Cite abstract as Author(s) (2009), Title, European Aerosol Conference 2009, Karlsruhe, Abstract T031A17

companyscourgeΤεχνίτη Νοημοσύνη και Ρομποτική

19 Οκτ 2013 (πριν από 3 χρόνια και 8 μήνες)

121 εμφανίσεις



Urban PM
10
modelling using neural network with weather forecast

K.B. Ang
1
, G. Baumbach
1
and K.G. Gutbrod
2


1
Institute of Process Engineering and Power Plant Technology (IVD), Department of Air Quality Control (RdL),
Universitaet Stuttgart, Pfaffenwaldring 23, 70569 Stuttgart, Germany
2
meteoblue AG, Clarastrasse 2, CH-4058 Basel, Switzerland
Keywords: PM
10
, neural network, weather forecaster, air quality modelling


Weather forecasters provide information on
meteorological parameters, such as probable wind
characteristics and precipitation amount for an area,
which would generally suffice for the general
population. If such forecasters could be extended to
provide additional information on probable ambient
PM
10
concentrations in a day in advance, the benefits
would be two-fold. First, such extended weather
forecaster could act as both an alarm for the quality
of ambient air and bad weather. Second, the
information derived from the extended weather
forecaster could aid in public education.
The potential of using neural network models
to predict the daily average PM
10
concentrations one
day ahead at an urban background site and a heavily
trafficked site in Stuttgart (Germany) was evaluated.
The two developed neural network models can be
simplified as follow:

Input parameters from 1 May 2007 to 1 May 2008:
PM
10, day 0
(Measurements from PM
10

gravimetric analyses)
Mixing height
, day 0
(Measurements from radio
soundings)
Temperature
, day 1
Weather forecast
Wind velocity
, day 1
Weather forecast
Wind direction
, day 1
Weather forecast
Precipitation
, day 1
Weather forecast

Target parameter: PM
10, day 1


The entire data set was divided into three
separate subsets (training, validation and test sets) for
the needs of the neural network model’s development
and evaluation. The early stopping technique was
implemented during the training of the neural
network. The training process was finalised when the
selected measure of errors for the validation data set
reached a minimum. For the evaluation of the neural
network models’ performance, several performance
indicators were computed, namely the mean absolute
error (MAE), the root mean square error (RMSE), the
fractional bias (FB), the index of agreement (IA) and
the squared correlation coefficient (R²).The results
for the developed neural network models’ for the two
sites are summarised in Table 1.
The time-series of predicted and observed 24
h PM
10
concentrations at the two sites are presented
in Figure 1 to illustrate the models’ performance. In
the urban background site, the results obtained with
the model demonstrate good agreement with the
observed PM
10
concentrations, with the IA and R²
values of 0.79 and 0.73 respectively. In contrast to
the results obtained at the urban background site, the
corresponding IA and R² values are clearly lower at
the traffic site. In general, the model for the traffic
site was able to reasonably reproduce the day-to-day
variation of PM
10
concentrations. However, part of
the lower measured values was underpredicted. This
result was expected, as the PM
10
concentrations at the
traffic site is strongly influenced by the local road
dust and thus making it more difficult for PM
10

prediction. At both sites, the local PM
10
pollution
episode during late December 2007 was also not
fully captured by the models.

Table 1. Performance indicators for the neural
network models
Site Urban
Background
Traffic
MAE in µg/m³ 6.57 12.42
RMSE in µg/m³ 1.51 0.29
FB in % -11.00 0.19
IA 0.79 0.67
R² 0.73 0.66

Figure 1. Time series of predicted and observed PM
10

concentrations at Bad Cannstatt and Neckartor

The overall models’ results illustrates a possibility of
effective use on the operational level for forecasting
PM
10
concentrations one day in advance using
Numerical Mesoscale Model (NMM) weather
forecasts as input parameters. The existence of
quality meteorological forecasts is an important
factor for the models’ successful implementation for
real-time predictions. With appropriate input data
sets, such models could be modified and adopted to
other locations as well.