A modeling effort was conducted in order to investigate the potential of artificial neural network
methods, as tools for the achievement of the difficult task of the prediction of PM10 hourly
concentrations, 24-h in advance. This difficulty has been demonstrated in previous research
studies and is attributed to the complexity and diversity of the mechanisms governing the
determination of atmospheric particle levels, and to the necessity of using large datasets with a
high level of internal noise for the development of the models. However, the existence of such
predictions is essential for a large area as the metropolitan complex of Athens, for which PM10 is
a priority pollutant, with concentrations exceeding legislated standards in the majority of
measuring locations, and presenting significant diurnal variability with important implications for
public health.
MLP neural networks were constructed using a methodology already successfully implemented
for the prediction of daily average PM10 concentrations14 and a range of predictor variables
including a set of meteorological parameters, which are routinely monitored and forecasted by
authorities. Additional MLP models were trained on a reduced number of input variables, preselected
using a genetic algorithm based process. This was done in order to validate the
assumption that the reduced size of the training dataset would result at an improvement of the
generalization ability. Multiple linear regression was used as a method for the comparison of the
performance of the ANN models.
The results produced by the ANN models were rather satisfactory, for the two selected
measurement locations. The used performance indices point out a measurable improvement when
applying the MLPf and GA-MLP models instead of the regression based MLR model. This result
is expected due to the existence of strong non-linear associations between PM10 and the used
predictor variables. However, in real-time forecasting conditions, some compromise in
performance can be expected, due to the possibility of less accurate meteorological forecasts. The
existence of quality meteorological forecasts is a matter of great importance, since it proved that
the incorporation of such predictors leads to a significant improvement in the predictive power of
the models. In view of this, the research effort in the immediate future should be supplemented
by a sensitivity analysis and by the development of models based on forecasted values of the
meteorological parameters.
Regarding the genetic algorithm input selection procedure it did not lead in a dramatic
improvement in the generalization ability of the MLP models. However, despite the considerable
reduction in the number of used input variables the results were comparable or slightly superior
to those of the models using the entire set of input variables. The reduction of the required
meteorological parameters might prove significant in operational conditions with the removal of
factors introducing uncertainty in the model. Moreover, the method is expected to produce better
results in the case of datasets with fewer training cases or in the intended future utilization of
radial basis function networks for the modeling of particle concentrations.