The obtained values of MAE were slightly higher than the 20% of the mean for the MLPf models
in both sites. In addition, especially high values of the correlation coefficient and the index of
agreement were found. The correlation coefficient exceeded 0.8 at both sites and the IA values
were circa 0.9, for MLPf models. These results are comparable with those obtained in other
studies for predictions of daily PM10 concentrations11,12,15.
Regarding the performance of the GA-MLP models, the reduction of the number of input
variables did not appear to yield in a drastic improvement of the networks generalization ability.
However, the results remain satisfactory for the GA-MLP models, with the values of the statistic
indicators being slightly better at both sites, than the respective values for the MLPf models. The
method appears to be promising and can have significant applicability in the case of datasets with
a smaller number of training cases and large number of considered input variables, like the
prediction of daily average PM values.
In comparison to the traditional MLR models the performance of MLPf models was significantly
improved. This improvement is interpreted to 18% lower RMSE values at MAR while the
respective difference for the ZOG station is 9.2%. Figures 2,3 present the scatter-plots between
predicted and observed values of MLPf and MLR models, along with least squares equations, for
the two monitoring stations. It can be seen that the values of the coefficient of determination (R2)
are higher for the MLPf models for both sites by 10% and 6%. The GA-MLP models also
outperform the MLR models especially in the site of MAR. This improvement in the
performance is justified since the neural networks are trained to model the highly non-linear
relationships between particle concentrations and meteorological variables, which constrain the
predictive ability of multiple linear regression techniques6,