Air quality data remediation by means of ANN

We present an application of neural networks to air quality time series remediation. The focus has been set on photochemical pollutants, and particularly on ozone, considering statistical correlations between precursors and secondary pollutants. After a preliminary study of the phenomenon, we tried to adapt a predictive MLP (multi layer perceptron) network to fulfill data gaps. The selected input was, along with ozone series, ozone precursors (NO_x) and meteorological variables (solar radiation, wind velocity and temperature). We then proceeded in selecting the most representative periods for the ozone cycle. We ran all tests for a 80-hours validation set (the most representative gap width in our data base) and an accuracy analysis with respect to gap width as been performed too. In order to maximize the process automation, a software tool has been implemented in the Matlab™ environment. The ANN validation showed generally good results but a considerable instability in data prediction has been found out. The re-introduction of predicted data as input of following simulations generates an uncontrolled error propagation scarcely highlighted by the error autocorrelation analysis usually performed.