Pattern selection strategies for a neural network-based short term air pollution prediction model

Abstract

SO/sub 2/ air pollution around coal fired thermal power plants is still one of the biggest environmental problems in Slovenia. A multilayer perceptron neural network based model for short term predictions of ambient SO/sub 2/ concentrations has been developed for locations around the Sostanj Thermal Power Plant (M. Boznar et al., 1993). Selection of the patterns used for neural network based model training is one of the most important tasks that should be solved in order to achieve good generalising capabilities of the model. Two different types of pattern selection strategies were developed: a meteorological knowledge based strategy and a Kohonen neural network based strategy. The strategies are explained in the case of prediction models for the Zavodnje automatic measuring station in the surroundings of the Sostanj Thermal Power Plant. The pattern selection strategies developed for the air pollution forecasting model can easily be adapted for use in other fields where models are built using large databases.