Artificial Neural Network versus autoregressive approach: Prediction of total ozone time series

Abstract

The central premise of the present research is to judge the performance of Artificial Neural Network against that of the conventional statistical autoregressive approach in predicting the mean monthly total ozone concentration one month in advance over Arosa, a locality in Switzerland (46.8°N/9.68°E). Prior to the implementation of neural net methodology to the dataset, some significant developments in the application of Artificial Neural Networks to the pollution study have been reviewed. Basic principles of feed forward neural nets are also briefly canvassed. In the implementation phase, instead of considering meteorological parameters, the past values of the given variable have been considered as predictor. After rigorous study it has been established that a three hidden layers Artificial Neural Network with Backpropagation algorithm produces better forecasts than a linear autoregressive procedure.