Coupling of Neural Network and Dispersion Models: a Novel Methodology for Air Pollution Models

INTRODUCTION Air pollution models have not so far been able to reproduce the ground level concentrations because, for example, it is recognised that deterministic models many times cannot provide an adequate correlation between hourly predictions and observed data paired in time and space. A supervised Neural Net (NN) model in forecasting concentrations levels, have to take in to account the influence of the system variables, such as source emission factors, turbulence conditions, local topolography, reactions rate, by using an appropriate training on the available experimental data. The proposed approach deals with the development of an integrated model that optimises the performances of each methodology (NN and dispersion models). We have applied a Neural filter to an operative model (VHDM: Virtual Height Dispersion Model). VHDM (Tirabassi and Rizza, 1994) is used for evaluating ground level concentrations from elevated sources that applies a new Gaussian formulation, where the source height is expressed by simple functions of the vertical profiles of wind and turbulent diffusivity. The dispersion model can be applied routinely using as input simple ground level meteorological data acquired by an automatic network.