Artificial Neural Networks - a Useful Tool in Air Pollution and Meteorological Modelling

Abstract

Artificial neural networks have become a widely used tool in several air pollution and meteorological applications.Yi and Prybutok (1996) used MPNN for surface ozone predictions, as well as Comrie (1997).Several prediction models were also made for other pollutants; for instance for SO 2 (Božnar et al., 1993) and for CO (Moseholm et al., 1996).Marzban & Stumpf (1996) used MPNN for predicting the existence of tornadoes.A review article by Gardner (1998) described a variety of applications, mainly in the field of air pollution forecasting and pattern classification.Though the number of applications is growing, especially in recent years, no special attention has been paid to the principles of artificial neural network usage in environmental applications.Our group first established a method for short term forecasting of SO 2 concentrations on the basis of a multilayer perceptron neural network (Božnar et al, 1993), but in the following years we use an artificial neural networks in several other applications that differ very much each another.In this article we intend to show examples of a variety of applications of artificial neural networks in air pollution and the meteorological field.Examples are taken from our past experience, extending over a decade.Several applications in this field start from fundamentals and too much attention is paid to optimization and speeding up of the learning algorithms.From our experience this should be a minor problem for an environmental modeller and does not significantly affect the final model quality if modern tools are used.In the process of model construction other factors are much more crucial -such as feature determination, pattern selection, and learning process optimization.These are the methods that are derived from the basic principle of artificial neural networks -that is the ability to learn information from given examples.In this article we intend to show some solutions for the effective transformation of measured information into air pollution and meteorological models.We hope that the variety of examples will inspire new applications and methods that will serve the air pollution modelling community.The mystique of artificial neural networks, derived directly from their name, prevents many modellers from using them.It is the purpose of this article to demystify this useful mathematical tool and in this way encourage its usage. www.intechopen.comAdvanced Air Pollution 496 Artificial neural networks -several types for different purposesArtificial neural networks can be divided into several groups according to their topology.The tool was firstly widely used in the pattern recognition field.The topologies vary from feed forward neural networks with several hidden layers, to topologies with backward loops that make the result sequence dependent, to fuzzy logic and several automatic sorting tools.A detailed explanation of this groups is far beyond the scope of this article.The reader interested in this issue can get information from several books (Lawrence, 1991).In this article we focus on two main "species" of artificial neural networks that can cover a huge variety of air pollution and meteorological modelling applications.The two selected are the Multilayer Perceptron artificial Neural Network (MPNN) and the Kohonen neural network (KNN).Both can be replaced by other artificial neural networks for the same purpose, but this does not change the method of using these tools.In this article MPNN and KNN can both be treated as one of the best possible solutions.The authors of this article have no intention to argue about the qualifications of other topologies.In this article it will be shown what the most suitable applications of MPNN and KNN are.The latter is not so widely used although it has great potential in environmental problems.MPNN is mathematically speaking a universal approximator (Hornik, 1991;Kurkova, 1992).It can reconstruct arbitrary multivariable and highly non-linear functions.Therefore it is a suitable tool for modelling atmospheric phenomena whose behaviour has not yet been described by formulas but is only known from measured examples.KNN, on the other hand, is a structure capable of sorting a multitude of multivariable samples or patterns into groups of similar ones.It is important that it can find these groups without a teacher -so-called unsupervised learning.This ability becomes extremely important when dealing with multivariable patterns where similarity rules are not obvious.