PREDICTION OF MAXIMUM DAILY OZONE LEVELS USING NEURAL NETWORK MODELS IN BANGKOK

apricotpigletΤεχνίτη Νοημοσύνη και Ρομποτική

19 Οκτ 2013 (πριν από 4 χρόνια και 2 μήνες)

90 εμφανίσεις

PREDICTION OF MAXIMUM DAILY OZONE
LEVELS USING NEURAL NETWORK MODELS

IN BANGKOK

by

L.H. NGHIEM and N.T. KIM OANH

Environmental Engineering and Management, SERD,

Asian Institute of Technology
-

AIT

OUTLINE OF THE PRESENTATION


Overview

ozone

pollution

in

Bangkok


Overview

of

(ANN)

model

and

its

application

in

air

quality

modeling


Develop

a

Artificial

Neural

Network

model

for

predicting

daily

maximum

1
-
hr

ozone

levels

at

Bangkok

urban

area

for

ozone

season

(January
-
April)
.


Analysis

of

ozone

pollution

using

the

available

monitoring

data

shown

that
:



The

curbside

stations

are

characterized

by

lower

frequency

of

O
3

exceeding

the

Thailand

AAQS

than

general

ambient

stations
.



Highest

maximum

1
-
hr

hourly

O
3

levels

and

highest

frequency

of

O
3

exceeding

the

standard

were

recorded

at

ambient

stations

at

a

distance

from

the

city

center
.


OZONE POLLUTION IN BANGKOK (1)


There

are

13

stations

of

the

ambient

air

quality

monitoring

network

in

Bangkok
:

3

curbside

stations

+

10

general

ambient

stations


Only

11

stations

have

O
3

monitoring

data
.



OZONE POLLUTION IN BANGKOK (2)



High

O
3

pollution

in

Bangkok

occur

mainly

in

the

period

from

January

to

April

(winter

and

local

summer)

and

lowest

during

mid
-
rainy

season

(
Zhang

and

Kim

Oanh,

2002
)



Highest

O
3

concentrations

in

Bangkok

occur

in

the

period

from

13
:
00
-
15
:
00

(LST)
.



Ozone
season

in BKK

OVERVIEW OF ARTIFICIAL NEURAL NETWORKS
(ANN)


ANN

are

computer

programs

designed

to

simulate

biological

neural

networks

(e
.
g
.

the

human

brain)

in

terms

of

learning

(training)

algorithms
.


The

most

popular

topology

is

feed
-
forward

neural

network,

multi
-
layer

perceptron

(MLP)
.

An

overview

of

MLP

applications

in

atmospheric

science

can

be

found

in

Gardner

and

Dorling

(
1998
)
.



In

recent

year,

ANN

have

been

investigated

for

use

in

air

quality

modeling

and

given

acceptable

results

for

atmospheric

pollution

forecasting

of

pollutants

such

as

SO
2
,

ozone

and

PM
10
.



ANN

can

be

trained

to

identify

patterns

and

extract

trends

in

imprecise

and

complicated

non
linear

data

(allows

for

non
-
linear

relationships

between

variables)


Ozone

in

the

lower

atmosphere

is

a

complex

non
-
linear

process
.

Therefore,

ANN

is

a

well
-
suited

method

ozone

prediction
.

Processing at each hidden node:

1
.

Weight

input

variables

and

sum
:


2
.

Transform

sum

using

transfer

function
:


Processing at each output node:

1
.

Weight

the

transformed

hidden

layer

output

and

sum
:


2
.

Transform

sum

using

transfer

function

and

output

ozone

prediction
:



INPUT LAYER

OUTPUT LAYER

HIDDEN LAYER

1

1

Output
Node

2

2

M

I

W
o
mo

W
o
2o

W
h
12

i

m

I
1

Meteorological and
air quality input data

W
o
1o

W
h
11

W
h
1m

W
h
i1

W
h
i2

W
h
im

Ozone
prediction

Node (Neuron)

I
2

I
i

Transfer (activation) function

1
.

Logistic

Sigmoid

function

with

the

range

[
0
,

1
]
:

2
.

Hyperbolic

tangent

function

with

the

range

[
0
,

1
]
:







3
.

Gaussian

function
:



4
.

Linear

function
:



DEVELOPMENT ANN MODEL FOR PREDICTION
MAXIMUM 1
-
hr OZONE LEVEL IN BANGKOK

(
INPUT DATA)

I
.

Air

quality

data
:




Collected

from

Pollution

Control

Department

(PCD)

during

1

January

to

30

April

for

the

years

2000
-
2003
.



The

highest

values

of

daily

maximum

1
-
hr

O
3

concentrations

observed

among

of

the

monitoring

stations,

i
.
e
.

the

domain

peak

O
3




The

domain

average

values

for

THC,

NO

and

NO
2

between

6
:
00

a
.
m
.

and

9
:
00

a
.
m
.

of

the

day

of

interest
.

II
.

Meteorological

data
:

(
the

same

period

as

the

air

quality

data)



Observations

at

the

Bangkok

Metropolis

station

(in

the

city

center)

were

obtained

from

Thai

Meteorological

Department

(TMD)
.



The

selected

meteorological

variables

included

wind

speed

(m/sec),

wind

direction

(WDI),

relative

humidity

(
%
),

solar

radiation

(W/m
2
),

and

daily

maximum

temperature

(
o
C)
.



Utilized

the

average

values

for

the

selected

variables

between

6

a
.
m
.

to

10

a
.
m

of

hourly

observations

in

the

morning

of

the

day

of

interest
.

CONSTRUCTION OF ANN MODEL IN THIS STUDY


MLP

network

were

selected

to

develop

the

prediction

model

for

maximum

O
3

level
.



The

MLP

network

was

trained

using

Levenberg
-
Marquardt

back
-
propagation

of

MATLAB

Neural

Network

Toolbox
.



The

input

data

set

including

481

rows

(patterns)

were

RANDOMLY

split

into

two

sets
:

training

set

of

361

patterns

for

training

the

network,

and

the

remaining

dataset

of

114

patterns

for

the

testing

the

network
.




Number

of

hidden

layer

and

hidden

nodes,

and

connection

weights

between

neurons

of

the

MLP

network

were

determined

by

an

iterative

process

in

training

(learning)

stage





ARTIFICIAL
NEURAL
NETWORK
MODEL

THC

NO

NO
2

T
max

WS

WDI

Maximum O
3

level

Eight input
variables

RH

SOLAR

EVALUATION OF PERFORMANCE


OF THE ANN MODEL

Performance statistics


Mean Absolute Error (MAE)




Root Mean Square Error (RMSE)





Coefficient of Determination (R
2
)




Index of Agreement (
d
)

RESULTS

ARCHITECTURE AND PERFORMANCE OF MLP

Items

MLP Network

Architecture

8
-
8
-
12
-
1

8
-
8
-
14
-
1

8
-
10
-
12
-
1

8
-
10
-
14
-
1

Training

Testing

Training

Testing

Training

Testing

Training

Testing

MAE (ppb)

17.79

19.71

15.98

17.68

13.26

15.8

8.64

10.26

RMSE (ppb)

23.17

25.26

19.07

21.21

18.3

20.7

11.84

13.53

R
2

0.69

0.66

0.73

0.69

0.76

0.71

0.89

0.85

d

0.72

0.68

0.74

0.71

0.78

0.73

0.92

0.89

MLP 8
-
10
-
12
-
1

MLP 8
-
10
-
14
-
1

The Best MLP

RESULTS OF MLP NETWORK WITH
ARCHITECTURE 8
-
10
-
14
-
1

Comparison of predicted and observed ozone levels for the testing dataset

Scatter plot of predicted
versus observed values

Testing dataset

Training dataset

RESULTS OF LINEAR REGRESSION MODEL
(1)


The final regression model using stepwise procedure as follow:




Stepwise regression results is shown in the table:

Steps

Set of variables

Coefficient of correlation,
R
r
2

1

NO
2

0.200

2

NO
2
, T
max

0.273

3

NO
2
, T
max
, WS

0.315

4

NO
2
, T
max
, WS, RH

0.351

5

NO
2
, T
max
, WS, RH, THC

0.371

6

NO
2
, T
max
, WS, RH, THC, WDI

0.396

COMPARISON OF OZONE PREDICTION
MODELS ON THE TESTING DATA SET

Indicators

MLP

LR

Training

Testing

Training

Testing

MAE (ppb)

8.64

10.26

16.91

16.42

RMSE (ppb)

11.84

13.53

21.4

22.42

R
2

0.89

0.85

0.39

0.34

d

0.92

0.89

0.74

0.68

DISCUSSION


The

MLP

model

was

developed

for

Bangkok

urban

area

and

should

be

consider

specific

to

this

area
.



The

specific

model

was

developed

on

data

cover

period

January

1
-
April

30

ozone

season

in

Bangkok
:


Minimized

effect

of

season

factors

in

the

model



May

not

be

appropriate

to

use

the

model

for

other

seasons


The

results

of

ANN

model

in

Bangkok

are

with

in

the

range

of

the

results

reported

in

previously

published

studies

CONCLUSIONS AND RECOMMENDATIONS



This

study

shows

that

the

ANNs

can

be

used

in

air

pollution

modeling,

e
.
g
.

predicting

the

daily

maximum

1
-
hr

ozone

levels
.


These

ANNs

can

be

a

simple

alternative

model

to

provide

reliable

estimates

of

pollution

by

using

only

limit

information
.



Modification

and

improvement

of

the

models

should

be

done

to

develop

a

reliable

model

for

ozone

forecasting

in

Bangkok

urban

area
.