PMThe results on the prediction computed at 9a.m.
for the current day are clearly satisfactory, with a
true/predicted correlation of about 0.94 and a index
of correctly predicted exceedances higher than
0.80. Data deseasonalization seems a valuable approach
to increase of some points the average performances
indicators; improvements are however
less clear on the threshold exceedances prediction
indicators. The 2-days prediction appears as an
open problem, and the extension of air quality forecast
horizons is likely to require a great research
effort. In our opinion, dramatical performances
improvements are not to be expected by studying
new prediction algorithms: indeed, neural networks
constitute a flexible non-linear modelling approach,
able to learn very complex relationship from data.
On the other hand, the availability of more advanced
meteorological data, able to describe the air masses
motion in the atmosphere (e.g. vertical profiles of
wind speed and temperature, mixing height), can
greatly increase the informative content of the input
variables set, and may thus allow more significant
improvements of air quality predictions.
We remark that, although presently no clear trend
is detected on the PM10 time series, the situation
may evolve over time, thus requiring a retraining of
the predictor. Neural networks cannot be easily updated,
and in fact it will be necessary to identify
ex novo both the structure and parameters of the
network, in order to have an up-to-date predictor.
From this point of view, it is worth to mention that
lazy learning, a local linear modelling approach, can
constitute a viable alternative to neural networks;
in fact, according to (Birattari et al. [1999]), this
method may provide comparable prediction performances,
allowing at the same time a quicker design
and an easier update of the predictor.