Prediction of PM 2.5 concentrations several hours in advance using neural networks in Santiago, Chile

Abstract

Hourly average concentrations of PM 2.5 have been measured at a fixed point in the downtown area of Santiago, Chile. We have focused our attention on data for the months that register higher values, from May to September, on years 1994 and 1995. We show that it is possible to predict concentrations at any hour of the day, by fitting a function of the 24 hourly average concentrations measured on the previous day. We have compared the predictions produced by three different methods: multilayer neural networks, linear regression and persistence. Overall, the neural network gives the best results. Prediction errors go from 30% for early hours to 60% for late hours. In order to improve predictions, the effect of noise reduction, rearrangement of the data and explicit consideration of meteorological variables are discussed.