amount in outputs. A slightest variety of

allocation reduces the likelihood of neuron saturation of a

system to thus eliminating any mistakes happening [18].

input vectors by selecting

a suitable weight related

group and transferring any features [9]. This has

interrelated 'node ' or 'neuron' layers as shown in Fig.

.1 ANN Modeling Method

The method arrange procedure based on six serial stages:

i) choosing of ideal ANN-based model manner; (ii) choosing

stimulation factor; (iii) selecting the best parameters for

teaching, training rate and impulse frequency; (v) weights of

initialization and network preferences; (vi) testing as well as

analysis of the model; and (vii) model estimation.

(

.1.5 Practice and Monitoring

Its ‘supervised’ learning algorithm is mostly used to trained

the network. It is skillful by producing identified output and

input figures in a well-arranged method toward system [17].

Learning includes finding a number of network weights so that

the system can depict the basic models in the exercise

information. It is achieved by reducing its design mistake to all

models of input and linked result [7]. Its network ' setups ' ' in

teaching ' a coaching algorithm in ' over-training ' and ' local '

minima outcomes in mistakes of elevated design forecast [9].

.1.1 Ideal ANN Model Manner Choosing

Amount with variables initially equivalent to an amount for

neurons of the initial layer (i.e., variables are weather and

pollutant concentrations in the current effort). The resulting

layer includes single neuron, i.e. contaminant quantity. The

resulting layer includes single neuron, i.e. contaminant

quantity. Amount in neuronal in a secret coating is measured

391

Journal of Environmental Treatment Techniques

2020, Volume x, Issue x, Pages: 1390-1399

Figure 2: Location map of refineries in Alahmadi city

Table 1: some of industrial facilities in Ahmadi area.

The processing capacity

Petrochemical facility

Mina Al-Ahmadi (MAA)

Shuiba (SHU)

Number of Stacks

466,000 bpd

200,000 bpd

Mina Abdullah (MAB)

Kuwait Petrochemical

Industries Company (PIC)

EQUATE Petrochemical Company

270,000 bpd

The production capacity of urea is 1,040,000 TPA

The production capacity of ammonia is 620,000 TPA

Ammonia plant

includes 10 stacks

The annual production of this company is over 6 million tons

Too much training maybe prevented with testing a system

with subgroup of production/entry to specify measurements

with testing the rest of sample information and check the design

predictions precision [19]. As a consequence, it is definitely the

amount of 11. Learning algorithm for back-propagation has

been the best in modeling research of air pollution [7]. The

method separates all information to 3 sections: ' teaching

information, ' and ' test information set ' and the ' assessment

information set. Most of measurements used for training target

is formulated in the ' training data set; ' and ' test data set ' for

checking Simplification in NN system taught. Its learning

remains at

a standstill while 'sample information set'

introduction leads to the smallest design mistake. Finally, the

design is certified using the ' assessment information collection

(Dorling and Gardner, 1998). The back-propagation learning

algorithm's step-by-step method as shown below. The entry is

multiply by a random main weight and add the outcome as

follows:

P j= h/j=i∑ W I j ( xi+ b j ); i =1,2,…,n ; j=1,2,…,H=1

(1)

(1) where Pj is the input to the ‘j’ hidden layer neuron, xi the

numerical value of the (i) input layer neuron, (wi) the weight of

the (i) input layer neuron to (j) the hidden layer neuron, n is the

number of the input layer neurons, H the number of the hidden

layer neurons and (bj) is the favoritism value for the (j) th

hidden layer neuron. (ii) Change the hidden layer output by a

Figure 3: Common control system neural network construction

392

Journal of Environmental Treatment Techniques

2020, Volume x, Issue x, Pages: 1390-1399

sigmoid transfer function f (P j). (a) Logistic function:(b) Q1=

5.2 Software

,2/1+e-pj, j =1,2,….,H. (c) (b) hyperbolic tangent. Where Q(j)

MATLAB NN has been applied as adaptable as well as easy

start to improve quality of the air forecasting system. Range

data with NN enhancement is easily selected from the NN

Toolbox proposal.

is the output of the neuron with a hidden layer ‘j’. (iii) Multiply

the hidden output layer weight by the hidden layer outputs and

sum as:

RK = h/ji∑ W j K Q j+ B k K= I j; = 1,2,…,H

5.3 Air Quality Prediction Using ANN Model

As network construction, a 3-layer perceptron model was

used. Initial entry coat comprises the entry data from network.

In the input layer there were seven neurons as well as three

where R k is the input to the (k) output layer neuron, w j k the

weight of the (j) the hidden layer neuron to the k output layer

neuron, m the number of the output layer neuron and b k is the

preference value for the k output layer neuron. (i) Convert the

output, R k by the transferal function to achieve the network

out puts (YK). (ii) The network outputs are then related with

observed values, and an error at the k output neuron is

calculated: EK=Ti-Y k. Where T k is the practice (real) value.

The general standard used in the back-propagation learning

method is the ‘delta rule’, which is founded on the decreasing

the sum of squares of the error obtained in Equation.

contaminants which is CO , ammonia and benzene four

variables which is relative humidity, temperature, air pressure,

and wind velocity. Amount with secret layers the limitations to

be selected in the model are the quantity for neurons for a secret

layer. Therefore, to enhance Artificial Neural Network

efficiency, one or two hidden layers as well as varying neuronal

measures have been selected. Final surface was production

surface that is the estimation of template focus. CO , ammonia

and benzene were used as the variables in the output. Sooner

than usual end, the registry has been separated in to three parts.

Fifty percent of information have been used for system

teaching, twenty-five percent for verification, as well as the

residual twenty-five percent for network testing. Correlation

coefficient R has been selected for numerical parameters for

calculating system efficiency.

Objectives

The objectives of this study are: (a) to predict

concentrations of Benzene, CO and Ammonia in Ahmadi; (b)

to analyze the predicted concentrations of pollutants; and (c) to

compare the predicted concentrations with the standard limits

set by Kuwait Environment Public concentrations of pollutants

within the study area domain.

6 Results and Discussion

.1 ANN Modeling

In this study Feed-forward neural network was applied. The

Materials and Methods

number in the hidden layer were 21. The biases and weights

were changed due to gradient-descent back-propagation in the

training phase. The correlation coefficient R which is statistical

measure was chosen for determining of the network efficiency.

.1 Data Sets

In this study measurements involved includes daily outside

temperature, relative humidity, air pressure, wind direction and

speed, and daily concentration of CO , ammonia and benzene

The training, validation and test for CO in fig. 4, 5, 6 and 7.

in Al Ahmadi during 4 years period from 2014 to 2017. The

data were delivered by (DGCA) Department of Meteorology.

Measurements have been distributed in two sets, with artificial

neural network learning as well as evaluating are validating its

effectiveness and accuracy.

While the correlation coefficient R of Ammonia shown in

figures 8, 9, 10 and 11. The correlation coefficient R of benzene

shown in figures (12, 13, 14 and 15).

Figure 4: Correlation coefficient R of CO

for validation

Figure 5: Correlation coefficient R of CO for training

393

Journal of Environmental Treatment Techniques

2020, Volume x, Issue x, Pages: 1390-1399

Figure 7: Correlation coefficient R of CO

for test

Figure 6: Correlation coefficient R of CO

Figure 9: Correlation coefficient R of ammonia for test of ammonia for

validation

Figure 8: Correlation coefficient R for all

Figure 10: Correlation coefficient R of ammonia for test

Figure 11: Correlation coefficient R of ammonia for all

394

Journal of Environmental Treatment Techniques

2020, Volume x, Issue x, Pages: 1390-1399

Figure 13: Correlation coefficient R of benzene for validation

Figure 15: Correlation coefficient R of benzene for training

Figure 12: Correlation coefficient R of benzene for training

Figure 14: Correlation coefficient R of benzene for test

Throughout the exercise, the figure appears. It shows the

network time regression versus the performance evaluation has

been implemented with examination in relation depends on the

value of the correlation coefficient both real as well as expected

outcomes, R. Good correlation the value of the training was

indicated by a dataset as well as output R. In a regression plot,

the appropriate fit shows good correlation between the

predicted and targets is indicated by the solid line. The best

fitting produced by the algorithm represented by dashed line.

As shown, that model provide high value of correlation

coefficient R, which represents precision within measured as

well as predicted values. In addition, the percentage difference

was obtained between the real and predicted concentration of

pollutants to show the accuracy of the ANN model by using

equation. Percentage diff. = (predicted- real)*100/real. The

results shown in the table 2 reflect the high precision of the

ANN model of prediction. Figure 16 and Figure 17 represent

the daily concentrations of CO and O that measured by Kuwait

environment public authority. In addition comparing the

predicted concentrations with the standard limits set by Kuwait

Environment Public as shown in Table 3 and 4, concentrations

of pollutants within the study area domain shows that the

pollutants concentrations results less than the standard set by

KEPA, where ammonia (NH ) was 23.94 ppb and benzene

(C ) was 0.697 ppb.

Table 2: Percentage difference between predicted and real

value

Pollutants

Percentage diff. %

-.005

Ammonia

Benzene

3.572

9.3695

395

Journal of Environmental Treatment Techniques

2020, Volume x, Issue x, Pages: 1390-1399

Table 3: Ambient Air Quality Standards for Residential Areas

Unit Hour* 8 hours Day**

Year

Pollutant

Ppb

µg/m³

Ppb

µg/m³

Ppb

µg/m

Ppb

µg/m³

Sulfur Dioxide (SO

Nitrogen Dioxide (NO

Carbon Monoxide (CO)

Ozone (O

Suspended particulate matter (PM-10)

)

170

444

225

34000

157

112

9000

)

100

3000

10,000

11500

120

8000

)

350

Average hour should not occur more than twice during the period of 30 days on the same site.

Daily average (24 hours) should not occur more than once during the year.

Should not occur more than once per year. – Should apply in residential dominated areas that lie on the border of industrial areas.

Table 4: Criteria Pollutants

No.

Pollutant

Averaging Time

Limit

800 ppb

144 ppb

100 ppb

20 ppb

Classification Method

Hour

4 Hour

Hour

H S

4 Hour

Annual

Hour

1.6 ppb

0.24 ppm

NMHC

------

Where; A_99 percentile of maximum concentration over on year, B- Not exceed more than 1 time per year in the same location, C- Not to

exceed, D- Not exceed more than 3 times per year in the same location.

Figure 16: Concentration of O

(ppm)

Figure 17: Concentration of CO (ppm)

.2 Optimal Back-Propagation Network Interface Setup

For this analysis, ANN 's task for estimating trends of

initial weight was examined. Normally, as the error in the test

collection rises, the testing cycle stops. The efficiency of

forward propagation is influenced by all training sets (testing

and predicting), irrespective of their scale. The forward

propagation algorithm was required to determine the optimal

values for the parameters of the neural network, such as the set

percentage, number of concealed neurons, number of initial

weights and learning rate and momentum rate.

Logiest networks and Gaussian functions have been

equipped, evaluated and predicted with 0.1 for both learning

and momentum. A forward propagation algorithm is used. The

weight was 0.3 in the first place. The estimated average errors

of 4 and 5 ppm respectively for the expected and preparation

compartments. It was important to ensure, rather than saving

from training examples, that each network knew the functional

relations between the variables input and output. The data sets

changes in concentration of pollutants is to use a forward

propagation algorithm. The daily outside temperature, relative

humidity, air pressure, wind direction and speed parameters

were used as input data in the model to show the potential of

the proposed ANN for modeling the pollutants concentration.

There were 120 variations. Based on the minimum MAE of the

training and prediction package, the real design and the

parameter variance for the forward propagation model were

chosen. The following approach has been adopted to simplify

the forward propagation design. First, the model form has been

designed with the activation function and the test scale. Third,

for the maximal forward propagation model the learning and

momentum levels were significant. Second, there is an

optimization of the amount of neurons in the hidden layers. The

396

Journal of Environmental Treatment Techniques

2020, Volume x, Issue x, Pages: 1390-1399

were combined allegedly for this reason and ultimately divided

into three subsets of data (training, testing, and prediction). The

optimal design for the forward propagation model has been

learned by six different neural network architectures. The 6

architectures of the forward propagation were the following: 1)

one overloaded layer; 2) two series overshadows; 3) three

series overshadowed layers; 4) two parallel overshadows; 5)

three parallel overshadows; and 6) two parallel layers with hop

links. The optimum parameters are based on the minimal MAE

as obtained from the outcomes of the preparation, testing and

prediction data subsets. Table 5 shows MAEs from the six

network architectures, with 4 separate evaluation percentages

with 5 %, collected at the end of 10.000 testing epochs. The

Test Collection of 5% is the following distributions with one

array: [80-9-27]; e.g. 80 data testing subsets, 9 data evaluation

subsets, and 27 data estimation subsets. Table 5, also indicates

that the MAE forecast was somewhat higher than the MAE

testing but was in general of the same magnitude suggesting

sufficient network preparation. The presence of this significant

prediction bias is a large gap in the network's predictive

efficiency. For each of the six forward propagation architecture

the number of neurons exceeded 25, the average errors

decreased significantly, from 9.2 to 2.8 and from 12.4 to 5.1,

for both the training and the prediction sets, respectively. The

neural network containing 30 hidden neurons was chosen as the

best case (Table 6). In addition to the prior parameters, the

initial weights, the learning rate, and the momentum rate were

the other important parameters that affected the training

effectiveness of the forward propagation. However, for the

current case, the adjustment of the initial weight from 0.05 to

0.3 did not affect the MAE of the training and prediction

networks. At the initial weight of 0.5, the MAE increased to 8

and 11.0 for the training and prediction networks, respectively.

Above the initial weight of 0.5, the network error improved

(i.e., decreased) significantly (Table 5). Furthermore, varying

the momentum rate and the learning rate from 0.05 to 1, did not

affect the error of the training and the prediction networks.

Hence, the initial values of the learning and the momentum

rates, as set by the software, were used for all training networks.

Based on the results of the analyses, the optimal configuration

of the ANN model was chosen, as summarized in Table 6. It

can be seen from the data in these figures that the results of the

model predictive ability were exceptionally good.

based on the MAE , Table 6 presents the optimal test set. The

designed architectural framework was found to be the forward

propagation with two hidden parallel layers (Architecture 4),

which had a 10% check scale and logistics transition (TF).

Table 6, demonstrates also the influence of the MAE on

training and data estimation for the same optimum layout (10

percent scale and logistic TF) of the numerical secret neurons.

It can be easily seen that the MAE of the network was much

higher for the two and five numbers of neurons than those with

6.3 Review of Sensitivity

A sensitivity study was also performed to test the general

capability of the optimum forward propagation model, by

estimating the significance of each input variable individually.

This analysis was done by omitting each input from the model,

retraining the model, and calculating the percentage change in

T P

both the training (MAE ) and the prediction (MAE ) error

functions from that of the original model.

, 10, 15, and 20. At 25 hidden neurons, the error increased

significantly for both the training and the prediction sets. When

Table 5: Effect of transfer function on mean–absolute error (MAE) for 5% test set

Logistic TF

MAE

Training

Gaussian TF

Forward propagation

architecture

Model Number

MAE

Training

Prediction

One hidden layer

Two hidden layers

in series

Three hidden layers

in series

Two parallel hidden

layers

Three parallel

hidden Layers

Two parallel hidden

layers

6.6

8.0

10.0

with a jump

connection

Table 6: Optimum Test Set for Trained and Predictive Forward–Propagation Neural Network (BPNN) Architectures with Logistic TF

at 10,000 Training Epochs

Model

Number

Training

Optimum Test set (%) MAE

Prediction

Optimum Test set (%) MAE

Forward propagation architecture

One hidden layer

Two hidden layers in series

Three hidden layers in series

[76-31-23]

[80-9-41]

[60-16-54]

4.2

8.5

6.7

[74-34-22]

[90-16-24]

[80-17-33]

3.8

10.8

9.0

Two parallel hidden layers

[72-17-41]

[92-34-4]

[62-31-37]

4.0

6.0

[65-40-25]

[93-23-14]

[79-34-17]

5.0

8.5

5.0

Three parallel hidden Layers

Two parallel hidden layers

with a jump connection

397

Journal of Environmental Treatment Techniques

2020, Volume x, Issue x, Pages: 1390-1399

Table 7: Effect of Omitting a Specific Input from Forward–Propagation Neural Network Model on Mean–Absolute Error (MAE)

Missing input

% increase rate in MAE

Input Strength

0.25

0.02

0.11

0.10

Outside temperature

Relative humidity

Air pressure

Wind direction

Wind speed

0.61 68

0.82 26

0.61 46

0.65 54

0.45 49

0.76 49

0.27

0.58

Outside temperature and Wind direction

The change in the error function is a direct measure of the

predictive ability, only if the input variables are uncorrelated.

Table 7 shows the percent change in the training and in the

prediction model when an input variable is omitted from the

neural network. The daily outside temperature, wind direction

and wind speed, appear more important than the relative

humidity, air pressure and wind speed parameters, for

predicting the level of pollutants concentration. The least

important variables are the humidity and air pressure. However,

it would be incorrect to conclude that humidity and air pressure

are jointly unimportant. When both outside temperature and

wind direction are omitted from the neural network model, the

percent change in MAE was 35 for the training and 49 for the

prediction, which are lower than the percent change resulting

from omitting the outside temperature, wind direction, or wind

speed.

reflects a good association between the targets and expected

outputs. In other wise comparing the predicted concentrations

with the standard limits set by Kuwait Environment Public

authority (KEPA) concentrations of pollutants within the study

area domain shows very good results, where ammonia (NH )

6 6

was 23.94 ppb and benzene (C H ) was 0.697 ppb.

Ethical issue

Authors are aware of, and comply with, best practice in

publication ethics specifically with regard to authorship

(avoidance of guest authorship), dual submission, manipulation

of figures, competing interests and compliance with policies on

research ethics. Authors adhere to publication requirements

that submitted work is original and has not been published

elsewhere in any language.

When outside air pressure was omitted from the model, the

wind direction and wind speed’ weight increased to

compensate for the missing outside temperature, and when the

wind direction and speed were omitted from the model, the air

pressure weight increased to compensate for the effect.

However, when the outside temperature was omitted from the

model, no such compensation took place by any of the

remaining variables in the model. Thus, the variable, outside

temperature was more important for the training and for the

prediction sets than were either wind speed or wind direction,

when considered individually. It is clear from the sensitivity

analysis that the outside temperature, wind direction, and wind

speed have different effects on the level of pollutant

concentration, and for most practical purposes, the seasonal

variation (summer vs winter) would be considered more

important than the rest of the parameters. It is also obvious from

the analysis result that the outside temperature is associated

with a much larger level of change in the output than are either

the relative humidity or air pressure; for most practical

purposes, the outside temperature and wind speed would be

considered the most important input variables in the model

because they have a pronounced effect over a large interval. Air

pressure and wind direction were moderately important

because they have a small effect over a large interval. Finally,

the relative humidity was the least important, because they have

a small effect over a small interval.

Competing interests

The authors declare that there is no conflict of interest that

would prejudice the impartiality of this scientific work.

Authors’ contribution

All authors of this study have a complete contribution for

data collection, data analyses and manuscript writing.

References

https://www.ontario.ca/page/ontarios-ambient-air-quality-criteria-

sorted-chemical-abstracts-service-registry-number. Accessed

July 2017.

U.S. Environmental Protection Agency, "Protection of

environment”. http://www.ecfr.gov/cgi-bin/text-

idx?c=ecfr&SID=bdd20ec807fda95c550270c4328e6306&rgn=di

v9&view=text&no de=40:6.0.1.1.6.7.1.3.38&idno=40. Accessed

20 June 2017.

Kumar A, Singh BP, Punia M, Singh D, Kumar K, Jain VK.

Assessment of indoor air concentrations of VOCs and their

associated health risks in the library of Jawaharlal Nehru

University, New Delhi. Environmental Science and Pollution

Research. 2014 Feb 1;21(3):2240-8.

Ramadan A. Assessment of spatial variation of ambient volatile

organic compound levels at a power station in Kuwait. Journal of

the Air

& Waste Management Association. 2017 Nov

2;67(11):1170-9.

Ramadan A, Al-Rashidi M. Atmospheric monitoring at an oil field

in northern Kuwait using diffusive passive samplers. International

Journal of Environmental Science and Technology. 2019 Jun

6;16(6):2809-20.

Conclusion

The current effort provides animated data and given by

using the techniques ANN, Multilayer neural net-work Used

for short-term development in Artificial Neural Network-based

CO , ammonia, benzene for the air-quality forecasting

purposes in Al Ahmadi. The daily time series of these polluted

concentrations, meteorological variables were collected from

6. Al-Awadhi JM. Measurement of air pollution in Kuwait City using

passive samplers. Atmospheric and Climate Sciences. 2014 Apr

;4(02):253.

3. Gardner MW, Dorling SR. Artificial neural networks (the

multilayer perceptron)—a review of applications in the

atmospheric sciences. Atmospheric environment. 1998 Aug

014 to 2017. Based on the present work, applying neural

;32(14-15):2627-36.

network provide the satisfied efficiency of prognosis of air

pollution with value of R which represents the estimate

precision. As shown values of R of CO , Ammonia and

Benzene respectively are 0.93308, 0.81805 and 0.923. Which

14. Schalkoff, R..”Pattern recognition: statistical, structural and neural

approaches”. John Wiley & Sons, Inc. New York, NY, USA.,

384pp, 1992.

398

Journal of Environmental Treatment Techniques

2020, Volume x, Issue x, Pages: 1390-1399

5. Gardner MW, Dorling SR. Statistical surface ozone models: an

improved methodology to account for non-linear behaviour.

Atmospheric Environment. 2000 Jan 1;34(1):21-34.

6. Alsugair AM, Al-Qudrah AA. Artificial neural network approach

for pavement maintenance. Journal of computing in civil

engineering. 1998 Oct;12(4):249-55.

7. Hornik K, Stinchcombe M, White H. Multilayer feedforward

networks are universal approximators. Neural networks. 1989 Jul

;2(5):359-66.

8. Sarle WS. Neural network frequently asked questions. ftp://ftp.

sas. com/pub/neural/FAQ. html. 1998.

9. Rege MA, W. Tock R. A simple neural network for estimating

emission rates of hydrogen sulfide and ammonia from single point

sources. Journal of the Air & Waste Management Association.

996 Oct 1;46(10):953-62.

0. Haykin S. Neural networks: a comprehensive foundation.

Prentice-Hall, Inc.; 2007 May 1.

1. Rumelhart DE, Hinton GE, Williams RJ. Learning internal

representations by error propagation. California Univ San Diego

La Jolla Inst for Cognitive Science; 1985 Sep.

2. Battiti R. First-and second-order methods for learning: between

steepest descent and Newton's method. Neural computation. 1992

Mar;4(2):141-66.

3. McClelland JL, Rumelhart DE, PDP Research Group. Parallel

distributed processing. Explorations in the Microstructure of

Cognition. 1986;2:216-71.

4. Wasserman PD. Neural computing: theory and practice. Van

Nostrand Reinhold Co.; 1989 Apr 1.

5. Comrie AC. Comparing neural networks and regression models for

ozone forecasting. Journal of the Air & Waste Management

Association. 1997 Jun 1;47(6):653-63.

399