ESTUDIO DE SERIES TEMPORALES DE CONTAMINACIÓN AMBIENTAL MEDIANTE TÉCNICAS DE REDES NEURONALES ARTIFICIALES

Abstract

An artificial neural network for the forecasting of concentrations of fine particulate matter in the atmosphere was designed. The data set analyzed corresponds to three years of pm2.5 time series (particulate matter in suspension with aerodynamic diameter less than 2, 5 microns), measured in a station that belongs to Santiago's monitoring network (Red MACAM) and is located near downtown. We consider measurements of concentrations between May and August for years between 1994 and 1996. In order to find the optimal time spacing between data and the number of values into the past necessary to forecast a future value, two standard tests were performed, Average Mutual Information (AMI) and False Nearest Neighbours (FNN). The results of these tests suggest that the most convenient choice for modelling was to use 4 data with 6 hour spacing on a given day as input in order to forecast the value at 6 AM on the following day. Once the number and type of input and output variables are fixed, we implemented a forecasting model based on the neural network technique. We used a feedforward multilayer neural network and we trained it with the backpropagation algorithm. We tested networks with none, one and two hidden layers. The best model was one with one hidden layer, in contradiction with a previous study that found that minimum error was obtained with a net without hidden layer. Forecasts with the neural network are more accurate than those produced with a persistence model (the value six hours ahead is the same as the actual value). (English) [ABSTRACT FROM AUTHOR]