Chapter1 by taoyni


									  Meteorology-based Forecasting of Air Quality Index Using Neural Network
           Mukesh Sharma,* Sachin Aggarwal*, Purnendu Bose* and Ashok Deshpande+
                *Department of Civil Engineering, Indian Institute of Technology
                                     Kanpur, 208016, India
             +Distinguished Professor SIES-Indian Institute of Environment Management,
                                  Navi Mumbai, 400 076 India

Abstract: Air Quality Index (AQI), a system for transforming air pollution levels into a single number, aims at
providing information about air quality in simple terms to general public. Any advance information about AQI can
forewarn the public of unhealthy air and encourage people to voluntarily reduce emissions-producing activities and
avoid exposures to polluted environment. Two mathematical models (i) meteorology-based air quality level
predictions and (ii) meteorology forecasting, have been developed (based on four year data) using neural network to
forecast AQI for following three days. The AQI forecasting model was concluded as being satisfactory and useful
for               information               dissemination              to              general               public.

                                                          between the metrology and air quality in terms of
Introduction and Objective                                air quality index and develop a forecasting
                                                          mechanism using artificial neural network for the
Air pollutants at ground level can be harmful to          developed AQI.
human health if their concentrations exceed
certain acceptable levels. As pollutants                  A scheme utilizing neural network simulation is
accumulate in or near large metropolitan areas,           developed in the course of this work for
this typically exposes people to unhealthy                prediction of AQI for a location in the city of
pollutant concentrations In light of the health           Kanpur (longitude 88° 22' E and Latitude 26° 26'
effects of air pollutants, environmental agencies         N), India. The neural network model is trained
(e.g. USEPA, 1998) have been using air quality            and tested using the air quality and meteorology
index (AQI) for public information (about air             data. Two neural network models were
quality) and data interpretation. AQI is defined          developed one for developing a meteorology -air
as an overall scheme that transforms the                  quality linkage and second for forecasting the
weighted values of individual air pollution               meteorology three days in advance. Sampling
related parameters (e.g. SO2, CO, visibility, etc.)       location at Agricultural University (AU), Kanpur
into a single number or set of numbers. An AQI            was selected for this study as concurrent data of
is developed in Indian context (Sharma et al              air quality (from National Ambient Air quality
(2001)) which include three pollutants (sulphur           Monitoring Programme) and meteorology were
dioxide (SO2), nitrogen dioxide (NO2) and                 available for last several years at this site, which
Suspended       Particulate     Matter   (SPM)).          were essential for this study.
Segmented linear functions are used for relating
the actual air pollution concentrations (of each
pollutant) to a normalized number. The
                                                          Meteorology-Air            Quality         Model
categories of index system are:                           Development

0-100; Good, 101-200; Moderate, 201-300;                  Air pollution system has three components:
Poor, 301-400; Very poor, 401-500; Severe                 emission source, transport medium (atmosphere)
                                                          and receptor. Pollution reaching a receptor
Ideally, the developed AQI should have accurate           depends not only on the emitted quantity but also
forecasting capability. Any advance information           on the atmospheric dynamics. The impact on
about AQI can forewarn the public of unhealthy            receptor can be estimated by developing source-
air and encourage people to voluntarily reduce            receptor linkages through atmosphere.
emissions-producing activities and avoid
exposure to polluted environment. The purpose             A modeling system (source-receptor linkages) is
of this paper is to establish a mathematical model        necessary to predict impact on the receptor and
                                                          translate the impact into an air quality index.

Steps in Model Formulation                                monsoon and post monsoon). In other words, for
                                                          a given season and meteorological data, the
The impact on receptor at any point is a function         model should predict concentration of the
of source strength, and meteorological                    pollutants. For this purpose, a neural network
parameters like wind speed, wind direction,               model (presented below) was developed.
temperature, relative humidity, stability, rainfall
etc. Thus, the AQI forecasting modelling                  Artificial Neural Network (ANN) Model for Air
completion is a three-step process.                       Quality Modelling

(i) Air Quality Models: Establish the model that          For construction of neural network model, the air
can predict the air quality for various pollutants        pollution system is looked upon as a system that
in terms of meteorology                                   under varying sets of meteorological inputs (e.g.
                                                          weather conditions) will respond by producing
(ii) Meteorology Forecasting Models: Forecast             different sets of output. Such a model
meteorology for coming next days and combined             presupposes no prior knowledge about the
the forecasted meteorology to forecast air quality        structure of relationship that exists between input
using the model developed in step (i) above               and     output      variables     (e.g.   pollutant
(iii) AQI Forecasting: Estimate and forecast the
air quality index based on forecasted air quality         Neural network comprises a number of
levels in step (ii) above.                                interconnected entities, similar in many ways to
                                                          biological neurons. The choice of the
                                                          architecture of the network depends on the task
Air Quality Models
                                                          to be performed. For modelling physical system
                                                          such as air pollution system, a feed-forward layer
Since air quality data were available on three
                                                          is normally employed (Wasserman, 1989). It
pollutants, SPM, SO2, and NO2, these pollutants
                                                          consists of layers of input neurons, and one and
are selected for model prediction and index
                                                          more hidden layers. For this study, a software
                                                          "Neuro Genetic Optimizer" (NGO, version 2.6)
                                                          is used. The modelling results are shown in
 Four-year air quality data for the site                  Figure 1. The other details of model are given
(Agriculture University, Kanpur) for the years            below.
1997 to 2000 were collected from Central
Pollution Control Board, Delhi. Similarly,
                                                          ANN model - Steps Followed
meteorological data for the same time period and
at the same location were also collected.                 1. The air quality data were categorized into
Meteorological data included temperature, wind            four seasons as explained earlier.
speed and direction, relative humidity and                2. The meteorology data were regressed with
rainfall.                                                 air quality data. Only average wind speed and
                                                          average temperature of the day showed
Initial data analysis showed significant seasonal         significant correlation with air quality. Hence,
variability in air pollutant concentrations.              average temperature and wind speed of day was
Monsoon period was the cleanest and winter                only taken as input layer. Output layer consisted
months were the most polluted. Therefore,                 of concentration of SO2, NO2, and SPM
meteorological and air quality data were                  concentration.
segregated as per the following seasons:                  3. Four models one for each season were
                                                          developed for each pollutant. For this purpose,
Winter: December, January, and February,                  data set of each season for four years was
march                                                     randomly divided into training and testing record
Summer: April, may, June                                  (50% for each).
Monsoon: July, August, September and                      4. Input data was normalized between -1 and 1
Post monsoon: October, November                           and output was subsequently de-normalized
In all, twelve models need to be developed for            5. Weights were randomly initialized between
predicting air quality concentrations for each set        +0.3 to -0.3.
of meteorological parameters. The twelve                  6. Momentum constant was selected in the
models include three models (for SO2, NO2 and             range 0.1 to 0.3
SPM) for four seasons (winter, summer,

7. Multiple hidden layers were used; their                                                            end, two neural networks were developed one for
number was selected using genetic algorithms.                                                         forecasting mean temperature and another for
8. Selection of hidden layer and number of                                                            wind speed. It may be stated that meteorology is
hidden neurons was decided using genetic                                                              a very complicated phenomenon and explaining
algorithms                                                                                            it on basis of simple mathematical models is very
9. Limit on the number of hidden neurons was                                                          difficult and one has to resort to ANN. The
set as eight.                                                                                         following steps were undertaken to develop two
10. The minimum number of passes for each                                                             neural networks-based forecasting model one
network was 20 and maximum was 50.                                                                    each temperature and for wind speed.
11. Learning rate constants were selected in the
range 0.1 to 0.4.                                                                                                          Collection      of      data      :     four      years         of
                                                                                                                                      MONSOON MODEL FOR SO2
                               MONSOON MODEL FOR NO2                                                                                                                          R = 0.7814
                                                                                   R = 0.7613                          25

                                                                                                       PREDICTED SO2

                   25                                                                                                  15
                   10                                                                                                   5
                    5                                                                                                   0
                    0                                                                                                       0.0       5.0        10.0        15.0      20.0          25.0
                        0.0   5.0   10.0      15.0                   20.0   25.0       30.0                                                       ACTUAL SO2
                                           ACTUAL NO2
                                                                                                                            meteorological data for various parameters
                                                                             MONSOON MODEL FOR SPM
                                                                                                                                                R = 0.4262
                                              PREDICTED SPM

                                                                    0.0     50.0        100.0       150.0                     200.0     250.0      300.0
                                                                                                ACTUAL SPM

Figure 1 Neural Network Modelling of NO2, SO2                                                                               like temperature, wind speed and direction,
and SPM                                                                                                                     relative humidity, rainfall etc were collected,
                                                                                                                           Preparation of data: for advance forecasting
Figure 1 shows model performance and values of
                                                                                                                            for three days, input data for each set of
coefficient of correlation (R). The R-values were
                                                                                                                            output was prepared. A value of previous
found significant at 1% level of significance in a
                                                                                                                            four days was found adequate to forecast
statistical sense and model was accepted for
                                                                                                                            temperature and wind speed for next three
further analysis. It can be concluded that the
                                                                                                                            days. Therefore, the data were arranged in a
developed neural network model establishes a
                                                                                                                            set of four; total 1100 such sets were made
reasonable relationship between meteorological
                                                                                                                            and fed to the ANN model, and
inputs and air concentrations.
                                                                                                                           Other steps were identical to those presented
                                                                                                                            in the section on air quality modelling.
Meteorology Forecasting Model
                                                                                                      Based on the above exercise two models were
As identified that for predicting air concentration                                                   developed. Model for temperature showed much
only average temperature and wind speed of the                                                        higher accuracy in prediction than wind speed
day were adequate for predicting air                                                                  (Figure 2)
concentrations. It was decided that attempt
should be made to forecast AQI for next three                                                         The temperature forecasting results are excellent,
days to provide enough warning period to people                                                       and it demonstrates the high predictive capacity
likely to be exposed from air pollution. To this                                                      of ANN. Both for temperature and wind speed,

    the value of R was found significant at 1% of                                like good, moderate, poor, severe etc. The model
    level of significance, and model was accepted for                            strength must be tested on predictability of AQI
    further analysis.                                                            in the right band rather than in terms of absolute
    AQI     Forecasting:                              Results     and
    Discussion                                                                   Figure 4 compares the model-computed AQI
                                                                                 band vis-a-vis observed AQI bands. Considering
                                                                                 the AQI bands, it can be stated that the
    Having established two models one fore
                                                                                 performance of the model improves and it can be
    predicting pollutant concentration based on wind
                                                                                 used for public information as an advance
    speed and temperature and other model for
                                                                                 warning system. Results of other seasons are not
    forecasting wind speed and temperature, the
                                                                                 shown here. It may be mentioned that for
    values of model-computed AQI can then be
    estimated for all days of air quality monitoring.
    Figure 3 compares the AQI values based on                                                                           WIND SPEED FORECASTING DAY 1
    actual measurements vis-à-vis model computed
    AQI.                                                                                                                                                        R2 = 0.569

                                                                                           PREDICT ED VALUES
                          TEMPERATURE FORECAST FIRST DAY                                                       40
                                                                  R = 0.9447                                   30
                   50                                                                                          20

                   40                                                                                          10

                   30                                                                                          0
                                                                                                                    0       5      10         15           20         25
                                                                                                                                  ACYUAL VALUES

                   0                                                             summer and winter the model has performed
                        0.0        10.0        20.0        30.0         40.0
                                          ACTUAL VALUES                                                         ACTUAL Vs PREDICTED AQI FOR MONSOON
                                                                                                                                                                R = 0.5731
                                        Figure 2 Neural Modelling
                                                                                  PREDICTED AQI

                              results for temperature and wind speed                                      120
                              forecasting                                                                 100
    The model does not show very high correlation
    between computed and observed AQI (Figure 3).                                                          20
    The probable reason for such a behavior can be                                                          0
    attributed high background concentrations and                                                                   0             50                 100                   150
    low values of pollutant concentration that may be                                                                                   ACTUAL AQI
    fluctuating from one day to another and model is
    not capable of picking these subtle changes.                                 better than monsoon season (Aggarwal, 2001).
    However, the aim of this study is to develop a
    method, which could predict AQI into a band
                                                                                Figure 3 Comparison of AQI predicted and Actual
                                                                               for Monsoon.


                           Percent Category
                                               60%                               Poor
                                               50%                               Moderate
                                               40%                               Good
                                                     Observed       Forecasted

                 Figure 4. Comparison of Observed and predicted AQI Bands For Monsoon season.

Aggarwal Sachin (2001). Application of Neural
Network to Forecast Air Quality Index. Thesis
submitted in partial fulfillment of requirements
for a degree in Bachelor of Technology, April

NGO (version 2.6) (2000). Neuro-Genetic
optimizer URL

Sharma M., Sengupta, B., Shukla, B.P. and
Maheshwari, M (2000). Air Quality Index for
Data Interpretation and Public Information.
Presented in International Conference, Centre for
Science and Environment, New Delhi, June 6-8,

USEPA (1998). Federal Register Vol.63 No.
236/Wednesday, December 9, 1998

Wassermann P.D. (1989). Neural Computing
theory and Practice. New York Van Nostrand


To top