Application of Sampling Theory to Forecast Ozone by Neural Network

摘要

In the present work, we analyzed environmental data by using neural net techniques for ozone prediction. The data concerns a period of two years (2006 and 2007) and comes from a monitoring station of air quality of Rome. The aim of this paper is to suggest a strategy for choosing an optimal set of input patterns to optimize the learning process during training and generalization phase, and to improve computation reliability of a Neural Net (NN). The selection of patterns combined with NN improves capability and accuracy of ozone prediction and goodness of models obtained. In particular, the approach considers two different methodologies for selecting an optimal set of input patterns: random patterns selection and cluster (K-means algorithm) ones. Results show significant differences between the methodologies: the NN's performance is always better when the patterns are obtained using our method based on cluster analysis than the conventional random pattern choice.