Modelling study for forecasting gaseous pollutants levels in a urban area

Atmospheric pollution is an important topic in environmental sciences. Nowadays the quality and the quantity of the data from air quality monitoring networks are significantly increased, but, for an effective management and assessment of this information, innovative data analysis methodologies have been developed. Approaches coming from advanced statistical methods were introduced in modeling and forecasting procedure to define operational techniques for atmospheric pollutants characterization at different scales. In this paper we present an application of artificial neural networks (ANN) for forecasting atmospheric gaseous pollutants. Starting from hourly data collected in Basilicata (southern Italy), from 1998 to 2007, we select the best dataset in terms of minimum data missing percentage. The applied model is a feed-forward multi-layer perceptron with an only hidden layer. The conjugate gradient learning algorithm is used. The learning capability of the model and the average goodness of the prediction are evaluated by Mean Absolute Percentage Error. The goal is to evaluate the performance the ANN model for forecasting 24-hourly data on the base of only the 24-hourly data collected in the previous day and to quantify the improvement obtained with different input strategies (optimal mix of pollutants defined by data correlation structure analysis). The preliminary results suggest that the dynamical characteristics of the gaseous pollutants may play a fundamental role in the definition of the forecasting procedure. Moreover, results confirm that the correlation structure analysis may be usefully applied for identifying the optimal strategy of data input selection. Nevertheless the quality of data represents the main limit of forecasting techniques at local scale.