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

1
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

2
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).

3
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 O3 monitoring data.

• Analysis of ozone pollution using the available
  monitoring data shown that
• The curbside stations are characterized by lower
  frequency of O3 exceeding the Thailand AAQS than
  general ambient stations.
• Highest maximum 1-hr hourly O3 levels and highest
  frequency of O3 exceeding the standard were
  recorded at ambient stations at a distance from
  the city center.

4
OZONE POLLUTION IN BANGKOK (2)
Ozone season in BKK

• High O3 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 O3 concentrations in Bangkok occur in
  the period from 1300-1500 (LST).

5
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 SO2, ozone and PM10.
• ANN can be trained to identify patterns and
  extract trends in imprecise and complicated
  nonlinear 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.

6
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
7
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 O3
  concentrations observed among of the monitoring
  stations, i.e. the domain peak O3
• The domain average values for THC, NO and NO2
  between 600 a.m. and 900 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/m2),
  and daily maximum temperature (oC).
• 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.

8
CONSTRUCTION OF ANN MODEL IN THIS STUDY

• MLP network were selected to develop the
  prediction model for maximum O3 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

9
EVALUATION OF PERFORMANCE OF THE ANN MODEL

• Performance statistics
• Mean Absolute Error (MAE)
• Root Mean Square Error (RMSE)
• Coefficient of Determination (R2)
• Index of Agreement (d)

10
RESULTSARCHITECTURE AND PERFORMANCE OF MLP
The Best MLP
MLP 8-10-12-1
MLP 8-10-14-1
11
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
12
RESULTS OF LINEAR REGRESSION MODEL (1)

• The final regression model using stepwise
  procedure as follow
• Stepwise regression results is shown in the table

13
COMPARISON OF OZONE PREDICTION MODELS ON THE
TESTING DATA SET
14
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

15
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.