Exploring the use of soft-computing and artificial intelligence techniques in atmospheric pollution modelling

摘要

Many different mathematical models for monitoring and controlling the atmospheric pollution have been applied in recent years. In this paper we illustrate a method for the forecast of atmospheric pollution caused by particulate matter (PM_(10)). This model has been developed using Bayesian networks to represent the pollutant behaviour and then the parameters of these networks have been optimized using genetic algorithms. We adopted this approach because the probabilistic nature of the Bayesian models well reflects the uncertainly characterizing the atmospheric pollution data and meteorological parameters. Several experiments in this direction proved that the best networks parameters (number of nodes, discretisation ranges, number of evidence) cannot be easily determined a-priori. We, therefore, tune some of them with a genetic evolution of their values that uses the model error as a fitness function. The simulation results reported in the paper, show that a significant advantage could come from this approach.