Neural network model for forecasting atmospheric particulate levels

Procedures based on artificial neural network (ANN) have been applied with success to forecast levels of atmospheric pollutants. These techniques show a capability to make regressive approximation of non-linear functions in high-dimensional space and they are more flexible in comparison to traditional statistical techniques. In this paper we present a short review of recent applications of ANN models for forecasting atmospheric particulate levels and the results of a study carried out to forecast hourly levels of PM10 in urban area starting from data measured in N-previous days. In particular we analyze PM10 hourly concentrations measured from March 2001 to February 2002 in three stations of the air quality monitoring network of Potenza town (southern Italy). The applied ANN model is a feed-forward multi-layer perceptron (MLP) 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 (MAPE) and by the number of concentration values that the model is not able to predict (NP). The results indicate that, in the study area, a simple model of ANN is able to forecast PM10 hourly levels with a good approximation but the quality of data, in terms of presence of data missing, represents the main limit of these forecasting techniques at local scale. In order to improve the model performance increasing the number of input variables, the results suggest not only to take into account meteorological parameters but also to better characterize the dynamic features of emission source pattern.