The aim of the present work is to evaluate the potential of various developed neural network
models to provide reliable predictions of PM10 hourly concentrations 24-h in advance, a task that
has been demonstrated by research studies to present certain difficulties. The study area is the
Athens basin, which experiences a significant air pollution problem related to particulate matter.
Also, it has been shown that PM10 concentrations exhibit large variations within the day, with
peak hourly values being, frequently, several times higher than the daily average concentration. It
has been proposed that the diurnal variations of particle concentrations are very important for
exposure assessment and health effects are more linked with immediate exposure, rather than
with exposure during the past days. Thus, it appears that the availability of predictions of PM10
hourly concentrations could be of particular importance for authorities, at an operational level.
The results for two measurement locations of different type (a heavily trafficked location and an
suburban background location) are presented here. The PM10 data used cover the years of 2001-
2002. Artificial neural network models were developed using an ensemble of meteorological and
temporal input variables. A genetic algorithm procedure for the selection of the input variables
was also used and evaluated. The results of the neural network models were quite satisfactory.
Values of the coefficient of determination (R2) for independent test sets were 0.78-0.68 for the
two sites and values of the index of agreement 0.93-0.89 respectively. In addition, the mean
absolute errors were approximately 20% of the mean observed concentrations. The performance
of examined neural network models was superior in comparison to multiple linear regression
models that were developed in parallel (e.g. R2 for MLR models at the two sites 0.68-0.60
respectively). Their performance was also found adequate in the case of high concentration
events, with probabilities of detection exceeding 0.6 and low false alarm rates.
INTRODUCTION
Since