Air Pollution Monitoring and Modelling in the
Urban Area of Catania
V. Ferlito1, G. Nunnari2 & C. Oliveri1
1
XIII Direzione -Ecologia, Ambiente e Nettezza Urbana,
Comune di Catania, Italy.
Dipartimento di Ingegneria Elettrica, Elettronica e dei Sistemi
(DIEES), University of Catania, Italy
2
Abstract
This paper presents the network for air pollution monitoring implemented in the
city of Catania (Italy) since 1993. Information concerning the global level of air
pollution and trends in the area, as it appears from analysis of data recorded until
2002, is given and discussed in the light of control action undertaken by the
municipal authorities in order to reduce the level of pollution. Finally some
recent research initiatives, undertaken by the DIEES of the University of Catania
in cooperation with the Ecology and Environment Department of Catania City
Council, to model pollutant time series recorded by the described monitoring
network are described. The purpose of the research is to implement prediction
models that could serve as a decision support system (DSS) to control air quality
in the area.
1 Introduction
Air quality in the urban area of Catania is monitored by a network installed in
1993, now consisting of 17 air pollutant recording stations. A general view of
the monitoring network is shown in Fig. 1, while the Latitude, Longitude, and
altitude a.s.l. for the recording stations are given in Tab. I. The Catania
monitoring network is an exclusively urban one since most of the stations are
located in the town centre and suburbs. Air pollutants such as NOx, NO2, NO,
CO, PM10, CH4, and NMHC are recorded at all the stations, while SO2 is
recorded at all the stations except for those referred to as Messina and Moro in
Tab II. O3 is recorded at the Librino and Moro
Fig. 1: The Catania air pollution monitoring network: the circles indicate the
positions of the recording stations
Tab. I - Name, geographical coordinates and altitude of the recording stations
Name of the Station Longitude
Librino
15º 02’ 52,29"
Giovanni XXIII
15º 06’ 02,04"
Messina
15º 06’ 54,32"
Moro
15º 05’ 17.08"
P. Gravina
15º 04’ 20.00"
Fontana
15º 02’ 41.12"
Veneto
15º 05’ 52.54"
Europa
15º 06’ 23.80"
Gioeni
15º 04’ 58.16"
Cristallo
15º 05’ 38.67"
Michelangelo
15º 05’ 48.79"
Stesicoro
15º 05’ 15.00"
Giuffrida
15º 05’ 30.18"
Garibaldi
15º 04’ 33.79"
Z. Industriale
15º 03’ 33.32"
Risorgimento
15º 04’ 16.85"
Regione
15º 04’ 25.53"
Latitude
37º 29’ 9,95"
37º 30’ 33.68"
37º 32’ 05.2"
37º 31’ 39.55"
37º 32’ 24.15"
37º 30’ 57.60
37º 30’ 59.03"
37º 31’ 05.53"
37º 31’ 40.57"
37º 32’ 21.56"
37º 31’ 28.66"
37º 30’ 32.61"
37º 31’ 37.79"
37º 30’ 47.76"
37º 27’ 07.56"
37º 30’ 11.56"
37º 29’ 44.07"
Height a.s.l (m)
70
25
50
90
180
150
50
10
100
170
55
25
80
50
15
60
15
stations only. Finally three stations (Librino, Stesicoro and Giuffrida) also record
Toluene, E-Benzene, P-Xilene, H-Xilene and O-Xilene. The efficiency of the
monitoring network in terms of the percentage of valid data recorded during one
year was about 77.63 % and 79% during 2000 and 2001 respectively. This result
is considered acceptable bearing in mind that most of the measuring equipment
is now quite old, since it was installed about ten years ago. However, some
technical action was recently adopted in order to improve the global efficiency
of the monitoring network. In the urban area of Catania air pollution is mainly
due to vehicle exhaust emissions and episodically to the presence of Mt. Etna,
the largest active volcano in Europe.
2 Data analysis from 1993 to 2002: some results
The contribution to air pollution due to domestic heating is quite low thanks to
the favourable meteo-climatic conditions that characterise the area. Industrial
activities are mainly concentrated in one area about 6 Km from the city centre
which seems to have a limited impact on air quality. In the last few years action
undertaken to reduce the level of pollutant emissions, in both Italy and most
other European countries, mainly based on the use of higher-quality fuel with a
low Sulphur content, has made a considerable contribution towards attenuating
the problem. Moreover, the Catania municipal authorities have adopted
initiatives to improve the flow of traffic. As a consequence, the pollution levels
recorded in Catania over the five-year period were relatively low compared with
the attention and alarm levels laid down under national and European law. This
decreasing trend was observed by most of the recording stations. As an example,
the annual concentration of Benzene from 1998 to 2002, recorded at the three
stations referred to as Stesicoro, Giuffrida and Librino, is shown in Fig. 2
Annual average Benzene concentration trend
microgrammi/m3
14
12
1998
1999
2000
2001
2002
10
8
6
4
2
0
P. Stesicoro
V. Giuffrida
Librino
Fig. 2: Annual Benzene concentration from 1998 to 2002
It is possible to observe that at the Stesicoro and Giuffrida stations, affected by
heavy traffic, the annual average concentration has continuously decreased,
while at the Librino station, located in the suburbs, the concentration is almost
constant during the considered time interval. Similar trends were observed for
other pollutants such as SO2 and CO. However, it should be observed that for
some other pollutants, such as for instance Ozone, a clear trend was not observed
(see Fig. 3)
Annual average Ozone concentration trend
ug/m3
45
40
35
1998
1999
2001
2002
2000
30
25
20
15
10
5
0
P. A. Moro
Librino
Fig. 3: Annual average Ozone concentration from 1998 to 2002
The annual average concentration computed for each pollutant recorded during
2002 is given in Tab. II. From this table it is possible to see that the annual
average level of air pollution is at present within the standard quality limits.
However, some pollutants, such as NO2, recorded at specific points, namely at
Giovanni XXIII, Michelangelo and Risorgimento exceed the threshold
introduced in a recent directive (directive N.60 dated 4/4/2002 issued by the
Italian Ministry for the Environment and Territory).
Tab. II - Average annual concentration of the main pollutants during 2002
CO
Librino
Giovanni XXIII
Messina
Moro
Gravina
Fontana
Veneto
Europa
Gioeni
Cristallo
Michelangelo
Stesicoro
Giuffrida
Garibaldi
Zona Industriale
Risorgimento
Regione
3
NO2
3
(mg/m )
(µg/m )
0.52
1.76
1.87
0.73
1.07
0.78
1.88
1.48
1.91
0.55
1.87
2.21
1.57
1.49
0.34
2.03
1.32
38.52
80.63
SO2
PM10
O3
(µg/m )
(µg/m )
(µg/m )
(µg/m )
24.86
21.77
23.59
17.02
30,84
1,89
3
4.52
52.97
47.34
86.79
55.60
57.21
51.84
87.64
55.30
59.95
50.61
54.02
43.30
2.26
4.86
4.29
1.94
2.62
2.98
1.40
4.71
4.94
2.42
1.02
3.95
1.64
3
3
Benz
3
39,00
26.30
40.63
36.74
31.16
28.51
32.91
29.05
30.95
18.86
37.56
33.44
8,61
4,42
From the data given in Tab. II, the monitoring stations can be ordered into three
sets. The first subset, characterised by the highest level of pollution, includes the
stations referred to as Stesicoro, Veneto, Giovanni XXIII, Risorgimento, Gioeni,
and Michelangelo. The second subset, containing stations characterised by an
intermediate level of pollution, includes the stations referred to as Europa,
Messina, Garibaldi, Gravina, Giuffrida, Regione and Fontana. Finally, the third
subset, containing the stations with the lowest level of air pollution, includes the
stations referred to as Cristallo, Librino and Zona Industriale.
3 The role of the wind in the dispersion of pollutants in Catania
Catania is a typical Mediterranean town characterised by a sea breeze wind
regime. Analysis of a typical day referred to both pollutants and wind allows
some understanding of the role of wind speed in the pollutant dispersion process.
A typical day concentration is calculated on the series of average hourly values
C as follows: 24 values, one for each hour of the day being considered, for each
day of the year. Each average is therefore calculated on 365 values (366 in a leap
year) recorded at the same times; this is expressed in a formula as:
C (k ) =
1 364
∑ C (k + h ⋅ 24),
364 h =0
k = 1,...,24
(1)
As an example the typical CO (Carbon Monoxide) days computed for the years
1996, 1997 and 1998 at the Stesicoro station are shown in Fig. 4.
Fig. 4: typical day for CO during 1995, 19996 and 1997
Fig. 5: CO and WS typical days
It is possible to observe that there are two maximum concentration values
approximately at 8 a.m and 8 p,m on a typical day for CO. These two
maximums can easily be related to pollution due to traffic.
However, from Fig. 4 it is possible to observe that a minimum in the daily CO
concentration is reached at about 1 p.m. which cannot be explained in terms of
vehicular traffic, since 1 p.m. is usually characterised by heavy traffic. This
apparently strange behaviour could be explained by looking at Fig. 5, where the
CO and wind speed (WS) on a typical day are compared. Lower CO
concentrations are recorded when WS assumes the maximum value, while
higher CO concentrations correspond to lower WS values. From this simple
analysis it seems reasonable to relate the daily behaviour of CO to local
atmospheric conditions that are established in the area, and in particular to the
course of the wind. Higher wind speed values contribute to the dispersion of air
pollutants and hence to lower pollutant concentrations. Wind speed, together
with other meteorological variables such as atmospheric pressure, solar
radiation, temperature and relative humidity, are natural candidates to model the
pollutant time series recorded by the monitoring network.
4 Statistical Modelling of Air Pollution Time series
Statistical models are based on semi-empirical relationships between available
pollution data and other available measurements. They are frequently used in air
pollution studies for short-term (e.g. 1 day) forecasting applied to real-time
control of emissions or to air quality assessment. Statistical methods, unlike
deterministic methods, do not aim to describe the level of pollution as a
phenomenology driven cause-effect problem. Statistical models are characterised
by their direct use of air quality measurements to infer semi-empirical
relationships. In particular, they are useful in situations where the information
available from measured concentration trends is generally more relevant than
that obtained from deterministic analyses. Some general information about basic
statistical methods for air pollution data is given in Zannetti [1]. More recent
statistical techniques are described by Finzi et al. [2].
The most popular statistical models considered for modelling air pollution time
series are ARMAX and NARX models. ARMAX (Auto Regressive Moving
Average with eXogenous inputs) models [3], have been considered to be one of
the most cost-effective approaches for time series analysis and many Authors
have been inspired to apply this technique in developing pollutant forecast
models for SO2 time series as well as for other pollutants such as Ozone, NOx
etc. (see, for instance, Finzi et al. [4]. NARX (Non-linear Auto-Regressive with
eXogenous inputs) models can be considered as the generalisation of ARX
models to the non linear case. The general form of a NARX model is the
following:
y(t + 1) = f ( y(t ), y(t − 1),...y(t − n y + 1), u1 (t ), u1 (t − 1),...,
u1 (t − n1 + 1),..,u q (t ), u q (t − 1),...,u q (t − nq + 1)) + e(t )
(2)
where f is an unknown non-linear function, u1, …,uq are the exogenous model
inputs and ny, n1, … nq are integer numbers related to the model order. Recent
studies show that NARX models are more appropriate than ARX for pollutant
time series modelling [5], [6]. A straightforward way to identify a NARX model
is the Multilayer Perceptron (MLP) neural network based approach that has been
used by several authors such as Boznar [6], Arena et al. [7], and Gardner and
Dorling [8], [9]. In this paper an MLP neural approach was used to identify a 1day-ahead prediction model for the daily maximum CO concentration at the
station referred to as Stesicoro, located in the city centre of Catania. To this end,
data recorded from 1996 to 2001 were considered. In more detail, the data set
was arranged into two subsets: the first set, referred as the learning set, includes
data recorded in 1996, 1998, 1999 2000 and 2001; the second subset, referred to
as the testing set, includes data recorded during 1997. The daily maximum time
series were extracted from the original hourly CO time series by taking the
maximum hourly average concentration of CO for each day. The prediction
model was organised as follows. Based on the considerations made in the
previous section about the role of wind speed, the daily average value of this
variable was considered as the only exogenous model input variable. Moreover,
it was assumed that the dispersion process of CO is characterised by a short
degree of memory and hence it was decided to use only WS(t) and WS (t+1) as
model input values, t being the day before the prediction. Furthermore, since the
model is an autoregressive one, one regression of the output variable was
considered as model input value. The model target was the daily maximum
concentration of CO on day (t+1). A number of trials were carried out to find the
number of neurons in the single hidden layer of the MLP neural networks
considered. The performance of the each prediction model implemented was
evaluated computing an appropriate number of performance indexes. These
indexes were organised into two different sets. The first set, referred to here as
global fit indexes, evaluates the fitting capabilities of the overall time series. This
set includes the Bias (see expression (3), the RMSE (Root Square Mean Error)
(4), the MAE (Mean Absolute Error) (5) which give estimates of the average
error, and the index of agreement d (6) which is a bounded relative measure
capable of measuring the degree to which predictions are error-free.
1
N
Bias :
1
N
MAE :
N
∑ (P − O )
i
i =1
N
∑ P −O
i
i =1
(4)
i
N
1
N
RMSE :
(3)
i
∑ (P − O )
i
i =1
N
d:
1−
∑ (P − O )
i
i =1
N
_
i
(5)
2
i
∑( P − O + O
i =1
2
i
i
_
−O)
(6)
2
In the expressions above, O and P indicate the observed and predicted time
series respectively, the overbar indicates the mean value, and the suffix i a
generic element of the time series.
The second set of indexes was studied to specifically evaluate model capabilities
in predicting critical CO episodes. The most important index of this set is the
success index, SI which indicates how well the exceedances were predicted. It is
not affected by a large number of correctly forecasted non-exceedances and
therefore is useful for evaluating rare events. Other indexes in this second set are
the probability of detection index, SP, which assess the ability to predict CO
exceedances and the false alarm rate index FA, which indicates the percentage
frequency of instances when a forecasted of a pollutant concentration
exceedance did not actually occur.
These indexes are recommended by the ETC-AQ (see Van Aalst and De Leeuw,
[10]) and the EPA. Expressions for SP, SR, FA and SI are given
below:
SP% = 100
NP
,
NO
SI % = 100(
N P N + N P − NO − N F
+
− 1)
NO
N − NO
SR% = 100
NP
,
NF
FA% = 100 − SR%,
(7)
where No is the total number of observed exceedances, NP is the number of
correctly predicted exceedances, NF is the total number of forecasted
exceedances and N the total number of data points.
5 Results and Conclusions
The results obtained for some of the MLP models implemented are given in Tab.
III. Mod 1, Mod 2 and Mod 3 indicate three MLP neural models having 8, 10
and 14 neurons in the single hidden layer.
Tab. III - Global Performance Indexes
Mod 1
Mod 2
Mod 3
Bias
0.316
0.31
0.25
MAE
1.77
1.76
1.78
RMSE
2.32
2.31
2.31
d
0.66
0.67
0.71
The models implemented perform in a quite similar fashion in terms of global
performance indexes. However Model 3 shows a slightly better index of
agreement (d). The exceedance performance indexes for Model 3, computed for
thresholds of 6, 9 and 10 mg/m3 , are shown in Fig. 6. The same indexes
obtained for Mod 1 and Mod 2 are not given here for the sake of brevity but they
are similar to that of Mod 3
Exceedance Indexes
100
50
0
SP%
SR%
FA%
SI%
th=6
76.72
76.72
23.28
40.95
th=9
38.2
65.38
34.2
30.13
th=10
10
50
50
8.52
Fig. 6: Performance Indexes for the Model 3
From these results it seems that the models implemented perform satisfactorily
in terms of exceedance indexes only at the lower threshold (6 mg/m3). This
indicates that wind speed is probably not enough to explain the dynamic of the
pollution process considered. More accurate analysis and trials are required to
see if the use of other meteorological data can improve model performance.
Data about emissions is probably also necessary. Emissions are intrinsically
difficult to measure in this case, since they are of a distributed type (vehicular
traffic). To overcome this shortcoming, it is planned to use indirect measures of
emissions, i.e. measures of traffic intensity performed at selected points in the
network. This is planned to be done in the near future.
REFERENCES
[1] Zannetti P., Air Pollution Modelling, Theories, Computational Methods and
Available Software, Van Nostrand Reinhold, New York, 1990.
[2] Finzi G., Pirovano G., Volta M., (2001), Gestione della Qualità dell'Aria Modelli di Simulazione e Previsione, McGraw-Hill Libri Italia Srl, Milano.
[3] Box G. E. P., Jenkins G. M., Reinsel G. C., (1994), Time Series Analysis,
Forecasting and Control, Prentice Hall.
[4] Finzi G., Volta M., Nucifora A., Nunnari G., (1998), Real-Time Ozone
Episode Forecast: a Comparison between Neural Network and Grey-Box
Models, Proc. of International ICSC/IFAC Symposium on Neural
Computation - NC'98, Vienna, pp 854-860.
[5] Nunnari G., Nucifora A., Randieri C., (1998) The application of neural
techniques to the modelling of time-series of atmospheric pollution data
Ecological Modelling, 111, 187-205.
[6] Boznar M., Lesjak, and P. Mlakar (1993), A Neural Network-Based Method
for Short-Term Predictions of Ambient SO2 Concentrations in Highly
Polluted Industrial Areas of Complex Terrain, Atmospheric Environment,
Vol. 27B, No. 2, pp. 221-230.
[7] Arena P., Baglio S., Castorina C., Fortuna L., Nunnari G. (1996), A Neural
Architecture to Predict Pollution in Industrial Areas, Prooc. of the
International Conference on Neural Network - ICNN, Vol. 4, pp. 2107-2112,
Washington, USA.
[8] Gardner M.W., S.R. Dorling (1998) Artificial neural networks (the
multilayer perceptron), A review of applications in the atmospheric sciences,
Atmospheric Environment, Vol.32, N.14, pp.2627-2636.
[9] Gardner M.W., S.R.Dorling (1999) Neural network modelling and prediction
of hourly NOx and NO2 concentration in urban air in London, Atmospheric
Environment, Vol. 33, pp.709-719.
[10]Van Aalst R.M., F.A.A.M. de Leeuw (eds.) (1997) National Ozone
Forecasting System And International Data Exchange In Northwest Europe,
European Topic Centre on Air Quality.