Time series of climatological and meteorological parameters show a dynamic trend where the prediction accuracy decreases as the prediction time increases. Both NN and BJ models confirm this behaviour, although the neural network approach is slightly more efficient when predicting temperatures; moreover, the gap grows up as the time lag increases (Fig. 5 (click here)).
Table 2 (click here) shows that N2, N3, N4 and N5 approaches are quite similar from the standard error point of view. On one hand the stochastic component represented by MA could be negligible; on the other hand the exogenus X component associated to the variable pressure seems not to include information which are able to improve the temperature prediction capability. And those case studies carried out using the whole set of meteorological parameters collected by CAMC show that prediction output is worse when one of them is added to the input set; in this case we may suppose that the studied case is purely autoregressive.
A further facet of the NN behaviour is that best performances are obtained for a topology with a minimum number of hidden units (nhu = 1,2); note that the validation set is used both for evaluating performances as well as for optimising the nhu value. During calibration phase, whenever the number of hidden units has been increased, we found a better performance of the model with respect to the training set and a worse feedback when using the validation set. This could be explained in terms of a physical behaviour characterized by a certain degree of linearity. In fact, a NN topology with a single or 2 hu mimes a quasi-linear approach, possibly related to an effective linear behaviour of the selected time series.