Prediction of Particlulate Air Pollution using Neural Techniques

We have analysed the possibility of predicting hourly average concentrations of suspended atmospheric particulate matter with aerodynamic diameter less than 2.5 microns (PM2.5) several hours in advance using data obtained in downtown Santiago, Chile. By performing some standard tests used in the study of dynamical systems, we are able to extract some features of the time series of data. We use this information to estimate the amount of data on the past to be used as input to a neural network in order to predict future values of PM2.5 concentrations. We show that improvement of predictions is possible by using another neural network for noise reduction on the original series. The best results are obtained with a type of neural network which is equivalent to a linear regression. Up to six hours in advance, predictions generated in this way have significantly smaller errors than predictions based on the persistence of the long term average of the data.