Artificial neural networks prediction of PM10 in the Milan area

PM10 constitutes a major concern for Milan air quality. We presents a series of results obtained applying different neural networks approaches to the PM10 prediction problem. The 1-day ahead prediction shows a satisfactory level of accuracy, which may be further improved if a proper deseasonalization approach is adopted, thus transferring some a priori knowledge in the data pre-processing step. Then, we tackle the problem of the 2-days ahead prediction; in order to optimize the neural network architecture identification procedure, we try a pruning approach besides the usual trial and error one. Prediction performances are very close between the two models, and denote a significant decrease of accuracy with respect to the 1-day case, even though some meteorological improper (i.e. future measures) input is added to the model structure.