We have analyzed the possibility to predict hourly averages of sulfur dioxi
de concentrations in the atmosphere at a site not far from the downtown are
a in the city of Santiago, Chile. We have compared the forecasts produced a
ssuming persistence, linear regressions and feed forward neural networks. T
he effect of meteorological conditions is included by using forecasted valu
es of temperature, relative humidity and wind speed at the time of the inte
nded prediction as inputs to the different models. The best predictions for
hourly averages are obtained with a three-layer neural network that has ho
urly averages of sulfur dioxide concentrations every 6 h on the previous da
y plus the actual values of the meteorological variables as input. Training
the network with 1995 data, error in 8 h in advance prediction for 1996 da
ta is of the order of 30%. (C) 2001 Elsevier Science Ltd. All rights reserv
ed.