Air dispersion model and neural network: A new perspective for integrated models in the simulation of complex situations

Abstract

The study presents the development of an integrated model that optimises the performances of Neural Nets (NN) and dispersion models developed by the authors. Supervised NN models and dispersion models are two important techniques for evaluating air pollution concentrations. In our work we filter the concentration levels produced by an air pollution model with an NN to account for disagreement between the measured and predicted values. The performance of the new methodology was tested using two different formulations (VHDM and SPM) applied to the data set of the Prairie Grass, of the Copenhagen and of the Kincaid. The VHDM (Virtual Height Dispersion Model), a Gaussian formulation that take account vertical non-homogeneity of wind field and turbulence conditions, is applied to the Kincaid data set. The puff model (Skewed Puff Model) is based on the Monin-Obukhov similarity theory and is applied to the Prairie Grass and the Copenhagen data sets. Results show a marked improvement for all simulations when the neural network is added downstream of the dispersion model. This study demonstrated that the use of NN in order to correct the air dispersion model could be the reasonable model combination when the air pollution model gives some systematic error with respect to experimental data.