Author(s)

M. Boznar and P. Mlakar

Abstract

Pattern selection strategies for SC>2 forecasting models M. Boznar and P. Mlakar Jo zef Stefan Institute, Jamova 39, SI-1000 Ljubljana, Slovenia Abstract SO] pollution of the atmosphere is one of the problems that have not yet been solved in Slovenia, especially around large coal fired thermal power plants. In past years we have developed Perceptron neural network based models for short - term forecasting of ambient SO] concentrations. Basically the forecasting model is taught with patterns that represent SO] concentrations and the meteorological history of the observed polluted site. Therefore it is very important how these patterns are selected in order to obtain a model that is capable of accurate forecasting for as many different pollution situations as possible. Three types of pattern selection strategies were developed. The models built according to these pattern selection strategies were tested

Keywords