Development and Assessment of Neural Network and Multiple Regression Models in Order to Predict PM10 Levels in a Medium-sized Mediterranean City

Abstract

Suspended particulate matter is significantly related to the degradation of air quality in urban agglomerations, generating adverse health effects. Therefore, the ability to make accurate predictions of particulate ambient concentrations is important in order to improve public awareness and air quality management. This study aims at developing models using multiple regression and neural network (NN) methods that might produce accurate 24-h predictions of daily average (DA) value of PM10 concentration and at comparatively assessing the above mentioned techniques. Pollution and meteorological data were collected in the urban area of Volos, a medium-sized coastal city in central Greece, whose population and industrialization is continuously increasing. Both models utilize five variables as inputs, which incorporate meteorology (difference between daily maximum and minimum hourly value of ground temperature and DA value of wind speed), persistency in PM10 levels and weekly and annual variation of PM10 concentration. The validation of the models revealed that NN model showed slightly better skills in forecasting PM10 concentrations, as the regression and the NN model can forecast 55 and 61% of the variance of the data, respectively. In addition, several statistical indexes were calculated in order to verify the quality and reliability of the developed models. The results showed that their skill scores are satisfying, presenting minor differences. It was also found that both are capable of predicting the exceedances of the limit value of 50 μg/m3 at a satisfactory level.